BME2022 Paper Submission & Registration
9th Dutch Bio-Medical Engineering Conference

Powered by
© Fyper VOF.
Conference Websites
15:00   Poster session 1
Monitoring dynamic collagen reorganization in human dermis during uniaxial stretching with second/third harmonic generation microscopy
Mengyao Zhou, Yuanyuan Ma, Ludo van Haasterecht, Frank van Mourik, Paul van Zuijlen, Marie Louise Groot
Abstract: Skin is a complex and multi-layered organ, containing epidermis, dermis and hypodermis. The dermis is a flexible layer that can be significantly stretched and bent while resisting pressure, and will return to its original shape when the pressure is relieved. It is the main support of the epidermis, and plays a protective role to prevent the epidermis from rupturing or tearing. We aim to study its mechanical properties in relation to the 3D microstructure. In this work, we use second/third harmonic generation microscopy (SHG/THG) to visualize the collagen reorganization in human dermis under uniaxial stretching, and experimentally investigate the correlation between the mechanical properties of skin and its microstructural characteristics. A Matlab program is used to analyze the orientation of collagen fiber, and correlate it with stress/strain curve at macroscale. The results show that the mechanical properties of the skin are closely related to the orientation of collagen fiber. Under a small strain (λ<1.1), the orientation of collagen does not change, and correspondingly, the stress is almost zero. Then, as strain gradually increases, the collagen fibers will gradually align in the stretch direction, and the corresponding stress will also increase. We also analyzed the relaxation properties of the skin. We will discuss the skin relaxation behavior under different strains. Our research will be helpful for predicting the impact of microstructural changes of skin on macroscopic characteristics, and may be useful for clinical surgery and tissue engineering.
Instant histopathology assessment of fresh lung biopsy quality using third and second harmonic generation microscopy
Laura van Huizen, Kirsten Kalverda, Venerino Poletti, Peter Bonta, Chris Dickhoff, Frank van Mourik, Jouke Annema, Marloes Groot
Abstract: Background - For the diagnosis of interstitial lung diseases (ILD) several lung biopsies are taken. These come with considerable morbidity and risk of sampling error. Immediate feedback on the sample adequacy has the potential to reduce the number of biopsies taken and therefore reducing the likelihood for adverse events, and reduce endoscopy time. THG/SHG/2PEF microscopy - A promising imaging technique for rapid histopathological feedback on lung biopsies is third and second harmonic generation (THG/SHG) and two-photon excited autofluorescence (2PEF) microscopy. This technique is non-invasive, label-free and provides 3D images with a high, sub-cellular resolution, within seconds. In a previous study, we showed that using THG/SHG/2PEF microscopy, we could successfully reveal alveolar structures and histopathology hallmarks of unprocessed lung tissue, including cell morphology and general tissue architecture (collagen and elastin organization) [1]. First results - Here, we used for the first time, to the best of our knowledge, a compact, mobile THG/SHG/2PEF microscope (Flash Pathology B.V.) in the clinic to image fresh transbronchial cryobiopsies. We imaged 55 biopsies of 15 patients, each imaged within a few minutes. Independent lung pathologists evaluated the lung biopsies based on the THG/SHG/2PEF images, using the standard histopathology as gold standard. The first results show that using THG/SHG/2PEF microscopy distinct histopathological ILD features can be identified, such as alveolar structures, fibrosis and inflammation. Further analysis is in progress to determine quantitatively the reliability of the sample quality as assessed by THG/SHG/2PEF microscopy and the potential clinical impact by reduction of the number of extracted biopsies for diagnosis. [1] L.M.G. van Huizen, et al. Translational Biophotonics (2020)
Certainty about uncertainty in sleep staging; A theoretical framework
Hans van Gorp, Iris Huijben, Pedro Fonseca, Ruud van Sloun, Sebastiaan Overeem, Merel van Gilst
Abstract: Sleep stage classification is an important tool for the diagnosis of sleep disorders. Because sleep staging has such a high impact on clinical outcome, it is important that it is done reliably. However, we know that uncertainty exists in both human scorers and machine learning models. On average, agreement between human scorers is only 82.6%. In this manuscript, we provide a theoretical framework to enable discussion and further analysis of uncertainty in sleep staging. To this end, we introduce two variants of uncertainty, known from the machine learning community to sleep staging: aleatoric and epistemic uncertainty. We discuss what these types of uncertainties are, why they are useful, where they come from in sleep staging, and give recommendations on how this framework can improve sleep staging in the future.
Improving ultrasound and photoacoustic volume imaging quality using singular value decomposition
Roy van Hees, Jan-Willem Muller, Frans van de Vosse, Marcel Rutten, Min Wu, Richard Lopata
Abstract: Small animal studies have demonstrated the use of volume imaging and it benefits in both ultrasound (US) and photoacoustics (PA), with some studies demonstrating full body US/PA tomography. The recent introduction of volume imaging systems and matrix arrays by Verasonics has greatly increased scientific interest in non-invasive volumetric US and PA imaging for clinical applications. Unfortunately, image quality is suboptimal, due to limitations in channel count, for a single system. In this research, we propose a volumetric imaging system, exploring the performance of 256 to 1024 channels, connected to a single 3MHz matrix array (Vermon), directly or in a multiplexed fashion. Sparse random aperture compounding is used when multiplexing, and the clocks of the Vantage systems are synchronized using the Verasonics’ Multi-System Synchronization Module. For experiments, a vessel mimicking phantom with three channels and lumen channel in a mock loop set up is employed. A signal processing method is developed, that consists of a block-wise singular value decomposition algorithm to improve volume imaging quality for both US and PA. Experiments were carried out using different multiplexing configurations for the matrix array, as well as different number of steering angles for US imaging, to assess their influence on the generalized contrast to noise ratio (gCNR). In addition, we compare the singular value decomposition algorithm to other SNR improvement techniques, such as simple averaging of PA data and displacement corrected averaging (DCA). The algorithm can successfully improve image quality for both US and PA imaging. An important advantage of the SVD based algorithm compared to other techniques such as averaging and DCA is that the SVD based algorithm can improve image quality for an entire sequence of US and PA acquisitions, as compared to the other techniques that typically reduce a sequence to a single imaging frame. The SVD based algorithm would be useful for a number of applications, such as increasing frame rate using less angles in plane wave imaging, or reducing the number of systems required by applying higher multiplexing ratios, at a minimal cost to image quality. Furthermore, it would potentially allow greater imaging depths.
Trick or treat bacteria and bone cells part ways on nanopatterned surfaces
Khashayar Modaresifar, Mahya Ganjian, Lidy Fratila-Apachitei, Amir Zadpoor
Abstract: Developing designer biomaterials with better control over their biological properties is proven crucial for enhancing the performance of implantable medical devices. For instance, bone implants need to be designed in such a way to promote host tissue regeneration and integration with the implant as well as to combat implant-associated infections, as one of the main causes of implant failure. In this regard, patterning the surface of the biomaterial with small-scale physical features has been proved a promising strategy for inducing both abovementioned functionalities 1, namely bactericidal and osteogenic properties, circumventing complications and challenges associated with the use of antibiotics and growth factors. However, the optimized geometrical parameters of these features (i.e., shape, dimensions, and spatial arrangement) and the underlying mechanisms by which they direct the fate of bacteria and bone cells are not fully understood yet 2,3. Here, we have explored the suitability of a variety of nanofabrication techniques (e.g., EBID, EBL, and ICP RIE) for the rational design of such surface nanopatterns. Furthermore, we investigated the effect of nanopatterns’ dimensions (e.g., interspacing) on their bactericidal activity against Gram-negative and Gram-positive bacteria by different assessment methods such as live/dead staining, CFU counting, and SEM observations. Overstretching of the bacterial cell wall on the nanopatterns and its disruption due to the direct penetration of nanopatterns were shown to be the dominant killing mechanisms of the nanopatterns 4-6. The interactions of the bone-making cells (e.g., MC3T3-E1 and hMSCs) with the nanopatterns were then studied to reveal the connection between the early stage adaptation of the cells to the surface and their osteogenic differentiation. To this end, morphological characteristics of the cells residing on the nanopatterns, formation of focal adhesions (FAs), organization of cytoskeleton, expression of osteogenic markers (e.g., Runx2 and osteopontin), and surface mineralization were evaluated 6. Changes in the cell spreading area and distribution of FAs were found to be the major regulators of the long-term cell response. Moreover, the observed nanopattern-induced osteogenesis was found to be dependent on the intracellular mechanotransduction pathways involving focal adhesion kinase (FAK), Rho-associated protein kinase (ROCK), and Yes-associated protein (YAP). References 1 Anselme, K. et al. The interaction of cells and bacteria with surfaces structured at the nanometre scale. Acta Biomater. 6, 3824-3846 (2010). 2 Modaresifar, K., Azizian, S., Ganjian, M., Fratila-Apachitei, L. E. & Zadpoor, A. A. Bactericidal effects of nanopatterns: A systematic review. Acta Biomater. 83, 29-36 (2019). 3 Dobbenga, S., Fratila-Apachitei, L. E. & Zadpoor, A. A. Nanopattern-induced osteogenic differentiation of stem cells–A systematic review. Acta Biomater. 46, 3-14 (2016). 4 Ganjian, M. et al. Nature Helps: Toward Bioinspired Bactericidal Nanopatterns. Adv. Mater. Interfaces, 1900640 (2019). 5 Modaresifar, K. et al. Deciphering the Roles of Interspace and Controlled Disorder in the Bactericidal Properties of Nanopatterns against Staphylococcus aureus. Nanomaterials 10, 347 (2020). 6 Modaresifar, K. et al. On the Use of Black Ti as a Bone Substituting Biomaterial: Behind the Scenes of Dual‐Functionality. Small, 2100706 (2021).
Combined phase- and amplitude- demodulation for size estimation of microbubbles
Sander Spiekhout, Jason Voorneveld, Guillaume Renaud, Martin Verweij, Nico de Jong, Hans Bosch
Abstract: An acoustical camera (AC) [1] is used to quickly study dynamics of many single microbubbles (MBs), but lacks absolute size information. To overcome this limitation, Fouan et al. [2] suggested a method, based on the time-of-flight difference from scattering of a vibrating sphere, to derive an absolute measure of the radial excursion. Combining this with the relative changes from amplitude demodulation, the resting radius, and thus, the complete radial dynamics can be measured with an AC. We optimized this method, obtained size distribution estimates from two MB populations of different sizes, and compared those with Coulter counter measurements. The AC setup consists of a pentagon-shaped watertank with three transducers of overlapping foci, which is filled with a heavily-diluted monodisperse MB solution. A low-frequency transducer drives a single MB with a sequence of 5 frequencies from 2.5 down to 1.0 MHz at 20 kPa, while the first high-frequency (HF) transducer sends a probing wave that is linearly scattered by the MB with its instantaneous radius. The second HF transducer receives a HF signal modulated in amplitude and phase due to the radial changes. By amplitude demodulating the HF scattered signal, the relative change dR/R0 is obtained. From phase-demodulation, by multiplying the phase change by the probing wavelength, the absolute change dR is obtained. Dividing the magnitudes of both demodulated signals gives the resting radius R0. The MB should vibrate sufficiently (dR/R0≥ 5%) to obtain reliable signals. The sizing method was tested on two monodisperse MB populations with a mean radius of 2.1µm and 3.5µm, showing sufficient vibration in 394/468 and 288/321 MB measurements for the small and large MB respectively. In both cases, the size distribution underestimates the Coulter counter distribution by 0.25 µm (12% bias for 2.1µm, 7% bias for 3.5µm), and shows a 50% higher FWHM. The shape of the distribution matches considerably better for the larger MBs, probably owing to stronger phase signals with increasing size. The results indicate the full microsecond-scale MB dynamics can be captured with a purely acoustic device, which provides a more scalable alternative to ultra-high framerate imaging [1] for studying single MB dynamics.
Polymer electrode arrays for cochlear implants
Alberto Miralles-Abete
Abstract: Present day cochlear implants use platinum electrodes incorporated in hand-made probes. There is a clear need to develop probes that can be manufactures using techniques developed in the silicon industry. An all-polymer probes has great advantages. This requires polymers which are conductive and insulating. An excellent candidate is PEDOT:PPS. This is normally a conductive polymer. However, it can be made insulating selectively by a number of methods. This work is focussed on the PEDOT:PPS material and its properties. Experiments have shown that, as deposited, the PEDOT:PPS has a conductivity of 230 Scm-1. This is sufficient for the electrode. PEDOT:PPS can also be made insulating. In this work looks at using UV radiation to turn the conductive polymer into an insulating polymer. The advantage is that we can use standard lithography to make selected areas insulating. For this we require a UV source with wavelength below 315nm. Above this wavelength the reduction of conductivity will be slow and not reach a good insulating layer. Using a UV light source with power 2000mWcm-2. the conductivity was reduced from 230 Scm-1 to 0.48 Scm-1 within 3 hours. This is sufficient to use PEDOT:PPS as the electrode and the surrounding insulating layer. The next issue is charge injection capacity. This is extremely important if we want to reduce the size of the electrodes in cochlear implants. If the electrodes are made too small, they are unable to inject sufficient current to stimulate the nerves. Experiments have shown that PEDOT:PPS has a charge injection capacity of 384Ccm-2, compared to 25Ccm-2 for platinum, an increase of a factor of 15. This means that the electrode can be significantly reduced in size, while maintaining sufficient current. This will allow us to increase the number of electrode and retain sufficient charge injection. This abstract presents method of treating the PEDOT:PSS to achieve electrodes and insulating layers for cochlear implants.
Burn segmentation on digital and laser Doppler imaging using deep learning
Vanja Miskovic, Thomas Rose, Carolina Varon, Carlo Saverio Iorio
Abstract: Aims: Burn assessment requires two parameters: the percentage of total body surface area and burn depth. This is done by a burn specialist, often supported by laser doppler imaging (LDI). Based on this assessment, specialists form the basis for the following treatment. Unfortunately, most burns occur in remote and resource-challenged areas that lack specialists and relevant diagnostic instruments. A computer-based automatic diagnosis could support the decision-making process for the non-specialist and predict the wound healing path. The first step in creating such a system is to achieve fully automatic detection (i.e., segmentation) of wound areas in natural images. This work proposes a framework for digital and LDI image segmentation using deep learning. Methods: The burn unit from the Queen Astrid Military hospital provided 546 digital and their corresponding 524 LDI burn images, with a resolution of 270x250 pixels. All the burns were manually annotated, using ImageJ software, which created a mask for each image. A convolutional neural network, U-net, specially designed for image semantic segmentation was trained on the digital and LDI images separately and their corresponding masks. The network was trained using 80 % (419) of the images and 10-fold cross-validation, and the remaining 20 % (127) were used as independent test set. The binary cross-entropy was used as a loss function with Adam optimizer. The efficacy of the image segmentation was evaluated using the Dice coefficient. Results: The Dice coefficients for both datasets were 65.16 % for digital and 69 % for the LDI dataset. Conclusion: Results suggest an over-segmentation, which is more evident for digital images in case of 2nd-degree burns, and for LDI images with the large blue region, color that on LDI images indicates both healthy skin and deep burns. This could be tackled by means of data augmentation and transfer learning.
Data loss and clinical interpretation of continue glucose measurements
Carlijn Braem, Niala Braber, Utku Yavuz, Goos Laverman, Miriam Vollenbroek, Hermie Hermens, Peter Veltink
Abstract: Introduction Continuous glucose monitoring (CGM) is an essential to improve glucose management for diabetic patients. Clinical parameters from CGM data include: the time in range (TIR, range 3.9-10.0 mmol/L), the time below range (TBR), and the time above range (TAR). The clinical target is to strive for >70% TIR with <4% TBR [1]. Data loss due to human, software, or hardware errors must be considered. Current literature arbitrarily recommend <30% data loss in a 14-day CGM recording [1]. Our aim is to test the influence of missing data, following the current guideline, on CGM parameters and the subsequent clinical decision making. Method From the Diabetes and Lifestyle Cohort Twente (DIALECT)-2 cohort 140 CGM recordings were used. Where with the FreeStyle Libre® sensor, glucose was monitored in type 2 diabetes patients for two weeks [2], [3]. Glucose parameters were compared between CGM data without data loss and the same CGM data with 30% artificial data loss. To this end, a dataset without missing data was generated, merging data from different subjects with the closest CGM charactheristics based on principal component analysis. Later, 30% artificial data loss was generated by simulating data loss with missing characteristics of the original recordings. The TIR, TBR, and TAR were computed for both data sets and the outcomes were compared with the mean absolute error (MAE). Results On average, the original CGM recording was 12.8±1.8 days long, with 11.8% [0-77%] data loss. Of 140 subjects, 9 subjects had more than 30% data loss. We generated 30 combined CGM recordings without missing data. The MAE between the data set with and without missing data was 1.20%, 0.49%, and 1.27% for TIR, TBR, TAR, respectively. In 4 out of 30 combined recordings, data loss resulted in change of meeting the <4% TBR target. Conclusions Although 30% of missing data results in small errors, it can alter whether patients meet TBR targets. Consequently clinical decisions could be altered by missing data. Therefore, current recommendations on missing CGM data may not be adequate. In future research, data imputation will be investigated to handle missing data while minimizing CGM outcome changes. References [1] T. Battelino et al., Diabetes Care, 2019, doi: 10.2337/dci19-0028. [2] C. M. Gant et al., Nutrients, 2017, doi: 10.3390/nu9070709. [3] N. den Braber et al., Nutrients, 2019, doi: 10.3390/nu11020409.
Study of linear and nonlinear cardiorespiratory interactions during sleep using transfer entropy
Andrea Rozo, John Morales, Jonathan Moeyersons, Rohan Joshi, Enrico G. Caiani, Carlo Iorio, Pascal Borzée, Bertien Buyse, Dries Testelmans, Sabine Van Huffel, Carolina Varon
Abstract: Aims: Changes in the dynamics of cardiac and respiratory signals, as well as in their interactions, have been observed in the sleep-wake cycle. The magnitude of the linear and nonlinear components of these interactions varies with the sleep stage, but most of the current methods for quantifying cardiorespiratory interactions do not differentiate them. In this study, one of the methods with the potential to quantify both components, namely transfer entropy (TE), is applied to identify their presence during light and deep sleep. Methods: Heart rate variability (HRV) and respiration (nasal airflow, NAS, and effort around the thorax, THO) signals recorded from 26 subjects participating in a sleep apnea study were used. The signals were extracted for light and deep sleep, divided into one-minute segments and resampled at 2 and 4 Hz. TE was computed between each respiration signal (NAS and THO) and HRV for each sleep stage, using 5 time lags. The significance of each component of the cardiorespiratory interaction computed with TE was validated through a surrogate data analysis. Differences between sleep stages were evaluated with the median TE of the significant segments by patient, using a Wilcoxon signed rank test. Results: A significant (p < 0.05) increment in linear and nonlinear interactions during deep sleep was found for different time lags. However, these differences were dependent on the type of respiratory signal and sampling frequency. Conclusions: The results suggest that there might be a higher influence of nonlinear interactions during deep sleep than during light sleep, at specific time lags. The change of magnitude of the TE with the time lag is in line with studies that have found that the intensity of the cardiorespiratory interactions varies at different times scales during sleep. Moreover, the results showed that the estimation of cardiorespiratory interactions is also affected by the sampling frequency. This highlights the importance of selecting appropriate signals and sampling frequencies for the study of linear and nonlinear cardiorespiratory interactions.
Smart module for removal of protein bound uremic toxins for artificial kidneys
Jeroen Vollenbroek, Fokko Wieringa, Karin Gerritsen, Joachim Jankowski, Lucas Lindeboom, Rosalinde Masereeuw,, Leonard van Schelven
Abstract: Most End Stage Kidney Disease (ESKD) patients rely on hemodialysis to remove toxins from their blood. During hemodialysis, toxins smaller than 20-45 kDa are filtered out of the bloodstream into the dialysate. Some small toxins, the so-called Protein Bound Uremic Toxins (PBUTs), are difficult to remove since they are bound to large proteins (e.g. albumin ~66 kDa) [1]. Whereas healthy kidneys can actively and very efficiently secrete these PBUTs, hemodialysis only removes the unbound fraction [2]. PBUTs-accumulation is implicated with injury to the heart, blood vessels, brain and nerves. Simply increasing filter pore size is not an option since the binding blood proteins are vital and should not be removed [1]. The Multi-compatible Implantable Toxin Removal Augmentation Module (MI-TRAM, KidneyX prize recently granted), currently being optimized by a multidisciplinary team, may greatly enhance PBUT-removal by loosening PBUTs from blood proteins, thereby increasing the free fraction that will be filtered out. The consortium partners are: RWTH Aachen, UMC Utrecht, Utrecht University, and IMEC. High strength, high frequency electromagnetic fields will be used to shake the electrostatic bonds between blood proteins and PBUTs [3-5]. In vitro experiments, performed by the group of prof. Jankowski, already demonstrated significantly improved PBUTs removal, but using bulky setups that are not compatible with wearable, portable or implantable artificial kidney systems [3]. The MI-TRAM chip, developed by IMEC, is a miniaturized version of this system that enables touchless ‘through-the-tubing’ capacitive coupling, for which a simple clip-on connection suffices. Furthermore, the MI-TRAM chip has integrated sensors to monitor hematocrit level, total body water and body temperature. In the near future, we will perform in vitro dialysis experiments with the MI-TRAM chip for performance testing, calibration, optimization of field strength and frequencies. Both conventional and novel dialysis membranes will be tested for PBUT removal efficiency enhancement using MI-TRAM. After successful in vitro testing, we will proceed with in vivo efficacy and safety experiments in a uremic large animal model [6], followed by clinical trials. MI-TRAM has the potential for a rapid clinical introduction of this technology to the benefit of ESKD patients.
Discriminative machine learning analysis of skin microbiome as biomarker in patients with seborrheic dermatitis and acne vulgaris
Hein van der Wall, Robert-Jan Doll, Adam Cohen, Gerard van Westen, Jacobus Burggraaf, Robert Rissmann, Tessa van der Kolk, Hans Pickaers, Martijn van Doorn
Abstract: In recent years the skin microbiome has become of high interest as a drug target and as a disease biomarker. Computational (machine learning) analyses of microbiome sequencing data from patients with skin disorders, (e.g., seborrheic dermatitis and acne vulgaris), could be helpful to identify discriminative disease biomarkers in their microbiome profile. The aim of the present study was two-fold, 1) to employ machine learning techniques to discriminate patients with acne vulgaris from seborrheic dermatitis based on their microbiome profile, and 2) to identify discriminative disease biomarkers in the microbiome of these patients. Data were collected from two different patient groups studied at a single centre in the Netherlands. The sample consisted of 30 patients with acne vulgaris and 37 patients with seborrheic dermatitis. In both groups, samples were taken from both lesional skin and non-lesional facial skin, resulting in microbiome profiles of 4 different skin categories. A random forest (RF) model was trained to distinguish between the skin microbiome of patients with acne vulgaris and seborrheic dermatitis versus the other profiles. Subsequently, the most important microorganisms for discrimination of the two skin diseases were determined by means of feature analysis and SHapley Additive exPlanations (SHAP) values. The trained RF classification model for discrimination between seborrheic dermatitis lesional skin and other skin categories, showed an accuracy of 83% and an ROC-AUC of 88%. For the comparison between by acne vulgaris and other skin categories there was 73% accuracy and and an ROC-AUC of 71%. For both conditions, the most important discriminative microorganisms, apart from Staphylococcus genus, had a relatively low occurrence, such as the Rubrobacter genus. Our preliminary results show that differentiation of skin conditions can be made by machine learning using the microbiome data, and can be of major importance in the development and application of new individualized therapies, involving modifications of the microbiome, as part of the next generation of ‘ecobiological’ anti-inflammatory treatments.
Validating the use of cerebral oxygenation (NIRS) to monitor blood pressure responses upon standing in older adults
Marjolein Klop, Rianne de Heus, Jurgen Claassen, Carel Meskers, Andrea Maier, Richard van Wezel
Abstract: Background: Orthostatic hypotension (OH) is common among older adults, with a prevalence ranging from 22% in community-dwelling older adults to 60% in geriatric inpatients. OH is classically defined as a blood pressure (BP) drop upon standing of at least 20 mmHg systolic and/or 10 mmHg diastolic, within the first 3 minutes after standing. OH can cause symptoms like dizziness, is an important cause of falls, and has been associated with low physical function, lower cognitive function, dementia, cardiovascular disease and mortality. Currently, diagnosis of OH is based on intermittent BP measurements using a sphygmomanometer. This method does not capture the full BP response, has poor reproducibility, and is not representative for OH in daily life. Near-infrared spectroscopy (NIRS) may be used as a proxy. NIRS uses near-infrared light to detect changes in oxygenated and deoxygenated haemoglobin in cerebral frontal lobe tissue. This non-invasive technique can potentially be used to measure continuously for a longer duration under daily life conditions. Aim: To validate cerebral oxygenation measured with NIRS as a proxy for BP changes during and after standing up. Methods: Cross-sectional study, including a population having both normal and impaired BP responses upon standing: 10 young (18-35 years) and 30 older (>65 years) adults, of whom at least 30% were geriatric outpatients with or without OH. Our protocol consisted of fast supine-stand and sit-stand transitions (5 minutes supine/sitting, 3 minutes standing); only for young healthy participants complemented with squat-stand manoeuvres (1 minute squat, 3 minutes stand). During these postural challenges, BP (by volume-clamp photoplethysmography: Finapres) and cerebral oxygenation were measured continuously. Correlations between BP and cerebral oxygenation were calculated. Results: Preliminary results show drops in both BP and cerebral oxygenation when standing up, during all challenges. Full results will be available at the time of the conference. Significance: Results of this study will indicate whether NIRS can be used as a proxy for OH and related outcome. This will allow for studying consequences of OH in older persons in their daily life setting, for example at home. Ultimately, this can contribute to an improved OH diagnosis and treatment.
Energy efficiency of non-rectangular stimulation pulses in biophysically realistic neuron models
Francesc Varkevisser, Tiago L. Costa, Wouter A. Serdijn
Abstract: The development of brain-computer interfaces and neuroprosthetic devices that interact with the cerebral cortex require implantable stimulators with hundreds to thousands of output channels. These devices have a limited power budget due to the wireless power link. Therefore, optimized power efficiency is crucial to facilitate as many channels as possible. Conventional electrical stimulation methods are not power-efficient, especially in multi-channel configuration. In the optimization process, the shape of the output waveform is often considered to be rectangular. However, previous work has shown that the conventional rectangular shape is not energy optimal. Hence, the effect of shape properties on the required activation energy is an essential parameter in optimizing stimulator power efficiency. This work explores the potential benefits of using non-rectangular pulses in cortical stimulation for reducing the required activation energy. For this, we use biophysically realistic single-cell models of cortical neurons in the NEURON v8.0 simulation software. We apply monophasic extracellular stimulation pulses to these models using a point-source electrode for different configurations changing electrode location, pulse shape, and pulse duration. The changes in activation current and energy are analyzed and related to the biophysical properties of the activated areas. The models show a decrease in activation energy with respect to rectangular pulses for half-sine (8.5 ±1.6%, mean ± std), centered triangular (9.3 ±2.1%), and Gaussian (8.7 ±3.0%) shaped pulses in the region of interest. Moreover, by comparing shape-specific strength-duration and energy-duration curves, it is shown that the chronaxie time and energy-optimal pulse duration depend on the temporal properties of the stimulus. This could be related to the matching between pulse and membrane dynamics. The outcomes of this work can be used as a toolbox in the design of power-efficient output stages for high-density multi-channel electrical stimulation. Considering the biophysical benefits of non-rectangular pulses will lead to well-founded novel circuit designs. Future work includes analysis of more complex configurations and in-vivo validation of the results.
Tensor-based deconvolution of the Hemodynamic Response in the mouse visual pathway using functional ultrasound
Aybüke Erol, Chagajeg Soloukey, Sebastiaan Koekkoek, Pieter Kruizinga, Borbála Hunyadi
Abstract: Functional ultrasound (fUS) is an emerging neuroimaging modality that records changes in cerebral blood volume and flow in response to neural activation with high spatiotemporal resolution. The dynamics of the acquired hemodynamic response (HR) signal depends on the unknown neural activity and the region-specific hemodynamic response function (HRF). An accurate estimation of the HRF is crucial to correctly interpret both the hemodynamic activity itself and the actual neural signals involved in evoking the neurovascular dynamics. Furthermore, the HRF has shown potential as a biomarker for pathological brain functioning [1]. In order to estimate both of these aspects of brain function simultaneously, we employ a multiple input-multiple output strategy and model the fUS signal as a convolutive mixture of underlying neural sources. First, we examine the validity of the linearity assumption that constitutes our modelling approach. To this end, we investigate the HR under various stimulus durations. Our results show that up to 4 seconds the HR changes linearly with respect to the stimulus duration, however, above 4 seconds it reaches a plateau with high frequency fluctuations before returning back to baseline. In order to get a strong enough response while avoiding such nonlinearities, we opt for a constant stimulus duration of 4 seconds (repeated multiple times) in a subsequent experiment for a reliable deconvolution. For deconvolution, we apply block term decomposition (BTD) on the tensor of lagged output autocorrelation matrices. The BTD approach is based on two assumptions: (i) HRFs are parametrizable, and (ii) source signals leading to fUS time series are uncorrelated. We focus on the HRs of three main regions of interest in the mouse colliculo-cortical, image-forming pathway, namely the lateral geniculate nucleus (LGN), superior colliculus (SC) and visual cortex (V1). When the estimated HRFs of these regions are compared, our results demonstrate that SC exhibits the narrowest and fastest response, whereas V1 has the latest and most rounded response. Last but not least, we show that our source signal estimation can be used to trace back the timings of the experimental paradigm. [1] A. R. Mayer et al. (2014). Investigating the properties of the hemodynamic response functionafter mild traumatic brain injury. Journal of Neurotrauma, 31(2), 189–197.
A CMOS integrated circuit to interface with a 2D array of ultrasound transducers for ultrasound brain stimulation
Hassan Rivandi, Tiago Lopes Marta da Costa
Abstract: A minimally invasive and precise neuromodulation modality is emerging by means of low intensity focused ultrasound (LIFU). However, existing ultrasound stimulation devices suffer from having bulky form factor and poor precision. This work aims to discover ways to design miniaturized ultrasound transducers and interfacing electronics towards 10x increase in precision while remaining minimally invasive. This demands the following three characteristics: firstly, a device miniaturized to the millimetre scale. This enables a form-factor compatible with in vitro electrophysiology and optical setups, and with wearable stimulators for in vivo studies with mice and rat models. Secondly, it includes electronics to implement a 2D phased array pitch-matched transmit beamformer, or 3D digital lens. This enables software controlled selection of the stimulation location in real time, emulating implantable multi-electrode arrays used in electrical stimulation. Finally, it generates ultrasound focal spots with volumetric spatial resolution below 500µm3 and with spatial peak pulse average intensity (ISPTA) above 1W/cm2, for successful neuromodulation. This will approach the volumetric spatial resolution achieved with implantable electrical stimulators without requiring surgical implantation in the brain. To obtain the optimal design described above, this works presents a design flow based on simulations with using Matlab’s acoustics toolbox (k-Wave) and OnScale multiphysics, and presents the first generation of driving circuits for a 2D phased array pitch-matched transmit beamformer. The simulation results show that a 64x64 array of ultrasound transducers with a pitch size of 50µm, corresponding to stimulation frequency of 15MHz, is needed to generate the required intensity with the desired volumetric resolution of less than 500µm3. The simulation results on OnScale software shows that the electronic circuit should drive piezoelectric ultrasound transducers with amplitude of 20V to achieve ISPTA above 1W/cm2. Furthermore, two different driving circuits were designed using TSMC 180 nm BCD technology, which are able to drive piezoelectric ultrasound transducers with 20V and 36V, respectively, while occupying an area of only 50x50µm2. These findings will pave the way for a full 64x64 2D phased array transducer and for the direct integration of piezoelectric transducers on top of the CMOS chip, towards in vitro and in vivo validation.
A brain-on-chip platform to study the optimal parameters of focused ultrasound neuromodulation
Gandhika Wardhana, Tiago Costa, Massimo Mastrangeli, Wouter Serdijn
Abstract: Brain stimulation techniques are essential tools to address various neurological conditions. Currently, approved neuromodulation techniques are either invasive, requiring patients to undergo risky surgery, or non-invasive but suffering from low spatial resolution. Focused ultrasound stimulation (FUS) is a promising modality for non-invasive brain stimulation which was shown capable of eliciting and suppressing neural activity [1]. In FUS, stimulation is delivered deep into the brain using acoustic waves with sub-millimeter spatial resolution, which is comparable to the resolution of invasive modalities [2]. However, the role of the modulation parameters in FUS is still heavily debated due to the limited performance of commercial FUS transducers and their inadequate pairing with typical in vivo neuronal recording technologies. Differences in experimental setups, in parameters used to characterize the ultrasound, and in parameter estimation have led to conflicting conclusions, and make prior studies difficult to compare [3]. There is therefore the need for a platform that can reliably correlate ultrasound stimulation parameters and their effect on neuronal tissues. Organ-on-chip (OoC), specifically brain-on-chip (BoC), is an emerging technology which may address this need. OoC utilizes microfabrication techniques to create devices that mimic human physiology in vitro in an organ-specific context to help develop drugs and treatments [4]. In this context, we aim at incorporating three-dimensional tissue constructs (such as organoids and neural networks) on a micro-electrode array (MEA) to monitor neuronal electrical activity in the presence of ultrasound stimuli. One major technological challenge to develop such BoC platform lies in the ultrasound transducer. To understand the effect of ultrasound on neuronal substrates, the ultrasound transducer has to be able to deliver focused stimulation with single-neuron resolution. This work presents the development of an ultrasonic piezoelectric transducer compatible with a BoC platform on a wafer-scale. Finite element modeling was used to explore different ways to improve the spatial resolution and acoustic pressure of the transducer. These findings were incorporated into the design and fabrication of the transducer and BoC platform using wafer-level micromachining techniques. While still a work in progress, the resulting platform is expected to push forward the understanding of the biophysical mechanisms of ultrasound neuromodulation.
The influence of extracerebral tissue on near-infrared spectroscopy in adults - A systematic review of in vivo studies
Nick Eleveld
Abstract: Background: Extracerebral tissue (scalp, skull) influences near-infrared spectroscopy (NIRS) measurements, but there is currently no consensus on the extent to which this is the case. The literature to date is synthesised in this systematic review. Methods: Four major databases were searched for original studies that investigated the influence of extracerebral tissue on continuous-wave or frequency domain NIRS with an in vivo study protocol in adult humans or large animals. Included NIRS-indices were oxygenated (OxyHb), deoxygenated (HHb), or total hemoglobin (tHB), or regional cerebral oxygenation indices. Only studies that performed a selective alteration of the intra- or extracerebral perfusion or that used reference techniques for both intra- and extracerebral perfusion were included. Critical appraisal was performed with a modified version of the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tool. Results: Seventy-nine studies were included. Thirty-eight studies investigated oxygenation indices and 41 looked into the concentrations of OxyHb, HHb, or tHB. Fifty-eight studies used a reference technique for extracerebral perfusion, the majority employing Laser Doppler Flowmetry (LDF) on the scalp. Equally, 58 studies employed an intracerebral perfusion reference technique, mostly Transcranial Doppler or functional Magnetic Resonance Imaging. Forty-two studies used both and seven employed no reference technique. Selective alteration of intracerebral perfusion (20 studies) consisted mainly of hyper-/hypocapnia protocols and clamping of the internal carotid artery during carotid endarterectomy. Extracerebral (24) alteration methods included cuff inflation on the forehead and external carotid artery clamping. Fifteen studies used both methods. Overall study quality was low. For each QUADAS-2 domain, over 75% of studies had a high risk of bias. Only one study had a low risk of bias on all QUADAS-2 domains. Preliminary synthesis: A subset of studies quantitatively expressed the extracerebral influence on NIRS as a correlation coefficient (R2). R2 between changes in intracerebral reference technique and NIRS varied strongly between studies, ranging from non-significant to 0.8. For the extracerebral reference, R2 with NIRS varied between non-significant and 0.5. Preliminary conclusion: The large variance between studies indicates no consensus on extracerebral influence on NIRS, for
Unobtrusive wearable sensing to estimate human circadian process
Nemanja Cabrilo, Charikleia Papatsimpa, Jean-Paul Linnartz
Abstract: The central biological clock in the brain has a near-24h rhythmicity that is a main determinant of individuals’ sleep/wake cycles. It also orchestrates the daily rhythms of hormonal secretion and behaviour such as subjective alertness and performance[1]. Having a solution to determine and track the state of an individual’s internal circadian state can allow us to make major steps to optimize human daily rhythms, from the use of light for improving sleep quality to precision medicine, diagnosing neurological disorders and optimizing drug delivery. Some estimates[2] of the circadian state demonstrated that, even if not up to clinical standards, can greatly enhance human-centric lighting to improve wellbeing of, for example, office, industry or warehouse workers. In a quest to investigate the use of wearable sensing for tracking the underlying circadian process, we developed a model that combines a physiology-based model of the human biological clock with non-invasive but possibly in-accurate ambulatory data (in particular actigraphy data) in a statistical framework[3]. Nowadays we are conducting a field study to validate our model in a real-life setting and to test to what extent this leads to meaningful estimates. We compare model-based predictions based on wearable data (CamnTech MotionWatch8) against the “gold standard” circadian state estimation, namely, individual subject’s bathyphase, or timing of the daily Core Body Temperature (CBT) nadir (minimum point) which is a widely accepted circadian phase marker[4]. CBT is continuously being recorded with BodyCAP e-Celsius Performance ingestible e-capsule(s). Beyond, we aim to assess further physiological signals and identify the best possible predictors, with reasonable user comfortability, as non-invasive circadian biomarkers, for instance: Skin temperature (DSL1922L temperature logger), Heart-Rate Variability (Movisense EcgMove4), Electro-Dermal Activity (Empatica E4 wristband). Accurate and unobtrusive estimation of the exact circadian phase can unlock the potential of numerous applications, possibly including personalized Human-Centric Lighting. Our measurement set-up is unique in its versatility of different sensor modalities and allows a cross-benchmarking. As measurements are being collected at the moment of writing this abstract, the conference paper will primarily address design consideration for the tests, discuss variability of measurements and sensor imperfections combined with early experiences and technical findings.
Edge enhancement applied in ent-endoscopic systems
Geert Geleijnse, Bernd Rieger
Abstract: Flexible endoscopes are essential to examine nose, throat and upper airway [1]. Digital endoscopes are connected to a video processor which can apply various operations to enhance the image without perceivable delay for the observer. Kawaida et al. reported a study on digital image processing of laryngeal lesions by electronic video endoscopy [2]. Although the nature or technique of the applied enhancement is not described, judging the presented images, it seems that they introduced some level of edge enhancement. Edge enhancement is a known technique to sharpen edges by adding an undershoot on the darker side of an edge and an overshoot on the brighter side. A major drawback of edge enhancement is that the operation cannot discern edges from noise and therefore enhances both. Today, edge enhancement is commonly applied in all types of endoscopes e.g., Otolaryngology, Urology, Gastroenterology and operating room, but the specific method and parameters are not disclosed (by the vendors) and literature to substantiate the default settings has not been found either. Some vendors give users the option to manually tune the degree of edge enhancement. Our hypothesis is that there exists an optimum for the level of edge enhancement with respect to the human interpretation. The absence or minor levels of edge enhancement yield images that are perceived as vague by Ear, Nose, Throat (ENT)-professionals. Excessive levels of edge enhancement, however, yield sharp images but, contain objectionable artefacts and too much noise. The purpose of this study in progress is to: (1) Objectively quantify the level of edge enhancement using a method that can be applied to endoscopic systems with undisclosed methods of edge enhancement. (2) Measure the effect of edge enhancement on sharpness and visual noise. (3) Find the subjectively perceived optimal level of edge enhancement. Edge enhancement is studied in four types of flexible digital ENT endoscopes. Objective measurements are performed using the Rez Checker Target Nano Matte [3] and subjectively perceived image quality will be studied using in vivo laryngeal images that are pairwise compared by ENT professionals. References [1] B. C. Paul, S. Chen, S. Sridharan, Y. Fang, M. R. Amin and R. C. Branski, “Diagnostic accuracy of history, laryngoscopy, and stroboscopy,” The Laryngoscope, vol. 123, no. 1, pp. 215-219, 2012. [2] M. Kawaida, H. Fukuda and N. Kohno, “Digital image processing of laryngeal lesions by electronic videoendoscopy,” Laryngoscope, vol. 112, no. 3, pp. 559-64, 2002. [3] G. Geleijnse, M. M. Hakkesteegt, G. J. de Groot and M. R. Metselaar, “Measuring Image Quality of ENT Chip-on-tip Endoscopes,” Journal of Imaging Science and Technology, vol. 65, no. 2, 2021.
A single cell resolution bi-directional neural interface for an artificial retina
Yi-han Ou-yang, Dante Gabriel Muratore
Abstract: Retinal prostheses have been proposed recently as a promising method to restore partial visual sensation in patients with degenerative diseases by stimulating the remaining healthy neurons in the retina [1-2]. However, current devices do not provide sufficient results for patients to become fully independent. A fundamental problem is that current stimulation strategies fail to respect the different retinal cell types that encode different scene information. Instead, they activate them indistinctively, sending a scrambled message to the brain. An artificial retina proposed in [3] is under development and aims at restoring vision with more precise control of natural neural code in the retina. This work focuses on an implantable chip that is clinically viable, fully wireless and has bi-directional capabilities when interfacing with neurons. It operates in three modes: cell calibration (to record the spontaneous activity and learn which cells and cell types are available to the device), dictionary calibration (to learn individual cell responses to the stimulation parameters), and runtime (to stimulate available cells optimally based on a cell-type specific dictionary). The envisioned chip consists of three major building blocks: 1) recording channels that can capture spike activities over a massively parallel microelectrode array; 2) stimulation channels that can activate neurons with single-cell resolution; 3) wireless power and data telemetry circuits. This work focuses on developing the wireless power and data system for the proposed implant. Power is delivered through a 2-coil transcutaneous inductive link with a 14-mm coil separation distance. With a carrier frequency of 40.68 MHz, this design targets a downlink (air to implant) transmission rate at 200 kbps using amplitude-shift keying modulation. The uplink (implant to air) transmits 20 Mbps data through an on-chip antenna. Compared to other state-of-the-art [1-2], this work features a high-speed uplink for closed-loop neuromodulation based on the dictionary approach described above. The prototype chip is under development using a 180-nm standard CMOS process technology. The complete system diagram with specific design considerations for wireless circuits will be presented during the conference.
Understanding the contribution of stretch-activated ion channels to cardiac arrhythmogenesis using computational modelling
Melania Buonocunto, Aurore Lyon, Tammo Delhaas, Jordi Heijman, Joost Lumens
Abstract: Cardiac electrophysiology and mechanics are strongly interconnected. Their interaction is mediated by cardiac mechano-electric feedback through stretch-activated ion-channels (SACs). These channels are also thought to contribute to the development of arrhythmias, but their precise role remains unclear. We aim to understand the contribution of SACs to arrhythmias using a novel computational model of cardiac electromechanics. The dynamics of cardiac electromechanics are modelled by a non-linear system of ordinary differential equations (ODEs). More specifically, we implemented two ODEs describing SACs in the existing O’Hara-Rudy model (ORd) (1) of human ventricular electrophysiology. We modelled one potassium-selective SAC and a non-selective SAC (conducting sodium and potassium). The model was calibrated based on experimental human and rodent data of cardiomyocytes undergoing stretch using the Nelder-Mead algorithm. Subsequently, we varied the amplitude, duration, and timing of the applied stretch to investigate their effects on action potential (AP) and calcium transient properties. The model reproduced APs measured experimentally. Early afterdepolarizations, delayed afterdepolarizations, and ectopic beats were observed when applying stretch with short duration (e.g. 20ms) and high amplitude (e.g. 40%). When varying the time of application, stretch closer to the subsequent beat (beat duration: 1000ms) also shortened the following AP duration (APD shortening of 30ms at t=600ms, 130ms at t=800ms). Milder effects (no APD shortening with stretch at t=600ms) were seen with a lower stretch amplitude (e.g. 15%). However, failure of repolarization occurred when 15% stretch was applied for a longer duration (1000ms). Applying stretch (e.g. 20ms, 40%, t=600ms) also affected the intracellular calcium concentration by increasing diastolic calcium concentration (control: 108nM vs. stretched: 352nM) and decreasing systolic calcium concentration (control: 369nM vs. stretched: 248nM). Using a novel human electromechanical computational model, we quantified the contribution of SACs to cardiac AP and calcium changes. We showed that variations of amplitude, timing and duration of cardiomyocyte stretch had different effects on cardiac electrophysiology. We also showed that SACs may lead to afterdepolarizations, and can shorten the subsequent AP duration, factors which may potentially contribute to the generation of arrhythmias.
Design of a regenerative ventricular assist device
Koen L.P.M. Janssens, Maaike Kraamer, Frans N. van de Vosse, Peter H.M. Bovendeerd
Abstract: Implantable left ventricular assist devices (LVAD) can provide a significant survival benefit and reduction in mortality for patients presenting with end-stage heart failure. However, current generation LVADs do not provide a long-term solution as adverse events or complications may lead to poor prognosis. Additionally, LVADs are unable to adapt to the patient’s varying physiological demands. Within the BRAVE consortium we aim to develop a personalized regenerative biological ventricular assist device (BioVAD) to provide long-term functional support to the infarcted heart. Other regenerative approaches have been proposed though none take into account the complex 3D geometry and structure of the heart. Our current design includes a 3D-printed scaffold, seeded with human induced pluripotent stem cells and matured in a bioreactor in order to replicate the properties of healthy myocardium. This cultured patch of artificial tissue is disposed over the injured area and contracts simultaneously with the heart to restore the pump function in a more natural manner. Within our group, we employ finite element models from [1] in an effort to identify key design criteria for optimal device function and personalization. We simulated both acute and chronic cardiac infarction by locally restricting contraction and increasing passive stiffness respectively. In both cases, relative loss in pump function exceeded loss in the amount of healthy tissue about two-fold, due to a change in mechanical load in the infarct boundary region. These analyses provided functional requirements for the BioVAD and at the same time elucidated unfavourable mechanical interactions to be avoided in the BioVAD design. In particular, different material properties for the central contractile region and the peripheral attachment region are required as poor choices may compromise device function or even the left ventricle itself. [1] Bovendeerd P.H.M., W. Kroon, T. Delhaas, Determinants of left ventricular shear strain, Am J Physiol 297:H1058–H1068, 2009.
Detection of acute cardiac ischemia by a novel portable electrocardiography device with four precordial electrodes in an animal model
Nynke M. de Vries, Alejandra Zepeda-Echavarria, Rutger R. van de Leur, Vera Loen, Marc A. Vos, Thierry X. Wildbergh, Joris E.N. Jaspers, Rien van der Zee, Pieter A. Doevendans, René van Es
Abstract: Background Home-based electrocardiography (ECG) has the potential to reduce time to treatment for patients suffering acute cardiac ischemia, lowering the morbidity and mortality of disease1. However, currently available portable ECG devices are designed and validated for rhythm disorders mostly and can therefore not be used for the detection of cardiac ischemia2. In the UMC Utrecht, a smartphone-sized ECG recording device called miniECG was developed to detect the full spectrum ECG abnormalities. This device uses four precordial electrodes to acquire a multi-lead ECG, without the need for additional equipment or a trained professional. Aim The aim of this study was to investigate the detection of ischemic ECG changes by the miniECG. Methods In eight pigs (64±1 kg), antero-septal myocardial infarction was induced by 75 minute occlusion of the left anterior descending artery, after the first or second diagonal, using a 3mm balloon catheter. A MiniECG recording was acquired every minute for the first 10 minutes post-occlusion and every 5 minutes until 40 minutes post-reperfusion. Gold standard 12-lead ECG was recorded continuously during the experiments. Ischemia related ECG features including ST-deviation and rhythm disorders were evaluated. Results MiniECG recordings showed large ST-deviation in the majority of leads, with maxima of 5.0±1.3mm in lead V3 and 6.7±3.4mm in lead M2 at 21 minutes post-occlusion. MiniECGs and 12-lead ECGs showed first signs of ischemia, defined as ST-deviation of ≥1mm in at least one lead, at 1±0.0 and 1.4±0.7 minutes post-occlusion respectively. All pigs showed ischemia induced rhythm disorders on both miniECGs and 12-lead ECGs. Conclusion The miniECG showed similar performance in early detection of ischemia related ST-deviation compared to 12-lead ECG. The miniECG is able to detect various rhythm disorders. This shows the potential of the miniECG for cardiological home monitoring. References 1. Diercks, D. B., Kontos, M. C., et al. Utilization and Impact of Pre-Hospital Electrocardiograms for Patients With Acute ST-Segment Elevation Myocardial Infarction. J. Am. Coll. Cardiol. 53, 161–166 (2009). 2. Bansal, A. & Joshi, R. Portable out-of-hospital electrocardiography: A review of current technologies. J. Arrhythmia 129–138 (2018) doi:10.1002/joa3.12035.
Generation of synthetic coronary artery geometries
Rajarajeswari Ganesan, Petrus L.J. Hilhorst, Marcel Van’t Veer, Pim A.L. Tonino, Frans N. van de Vosse, Wouter Huberts
Abstract: Background: In Silico Clinical Trials (ISCT) show a promising future as they could (partially) replace human clinical trials that aim to evaluate the safety, efficacy and usability of new diagnostic/interventional procedures or medical devices. The major challenge for the human clinical trials is the difficulty to take into account enough patients in order to represent the full population distribution of interest, i.e., data including all demographics, patient characteristics like age and gender, and physiological variations. In our work, we are focussing on generating synthetic datasets to include all variations within the population and physiological range. In this study, we focus on coronary artery disease. Method: Initially, a template 3D geometry of the coronary arteries will be generated based on real X-Ray angiograms. Subsequently, parameters such as radii, length and bifurcation angle are varied to generate different 3D geometries to represent the anatomical variations. 2D projections are generated from these different 3D geometries resulting in synthetic 2D angiograms. Geometric uncertainty will also be quantified. A machine learning model will be trained with the synthetic geometries and FFRs which are generated by using computational fluid dynamics simulations. Subsequently, the ML model will be used to estimate the FFR with a real patient geometry as input to demonstrate that developing a ML model on synthetic datasets are feasible. FFR has already been measured clinically for the real data. At the moment, we have already collected and processed >1000 real X-Ray Angiograms, which originate from the FAME I study. Conclusion: Here, we present our approach and preliminary results of the generation of synthetic datasets for coronary artery disease patients. These synthetic datasets will resolve the patient privacy issues and these datasets can be made open for public access. These datasets can also help in study of many rare diseases using artificial intelligence (AI) in future.
A closed-loop modelling framework for cardiac-coronary interaction
Anneloes Munneke, Joost Lumens, Theo Arts, Tammo Delhaas
Abstract: A computational model describing the interaction between cardiac (sarcomere) mechanics and coronary hemodynamics might be useful for studying coronary perfusion in case of pathophysiologies that effect cardiac mechanics (e.g. mechanical dyssynchrony). Therefore, a mathematical model of coronary mechanics was implemented in the previously published multi-scale CircAdapt lumped-parameter model of the closed-loop cardiovascular system. The coronary model consisted of a relatively simple one-dimensional network of the major conduit arteries and veins as well as a lumped parameter model with three transmural layers for the microcirculation. Intramyocardial pressure was assumed to arise from transmission of ventricular cavity pressure into the myocardial wall as well as contraction-related wall stiffening, based on global pump mechanics and local myofiber mechanics. Model predicted waveforms of global epicardial flow velocity, intramyocardial flow and diameter were qualitatively and quantitatively compared with reported data. Versatility of the model was demonstrated in a case study of aortic valve stenosis. The reference simulation correctly described the phasic pattern of coronary flow velocity, arterial flow impediment, and intramyocardial differences in coronary flow and diameter. Predicted retrograde flow during early systole in the aortic valve stenosis was in good agreement with findings in patients. In conclusion, we presented a powerful multi-scale modelling framework that enables realistic simulation of coronary hemodynamics. Although simplified, this modelling framework can be used for educational purposes and as research platform for in-depth studies of cardiac-coronary interaction in healthy and diseased conditions.
A strain energy limiter approach for atherosclerotic plaque rupture modeling
Aikaterini Tziotziou, Ali Akyildiz
Abstract: Atherosclerosis, defined as plaque formation inside the arterial wall, is one of the most prevalent cardiovascular diseases. Rupture in an atherosclerotic plaque usually leads to life-threatening clinical events, such as an ischemic attack or stroke [1]. Plaque rupture could be mechanically characterized as material fracture in the plaque tissue. Fracture in a material could be divided into two parts: the crack initiation and the crack propagation. The mechanisms in either of these phenomena in plaque rupture are scientifically not well studied. The greatest challenge involved arises as in-vivo, real-time observations and measurements of plaque tissue rupture are hard to make. Alternative is utilizing numerical damage models in order to describe and predict the fracture mechanisms of atherosclerotic plaque tissue [2]. This study focused on developing a theoretical and computational framework of Strain Energy Density Function (SEDF) with energy limiter damage model for atherosclerotic plaque rupture. The generated numerical damage model was implemented in Neo-Hookean and Holzapfel-Gasser-Ogden (HGO) SEDF via material user subroutines (UMAT) in the finite element software ABAQUS. With the developed computational models, rupture in tissue-engineered plaque analogs were simulated and compared against experimental test results. The conducted parametric study investigated the influence of various energy limiters in matrix and fiber components. The results demonstrated that the case-specific computational models with SEDF with energy limiter damage model reproduced the rupture in tissue-engineered fibrous plaque analogs, both for crack initiation and propagation phenomena. The results further revealed that the collagen fibers have a dominant role in pre-failure and failure mechanics of plaque analogs. The SEDF with energy limiter model could be further optimized by developing a fully automated damage model and validated with real plaque tissue. [1] Akyildiz, A.C., Speelman, L., van Velzen, B., Stevens, R.R.F., van der Steen, A.F.W., Huberts, W., Gijsen, F., 2018. Intima heterogeneity in stress assessment of atherosclerotic plaques. Interface Focus. 8, 20170008. [2] Holzapfel, G.A., Mulvihill, J.J., Cunnane, E.M., Walsh, M.T., 2014. Computational approaches for analyzing the mechanics of atherosclerotic plaques: A review. J. Biomech. 47(4), 859-869.
Improving cardiac strain imaging by compounding multi-probe axial displacements based on unit axial vectors
Peilu Liu, Hans-Martin Schwab, Richard Lopata
Abstract: Multi-probe ultrasound (US) imaging has been introduced to improve strain estimation by radially projecting and angularly compounding axial displacements in homogeneous phantoms and vascular applications. However, a different compounding approach is needed for cardiac strain imaging, to cope with the non-circular shape, asymmetrical and inhomogeneous deformation pattern, and high level strain in different regions of the left ventricle (LV). This study introduces a unit axial vectors based displacement compounding method to tackle the aforementioned issues, and improve myocardial motion tracking and strain estimation using multi-probe ultrafast US. In an ex-vivo experiment of a beating porcine heart (LifeTec, NL), two phased array probes (P4-2v, Verasonics) were attached on an arch to image the LV at a frame rate of 170 frames per second. Probe 1 (P1) was positioned to acquire the parasternal short axis view while probe 2 (P2) was rotated every 30˚ from 30˚ to 60˚ with respect to P1. Images of all probes were registered automatically and radio frequency (RF) based speckle tracking were performed subsequently. The two axial displacement datasets acquired by the dual probes were compounded based on the unit axial vectors, where the true displacements were measured by solving two linear equations with a unique solution. Furthermore, by weighted compounding the axial displacements of the triplex probes (P1, P2 30˚ and P2 60˚), a least-squares solution of the true displacements was derived. The performances of motion tracking and strain estimation were compared for single probe (SP), dual-probe (DP) and triplex-probe (TP) compounding approach, respectively. The largest improvements were found using TP compounding in terms of motion tracking, elastographic signal-to-noise-ratio (SNRe), strain magnitudes, and strain curve shapes for both strain components. After TP compounding, mean tracking error reduced significantly (-39%), SNRe increased by 226% and 51% for radial and circumferential strain, respectively, and strain curves revealed less noise for all regions. Compared to DP compounding (P2 60˚), TP compounding substantially increased SNRe by 12.5 (radial) and -8.8 in regions, where strain directions were close to the axial direction of the third probe (P2 30˚).
Evolution of right ventricular tissue abnormalities in early-stage Arrhythmogenic Cardiomyopathy: An imaging-based patient-specific modeling study
Nick van Osta, Feddo Kirkels, Tim van Loon, Maarten Jan Cramer, Arco Teske, Tammo Delhaas, Kristina Haugaa, Joost Lumens
Abstract: Introduction: Arrhythmogenic Cardiomyopathy (AC) is an inherited cardiac disease which is characterized by life-threatening ventricular arrhythmias and progressive cardiac dysfunction. Early disease detection and risk stratification is important as geno-positive subjects with and without symptoms may suffer from sudden cardiac death. We propose a patient-specific computer modelling approach to non-invasively reveal the development of myocardial disease substrates underlying the regional right ventricular (RV) deformation abnormalities in individual AC mutation carriers. Methods: In 2 individuals carrying a plakophilin-2 mutation, regional longitudinal deformation patterns of the three ventricular walls were obtained using speckle-tracking echocardiography at baseline during follow-up (4.2 and 9.9 years). Using multiple adaptive important sampling (which includes measurement uncertainty), Digital Twins were created by personalizing the CircAdapt model of the cardiovascular system. The Digital Twins at baseline and during follow-up reveal evolution of the investigated regional tissue properties myocardial contractility, compliance, activation delay, and work. Results: This framework was able to reproduce the regional deformation patterns measured at baseline and during follow-up. Both patients developed abnormal basal deformation patterns during follow-up. Our model revealed this was paired with an increase in heterogeneity in tissue properties. Patient 1 developed heterogeneity in contractility (75kPa/s [p=.228] at baseline to 347kPa/s [p<.001] at last follow-up). No activation delay was present in this subject (p=.188 and p=.242 at baseline and at last follow-up). Heterogeneity in compliance developed from 0.11 (p=.014) to 0.49 /kPa (p=.002). Heterogeneity in work-density increased from 1927 (p<.001) to 5497kPa (p<.001). Patient 2 did not develop a contractile substrate (p=.336 and p=.104 for baseline and last follow-up) or delayed basal activation (p=.336 and p=.190 for baseline and last follow-up). Heterogeneity in compliance and work did both develop from 0.04 (p=.070) to 0.1 /kPa (p=.004) and from 865 (p=.100) to 2504kPa (p<.001), respectively. Conclusion: Our patient-specific modelling approach is able to reveal individual myocardial substrates including uncertainty underlying the regional RV deformation abnormalities and allows investigation of substrate development. Early abnormalities in RV longitudinal strain are most likely caused by increased heterogeneity in tissue compliance. Future studies will investigate whether this approach can improve early detection and risk stratification in geno-positive subjects.
Surface electrocardiogram reconstruction using intra-operative electrograms
Hanie Moghaddasi, Alle-Jan van der Veen, Natasja M.S de Groot, Borbala Hunyadi, Richard C. Hendriks
Abstract: Atrial Fibrillation (AF) is the most sustained arrhythmia in the heart. On the surface electrocardiogram (ECG), AF is characterised by the irregular RR intervals and by fibrillatory waves or the absence of a P wave. Since AF is a progressive disease, timely and correct detection is crucial for AF treatment. Detailed insight into the areas of arrhythmia-related electropathology can be obtained by analyzing high-resolution (inter-electrode distance 2mm) electrograms (EGMs). However, these measurements are rather invasive. By integration of high-resolution epicardial mapping data and surface ECG data, we hope to learn how different stages of AF represent themselves on the ECG. Eventually this can help to guide to identify areas of electropathology as target sites of ablation therapy on the less invasive ECG. A first step in this direction is to learn how to reconstruct the ECG based on EGM measurements. In practice, however, EGMs are only measured at few atrial locations, not covering the complete atria. An important question therefore is: How can we reconstruct ECG based on the observations from a limited part of the heart? To answer this question, we propose two methods. In the first method, we increase the number of observations from a part of the right atrium (RA) to the whole RA by synchronizing EGMs that are measured at different moments in time based on the local activation time (LAT). In the second method, under the assumption that atrial EGMs measured at different spatial areas are linearly related, the conductivity matrix is estimated for the whole atrium which enables us to reconstruct the ECGs from the limited observations. The second method brings twofold benefits. First, the conductivity matrix can be used as a novel diagnostic tool to detect AF as well as areas of electropathology. Second, it provides a practical solution to reconstruct epicardial potentials from ECGs, non-invasively. The results show that method one increases the reconstruction accuracy. Furthermore, the conductivity matrix reveals the structural differences between sinus rhythm (SR) and AF episodes which could be the first step to interpret the underlying electropathology of AF.
Design comparison om implants for revision total knee arthroplasty: A finite element analysis
Simone Stocco, Navid SoltaniHafshejani, Dennis Janssen
Abstract: INTRODUCTION : Bone loss is a major concern in revision total knee arthroplasty (rTKA). Optimal fixation and stability in rTKA cannot be performed without adequate bone loss management. An overall comparative assessment of the implant design is the first step towards a homogenized treatment procedure that could help the preoperative analysis. Clinical and finite element (FE) analysis have been carried out to investigate the outcome of the different augment’s design, however there is no comprehensive research that compares all possible options. Computing the micromotions, this research aims to investigate the influence of the augment design on the primary fixation under different bone conditions and implant configurations. METHODS: FE Models were created of the implanted proximal tibia with different configurations including four modular augments: block, cone , sleeve and plain( as a reference). The bone was considered as a heterogenous elastoplastic material while the implant and augments were assumed to behave as elastic. A walking cyclic load was applied on the medial and lateral condyles of the polyethylene insert. Bone loss is simulated by deactivating specific elements of the resected bone (according to the AORI classification) and the micromotion of each bone condition was analyzed with all the implant configurations. RESULTS: The magnitude of the micromotions was slightly smaller with the AM and PM bone losses, in addition, the results variation among them was neglectable, except for the sleeve that caused less micromotions in the PM case. . While the cone and sleeve showed similar behavior, the block configuration had a noticeable reduction of micromotions. AM and PM bone losses, treated with cones, produced the worst result and the more significant variation occurred with ST bone losses. DISCUSSION: Despite having the best positive effect on micromotion reduction, the block, might not be clinically feasible due to the greater bone resection required. Interestingly in all cases, the peak micromotions were located on the stem, showing crucial influence in the outcome. Clinical aspects should be considered together with the effect of the augment on the potential for bone restoration. SIGNIFICANCE: The sleeve implants resulted in being the most suitable option for rTKA.
Reconstruction of 3D skeletal knee kinematics from sparse ultrasound data and motion tracking system
Dennis A. Christie, René Fluit, Guillaume V. Durandau, Saša Čigoja, Massimo Sartori, Nico J.J. Verdonschot
Abstract: Investigation on the human skeletal kinematics is crucial for a better understanding of human motion analysis. Due to the nature of human skeletal structures, which are surrounded by soft tissues, the measurement of true bone kinematics is challenging. Ultrasound together with a motion tracking system has the potential to reconstruct the skeletal kinematics in a non-invasive and non-radiative manner. When multiple Amplitude mode (A-mode) ultrasound transducers are used, multiple points of a bone landmark can be obtained. Together with a motion capture system, a set of points in 3D space is obtained. A unique 6-DoF rigid-body transformation can be found that best matches all points to the bone model. The proof of concept to obtain 3D skeletal kinematics was provided by [1]. However, the existing setup did not achieve the desired accuracy. This is because the measurements from A-mode ultrasound do not provide sufficient information on the bone geometric structure. Brightness-mode (B-mode) ultrasound, on the other hand, can be utilized to detect the curvature of the bone surface [2]. We hypothesized that fusing these two types of modalities provides rich enough information on geometrical features and landmarks of the bones. In this study, we will investigate the generalization of the ultrasound system in this setup, i.e., a hybrid of A- and B-mode. To address the 3D point cloud sparsity problem, a filtering-based technique, i.e., particle filtering [3], will be used for 3D point cloud registration. The main objective of this study is to aim for surgical navigation kinematics accuracy requirement, which is specified as 1-degree rotation error and 1 mm translation error. To demonstrate the technical and clinical feasibilities of this concept, an in-vivo experiment with healthy individuals will be conducted.
Computational modeling of the effect of continuous positive airway pressure on the work of breathing during invasive and non-invasive mechanical ventilation
Anouk van Diepen, Tom Bakkes, Ashley De Bie, Simona Turco, Arthur Bouwman, Pierre Woerlee, Massimo Mischi
Abstract: Patients with pneumonia, Acute Respiratory Distress Syndrome (ARDS) and COVID-ARDS frequently require positive pressure mechanical ventilation, either non-invasively or, when needed, also invasively. Strong spontaneous breathing efforts in the fragile lungs can lead to patient self-inflicted lung damage (P-SILI) [1]. Application of Positive End Expiratory Pressure (PEEP) or Continuous Positive Airway Pressure (CPAP) decreases the strength of spontaneous efforts and work of breathing (WOB). In this work we present a simulation environment to evaluate the effect of CPAP or PEEP on WOB during invasive and non-invasive mechanical ventilation, since objective WOB measurements are currently difficult to obtain breath-by-breath. Software that correctly estimates WOB could give insights in the mechanisms underlying the observed decrease in WOB. We used an in-silico non-linear lung-airway model [2] to calculate pleural pressure and lung volume during spontaneous breathing in non-invasive and invasive ventilation with 0 cmH2O and 10 cmH2O CPAP/PEEP. WOB was evaluated using Campbell diagrams that include elastic, resistive and visco-elastic work. We tuned the patient muscle strength such that for identical breath duration a tidal volume of 0.5 L was achieved. For healthy patients without PEEP/CPAP, WOB was estimated to be 0.52 J/L (consisting of 0.38 J/L resistive WOB and 0.16 J/L elastic WOB). This corresponds to values found in the literature for healthy patients [3]. At 10 cmH2O, WOB decreased to 0.44 J/L (resistive WOB 0.28 J/L and elastic WOB 0.16 J/L), a decrease of 15% in agreement with studies that found a similar decrease during CPAP [3]. This reduction is mainly caused by a decreased contribution of resistive WOB in the model. During invasive ventilation, WOB increased to 1.12 J/L (elastic WOB 0.17 J/L and resistive WOB 0.95 J/L) at 0 cmH2O and 1.04 J/L (elastic WOB 0.18 J/L and resistive WOB 0.86 J/L) at 10 cmH2O CPAP/PEEP. WOB is increased due to the extra resistance added by the tubing. WOB calculated by the model behaves as expected for different PEEP/CPAP values and gives values for WOB that correspond to values found in the literature. In the model, the decreased WOB during application of CPAP/PEEP is caused by decreased resistive work. [1] Carteaux, Guillaume, et al. "Patient-Self Inflicted Lung Injury: A Practical Review." Journal of Clinical Medicine 10.12 (2021): 2738. [2] Athanasiades, A., et al. "Energy analysis of a nonlinear model of the normal human lung." Journal of Biological Systems 8.02 (2000): 115-139. [3] Kallet, Richard H., and Janet V. Diaz. "The physiologic effects of noninvasive ventilation." Respiratory care 54.1 (2009): 102-115.
Label-free second and third harmonic generation microscopy for real time visualization of osteogenesis imperfecta fibroblasts culture
Yuanyuan Ma, Jasmijn M. Rootlieb, Lisanne E. Wisse, Ludo van Haasterecht, Dimitra Micha, Elisabeth M.W. Eekhoff, Peter Kloen, Thomas Rustemeyer, Marie L. Groot
Abstract: The inherited bone disease Osteogenesis Imperfecta (OI) can lead to brittle bones that fracture easily and serious skeletal deformities due to collagen defects. To further mimic and understand this disease, in vitro models of 2D monolayer fibroblast cultures obtained from subdermal skin biopsies were developed and extended to 3D models. The models allow for studying of the collagen formation and possible abnormalities in extracellular matrix (ECM) and will enable studies of the mechanisms underlying defects in the quality and quantity of collagen fibres in OI skin fibroblasts when compared to the healthy cells. The most common methods for measuring the components in fibroblasts cultures are collagen immunostaining, which is time consuming and prohibits time-lapse measurements, or second harmonic generation microscopy (SHG) which enables visualization of collagen fibres. Here we have extended SHG with third harmonic generation microscopy (THG) to visualize cells and cell nuclei (fibroblasts) and with multiphoton autofluorescence (MAF) signal from the cytoplasm of the fibroblasts. Our techniques allow for the visualisation of the cultures in 3D and in real-time, resulting in the possibility of making time-lapse videos. For both patient OI and control fibroblast cultures, fibroblasts and their nuclei were clearly visualised by the backward and forward THG signals; the collagen fibres by the SHG channel, and the cytoplasm signal with the autofluorescence channel. First results on the comparison between cultures of OI patients and controls reveal differences in collagen orientation and density. We further visualized dynamics induced upon a scratch assay in OI and control fibroblasts cultures, which will be presented at the meeting.
Sitting behaviour detection to reduce the risk of non-communicable disease in office workers – A study protocol
Linda Ong, Ming Cao, G.J. (Bart) Verkerke, C.J.C. (Claudine) Lamoth, Elisabeth Wilhelm
Abstract: Musculoskeletal disorders (MSDs) and metabolic syndrome (MetS) are two common health conditions in office workers. According to WHO, 1.7 billion people worldwide experience MSDs [1]. Around a quarter of cases occur in Europe [2] and MSDs frequently are related to the neck, upper, and lower back region. Furthermore, MetS, which is known as a risk factor for developing chronic diseases, such as diabetes type II, and cardiovascular diseases, occurs up to 32.6% in office workers in some European countries [3]. Currently, support for the prevention of these two health issues is limited. People only seek treatment once the symptoms are perceived and they are limited in daily functioning. Unfortunately, body function limitation indicates MSDs and MetS might be already in the chronic phase, which will be more difficult to treat and reverse the condition. Therefore, the prevention of developing MSDs and MetS is necessary. One of the main risk factors for developing MSDs and MetS is sedentary behaviour during performing work that requires static and prolonged sitting. Even upright sitting posture can cause serious health issues if one keeps the static posture for prolonged time. Therefore, the main aim of the study is to detect sitting behaviour, including sitting transition, postures, and duration. Existing sitting posture detection algorithms based on pressure mat data are developed in a controlled environment [4]. We will develop the first posture detection algorithm based on real-life environment measurements. The study will start at the end of 2021 by involving 55 participants that are at risk of developing MSDs and MetS because they are office workers. The sitting monitoring will be done for 7 hours daily for 5 consecutive days. 80% of the data will be used for training, 10% for development, and 10% for testing. Sitting posture classification will be developed based on novel machine learning approach, while sitting transition and duration will be identified by thresholding the result from posture classification. To be useful for health advice, the system should classify common sitting postures of upright, slouch, slump, lean left, right, and backward with accuracy and f1-score of at least 85% and 80% respectively. REFERENCES [1] World Health Organization, "Musculoskeletal Conditions," 8 February 2021. [Online]. Available: [2] EUROSTAT, "eurostat," 2010. [Online]. Available: [3] M. Strauss, P. Foshag and R. Leischik, "Prospective evaluation of the cardiovascular, cardiorespiratory, and metabolic risk of german office workers in comparison to international data," International Journal of Environmental Research and Public Health, vol. 17, no. 5, 2020. [4] A. M. Kappattanavar, N. Steckhan, J. P. Sachs, H. Freitas da Cruz, E. Böttinger, and B. Arnrich, "Monitoring of Sitting Postures With Sensor Networks in Controlled and Free-living Environments: Systematic Review," JMIR Biomedical Engineering, vol. 6, no. 1, p. 1, 2021.
Catch-up phenomenon in baseball pitching, fact or fiction?
Ton Leenen, Bart van Trigt, Marco Hoozemans, Dirkjan Veeger
Abstract: One of the most valuable skills in overhand sports, specifically in baseball pitching, is the ability of throwing a ball at speeds up to 100 miles per hour. This requires a highly coordinated whole-body motion requiring transfer of kinetic energy from the lower body, pelvis and trunk up to the upper body. This transfer of energy through body segments is often referred to as the kinetic chain and requires sequential motions of subsequent body segments. This appears to be a key factor for optimal throwing performance. Movements out of sequence may hamper optimal transfer of kinetic energy due to deficiencies in the kinetic chain that either results in reduction in performance or result in an – additional – compensation by the upper body segments to accommodate for the energy loss to maintain performance. This mechanism is referred to as the ‘catch-up’ phenomenon. The upper body segments are therefore conceivably more prone to (overuse) injuries. It can be expected that a kinetic chain disruption, in addition to a possible performance reduction, also results in upper body compensation that can be deduced from changes in upper body kinetics. The purpose of the present study was to examine the effect of experimentally limiting the kinetic chain in fastball pitching on ball speed, elbow and shoulder kinetics in elite youth baseball pitchers. The experimental design consisted of two (within subject) conditions. One condition acted as control condition where pitchers were unimpeded while pitching. The other condition involved a limitation to use the kinetic chain by taping the pelvis and trunk. In both conditions, pitchers were instructed to throw fastballs until a minimum of 25 pitches were captured. Inverse dynamic solutions were used calculate the internal elbow and shoulder moments of 11 elite youth baseball pitchers. The pitchers who were limited in using their kinetic chain when throwing fastballs showed significant lower ball speeds compared to the pitchers that were allowed to throw as they normally would. The internal moments showed no significant differences between the two conditions.
Dynamic balance quality evaluation: How does rotation of the trunk influence balance?
Junhao Zhang, Heike Vallery, E.H.F. van Asseldonk, P.H. Veltink
Abstract: Most human balance quality evaluation methods, which are based on the simplified linear inverted pendulum model, assume that the rotational inertia of the upper body does not influence balance. In our previous work, Refai et al. proposed a Portable Gait Lab (PGL) which uses three wearable Inertial Measurement Units (IMUs) to realize portable analysis of balance quality based on this assumption. However, in some scenarios, for example, when there is a large disturbance acting on the human body, an instantaneous torque generated by the hip muscles will create angular acceleration of the trunk to prevent a fall. Therefore, the rotation of the trunk definitely contributes to balance control, and the goal of our research is to derive a novel model to analyse the influence of trunk rotation in balance.
Joint stiffness estimation via neuromusculoskeletal modeling
Christopher Pablo Cop, Alfred C. Schouten, Bart F.J.M. Koopman, Massimo Sartori
Abstract: Background: Quantifying human joint stiffness in vivo during movement remains a challenge. We propose a data-driven neuromusculoskeletal model-based approach to estimate multi-muscle stiffness as well as resulting joint stiffness. Importantly, the proposed model-based approach does not require joint perturbations, in contrast to well established system identification methods. Musculoskeletal models often use a standard Hill-type muscle model that often neglects muscle architecture features, specifically pennation angles, when estimating equivalent muscle-tendon stiffness. Our goals are to include pennation angles into the current muscle-tendon unit stiffness model, and to validate the joint stiffness profiles predicted by our extended muscle model against an ensemble-based system identification method during ankle rotations. Methods: A young healthy voluntary participant was instructed to follow a sinusoidal plantar-dorsi flexion angle target (amplitude: 0.15 rad, frequency: 0.6 Hz). Electromyography (EMG) and kinematic data were used to estimate joint stiffness using an extended EMG-driven neuromusculoskeletal model that considers the muscle fiber’s pennation angle in the equivalent muscle stiffness computation: Keq= 𝜕𝜕Fm𝜕𝜕lm ∙ (cos𝛼𝛼)2+Fmlm∙(sin𝛼𝛼)2, where Fm and lm are the force and length, respectively, of the muscle fiber, and α represents the muscle fiber’s pennation angle. Joint stiffness predictions were compared to an ensemble-based system identification method. Results: Our extended model’s joint stiffness prediction was comparable to the system identification reference profile (R2 = 0.69 and root mean squared error, RMSE, = 3.33 Nm/rad). Discussion: We extended our joint stiffness formulation to consider the muscle’s pennation angle. The inclusion of physiological features in musculoskeletal models may be beneficial for joint stiffness estimation. Next steps include the implementation of history-dependent muscle properties, such as residual force enhancement and residual force depression, as well as a short-range stiffness module, in the Hill-type muscle model to improve its capabilities of force and stiffness estimation during complex tasks involving eccentric, concentric and isometric contractions. Relevance: Being able to decode instantaneous joint stiffness from EMGs and kinematic data, without the need of applying external joint perturbations, can radically change the way biomimetic controllers for orthoses and prostheses are designed and implemented.
Bionic H2020: Detection of fatigue in human biomechanics during a simulated work task
Luca Marotta, Bert-Jan van Beijnum, Jaap Buurke, Jasper Reenalda
Abstract: Rationale: Prolonged, repetitive movements (e.g. carrying loads) can cause biomechanical stress on a worker. In particular, elderly workers are at higher risk of overloading than younger workers [1]. Physical overloading and exposure to ergonomic risk factors during work activities could cause musculoskeletal disorders (MSDs) [2]. Detection of fatigue and timely feedback could prevent overloading and MSDs and is among the main goals of BIONIC, a European research project aiming to develop an unobtrusive system for real-time risk alerting and coaching [3]. Aim of this study is to assess the capability of a machine learning classifier to detect fatigue based on IMU-derived features during simulated work tasks. Methods: 10 healthy individuals (age range 40-65) will be recruited to participate in the study. Subject(s) are asked to perform: 1) a simulated work task, consisting of picking up a 5kg box and carrying it with forearms parallel to the ground for 10 meters, 10 consecutive times; 2) a fatiguing run to exhaustion; 3) the simulated work task as described above, now in a fatigued state. Subjects wear a suit with 17 IMUs (left and right foot, left and right tibia, left and right thigh, pelvis, sternum, left and right shoulder, left and right forearm, left and right wrist, left and right hand, head). Data are segmented in gait cycles based on peak tibial accelerations and features are extracted from joint angles and segmental accelerations. Preliminary Results: Initial results from a single subject case study (male, 46 years, 167 cm, 60 kg) indicate that fatigue is detected with an accuracy of 94.2% (sensitivity 92.5%, specificity 96.0%). RPE increased from 7 (rested condition) to 14 (fatigue condition) after running for 20 minutes at 13.3 km/h and performing the second work task). Conclusions: These results show the feasibility of using IMUs and machine learning to detect physical fatigue in work tasks. A leave-one-subject-out cross-validation approach will be used with the complete dataset in order to evaluate if the algorithm can potentially be generalized. Translation of the fatigue detection algorithm to real-life workplaces. enabling the design of workplace interventions adapted to the fitness levels of workers. [1] B. C. H. De Zwart, M. H. W. Frings-Dresen, and F. J. H. Van Dijk, “Physical workload and the ageing worker: A review of the literature,” Int. Arch. Occup. Environ. Health, vol. 68, no. 1, pp. 1–12, 1995, doi: 10.1007/BF01831627. [2] E. Valero, A. Sivanathan, F. Bosché, and M. Abdel-Wahab, Musculoskeletal disorders in construction: A review and a novel system for activity tracking with body area network, vol. 54. Elsevier, 2016, pp. 120–130. [3] “About – BIONIC.” (accessed Oct. 18, 2021).
Activation primitive-driven musculoskeletal modelling of human locomotion: towards model-based control of bionic legs
Federica Damonte, Guillaume Durandau, Herman van der Kooij, Massimo Sartori
Abstract: Powered prosthetic devices have the potential to improve the quality of life of lower limb amputees, improving their mobility [1]. However current control systems that try to mimic human control strategies to have better performances employ electromyographic signals (EMGs) [2] which are not practical in case of lower limb amputation. EMG-dependence in prosthetic controllers can be addressed by exploiting the concept of muscle synergy and modularity. The human neuromuscular system has an inherent redundant nature. The number of muscles that generates a movement around a joint is greater than its degrees of freedom. It is hypothesized that the nervous system organises motor control by activating fixed groups of muscles as individual units, or muscle synergies, determined by their functionality and/or location. Therefore, in order to generate a specific movement only a low dimensional set of basic patterns of excitations is required for a large number of muscles [3]. Gonzales-Vargas et al.[4] analysed EMGs during gait at different speeds and inclinations conditions, from which they built a model to reconstruct the activity of leg muscles as a linear combination of four basic patterns. The generated muscle excitation showed minimal loss of accuracy in estimating EMGs. The aim of the current research is to interface the synergies model [4] with existing musculoskeletal models to estimate muscle-tendon units forces and joint moments in real-time with no need for EMG sensing [5]. To assess the ability of the model to predict accurately joint moments, data from dynamic trials on healthy subjects walking on a treadmill will be used . At each timestep the synergies model requires an accurate gait phase estimation using sensory informations from the treadmill, specifically ground reaction forces together with speed and elevation . Results would be validated in comparison with estimated joint moments from inverse dynamics and angles from inverse kinematics. We assume as the main advantage of the proposed approach that the biomimetic feedforward control will improve the physical interaction between the user and a powered lower limb device , yielding to a more natural and robust control.
Accuracy and repeatability of joint sparsity multicomponent estimation in MR Fingerprinting
Martijn Nagtegaal, Laura Nunez-Gonzalez, Dirk Poot, Thijs van Osch, Jeroen de Bresser, Juan Hernandez Tamames, Frans Vos
Abstract: MR fingerprinting (MRF)1 is a promising method for quantitative characterization of tissue parameters, e.g. T1- and T2 -relaxation times and M0 magnetization. By using efficient acquisition schemes and highly undersampled images, quantitative information can be obtained in clinically feasible scan times. Traditionally, parameter estimates are obtained voxel-by-voxel, assuming a single tissue-type (T1,T2-combination) per voxel.These voxel-wise, single component estimations neglect the possibly of a mixture of tissues present within one voxel, leading to partial volume errors and less details in smaller structures. The Sparsity Promoting Iterative Joint Non-negative least squares Multi-Component MRF method (SPIJN-MRF)2 facilitates tissue parameter estimation per tissue type as well as partial volume segmentations for each of these tissues. The aim of this project was to evaluate the accuracy and repeatability of the SPIJN-MRF parameter estimations and partial volume segmentations in the human brain. This was done (1)through numerical simulations3 and (2)using in-vivo acquired MRF-data1 (acquisition time 5:36) from 5 subjects, scanned on the same week-day for 8 consecutive weeks. SPIJN-MRF partial volume segmentations were compared to those obtained by two conventional methods, FSL-FAST4 and SPM125, based on synthetic T1-weighted images. SPIJN-MRF showed higher accuracy in simulations compared to FSL- and SPM12-based segmentations: Fuzzy Tanimoto Coefficients(FTC) comparing the segmentations and Brainweb references were higher than 0.95 for SPIJN-MRF in all the tissues and between 0.6 and 0.7 for SPM12 and FSL in white and gray matter(WM,GM) and between 0.5 and 0.6 in CSF. For the in-vivo MRF data estimated relaxation times were in line with literature and minimal variation was observed. Furthermore, the coefficient of variation (CoV) for SPIJN-MRF tissue volumes were 10.5% for myelin water, 6.0% for WM, 5.6% for GM, 4.6% for CSF, and 1.1% for the total brain volume. CoVs for CSF for SPIJN-MRF were in line with those obtained with SPM12 and FSL and CSF-maps from SPIJN-MRF showed increased detail. SPIJN-MRF provides accurate and precise tissue relaxation parameter estimations taking into account intrinsic partial volume effects. Myelin water fraction maps were obtained as well, possibly relevant for evaluating diseases affecting the white matter such as MS, not available with conventional scans and methods.
Large-scale multi-channel electromyography and musculoskeletal modelling via wearable smart garments to support clinical decision-making
Donatella Simonetti, Bart F. J. M. Koopman, Massimo Sartori
Abstract: Introduction: Clinical decision-making requires above all rapidity. Currently, motor deficit evaluation is based on a simplistic and subjective assessment, i.e. a simple 10m walk. It fulfills the main requirement, but it is not accurate, and it is just based on the evaluator's knowledge and experience. Greater accuracy is achieved in biomechanical laboratories where advanced technology together with neuro-musculoskeletal modeling allows to quantify the subject impairment. However, this is made at the expense of rapidity. Our work is aimed to balance clinical rapidity and biomechanical accuracy. We propose to use advanced signal processing techniques and real-time neuro-musculoskeletal modeling integrated into a smart wearable garment. A simple lower leg sock instrumented with a large-scale multi-electromyography (EMG, 64 channels) grid and inertial sensors (IMUs) allowing to get over the lengthy setup and to prevent human error in the manual electrodes’ placement. The smart clothing together with advanced signal processing techniques provides muscle activation and muscle-tendon unit (MTU) kinematics necessary to finally model the subject-specific musculoskeletal properties. Methods: Three healthy subjects were equipped with 33 reflective markers and a lower leg flexible garment instrumented with 64 equally distributed EMG monopolar electrodes. The 64-electrode space is reduced in 5 muscle-specific clusters applying iteratively the non-negative matrix factorization (NNMF) [1] during slow locomotion at 1km/h. Afterward, 5 average muscle activations were extracted during locomotion at different speeds 1, 3, and 5 km/h, and used as input to an offline EMG-driven musculoskeletal model to estimate ankle torque. Results: The NNMF-approach was able to locate the muscles and extract averaged activations during each locomotion speed that resembled with good accuracy the activation recorded with bipolar EMG. Afterward, the musculoskeletal model driven by the automatically extracted muscle-specific activation reproduced experimental ankle torques during gait at different speeds. Conclusions: The combination of a soft sensorized garment and the automatic procedure for the extraction of muscle activations added to the framework for neuromuscular modeling has a good potential to become a resource for fast and more accurate clinical decision-making.
Biophysical neuronal model optimization for quantifying human motoneuron properties in vivo
Rafael Ornelas Kobayashi, Antonio Gogeascoechea Hernandez, Jan Buitenweg, Utku Yavuz, Massimo Sartori
Abstract: As the final common pathway of the central nervous system, alpha-motoneurons (MNs) integrate multiple inputs to produce a muscle activation command, the neural drive. Evidence suggest that the neural drive is the linear transformation of the common inputs received by the MN pool (Farina and Negro, 2015). However, the specific details of this neuro-muscular interaction remain unknown, mainly because the MNs output is determined by subject-specific electrophysiological properties impossible to measure in vivo (e.g. soma size, ionic channels conductance, etc.). In this project, we propose a novel approach that combines biophysical neuronal modelling, high-density electromyography (HD-EMG) decomposition and optimization for the non-invasive estimation of subject-specific MN properties in vivo. The method consisted on recording HD-EMG from the tibialis anterior muscle of a subject performing isometric ankle dorsi-flexion on a dynamometer chair (Biodex multi joint system) while tracking pre-defined torque profiles. In vivo MN spike trains were decoded from HD-EMG recordings using convolutional blind-source separation (Holobar, 2014). From the in vivo MN spike trains, the experimental neural drive, and hence the common synaptic input (CSI) to the MN pool (Farina and Negro, 2015), was estimated. For each decoded spike train, a digital MN model was created and driven by the same derived CSI. Using genetic algorithm optimization, soma size and ionic channels activation rates of each MN model were adjusted to minimize the error in discharge rate, recruitment time and temporal spike match. This way we create digital copies, with known electrophysiological features, of each in vivo decoded MN. Similarity between pairs of in vivo and digital MNs was quantified on basis of their coincidence factor (Lynch and Houghton, 201), whereas correlation between measured torque and estimated neural drive was calculated to compare the behaviour of the entire MN pool. Results demonstrate that the optimization of soma size and potassium channel dynamics enables to reproduce individual MN discharge patterns and force profiles. Moreover, resulting parameters distributions provide a quantitative map of the subject-specific neuronal features underlying human movement, which will enable the implementation of intuitive controllers for wearable robots, as well as the development of novel closed-loop neurorehabilitation technologies.
Revealing reflex modulation during goal-directed movement using continuous perturbations
Mark van de Ruit, Winfred Mugge, Frans van der Helm, Alfred Schouten
Abstract: Background: Reflex strength modulation plays a vital role during movement. The ability to rapidly modulate reflex strength determines the steadiness of our movements during our interaction with the environment when faced with unexpected changes. In this study we aimed to develop a new paradigm to assess reflex strength modulation during a goal-directed movement using continuous perturbations. Methods: Participants were comfortably seated with their right hand holding the handle of a robotic wrist manipulator. Participants made rapid goal-directed wrist flexion movements in a force field over 0.32 rad in 200-500 ms. During the main experiment, they received continuous pseudo-random binary sequence position perturbations (peak-to-peak amplitude: 0.06 rad, switching rate: 150 ms). In a validation experiment participants received transient ramp-and-hold perturbations (ramp: 40 ms at 2.0 rad/s) before and after the movement, and at three positions during the movement. Position of, and torque applied to, the handle together with electromyography (EMG) from the flexor and extensor carpi radialis muscles were recorded. Reflex strength was quantified using time-varying system identification (short data segment method [1]) in the main experiment, while for the validation experiment the mean EMG amplitude during the short and long-latency reflex response window was used. Results: Using continuous perturbations, reflex strength was found to be significantly reduced during the movement when compared to before and after the movement. The validation using transient perturbations revealed a significant increase in the short-latency reflex response during movement, whereas no significant change in the long-latency reflex response was found. Conclusion & significance: We successfully developed a paradigm to reveal reflex modulation during movement using continuous perturbations. The paradigm enables us to get information on reflex modulation with a higher temporal resolution and less data than when using transient perturbations. Using the developed paradigm we will be able to gain fundamental new insights in the pathways underlying (impairments in) reflex strength modulation during movement, and hence to truly understand movement disorders and improve sensorimotor rehabilitation. Reference(s): 1. Ludvig, D., and Perreault, E.J: IEEE Trans Biomed Eng, 2012, 59, pp. 3541-3549
Cortical activity related to sensorimotor synchronization guided by different types of external cues
Janne Heijs, Silvana Huertas-Penen, Richard van Wezel, Tjitske Heida
Abstract: External cues are frequently used as therapy in Parkinson’s disease to guide automatic, repetitive movements, such as gait. The effectivity of this therapy, also known as sensorimotor synchronization, depends on the characteristics of the cue. Some patients benefit from visual cues (e.g. bars on the floor), while others prefer the use of auditory cues (e.g. metronome). Moreover, cues with frequencies below 2Hz (discrete cues), and cues consisting of multiple patterns (polyrhythmic cues) appear to be more effective than cues with frequencies above 2Hz (continuous cues) and cues consisting of a single pattern (isorhythmic cues, e.g. metronome). We hypothesize that the effects of different cues are related to the synchronization of different neuronal networks, the Basal Ganglia-Thalamo-Cortical (BGTC)-network, and the Cerebellum-Thalamo-Cortical (CTC)-network. Therefore, the aim of the current study is to evaluate the cortical activity related to sensorimotor synchronization guided by different types of external cues, to assess the effect of the type, frequency and rhythmicity of the cues. It is hypothesized that continuous cues and isorhythmic cues will activate the cortical areas involved in the BGTC-network (e.g. the (pre-)frontal and sensorimotor cortex), while discrete and polyrhythmic cues will activate cortical areas involved in the CTC-network (e.g. the premotor and parietal cortex). We performed a finger tapping experiment in 21 healthy subjects, while providing different types of cues in a 2x2x2-design, to evaluate the effect of: 1) cueing type: visual cues (white, flickering circle) vs. auditory cues (repetitive tones); 2) cueing frequency: discrete cues (1Hz) vs. continuous cues (3.2Hz); and 3) cueing rhythmicity: isorhythmic cues (one pattern/rhythm) vs. polyrhythmic cues (two patterns/rhythms in a [2:3]-relationship). A 32-channel EEG system recorded the electrocortical activity. A 3D-accelerometer below the metacarpal joint of the index finger, recorded the finger tap movements. Preliminary results of the first 7 subjects showed a trend towards increased event-related synchronization in cortical areas associated to the BGTC-network when using continuous or isorhythmic cues, and in areas associated to CTC network when using discrete or polyrhythmic cues. No differences were found between visual and auditory cues.
Quantification of upper limb function using the PowerJar: An exploratory study in healthy subjects and patients with neuromuscular diseases
Ingrid S. van den Heuvel, Marcus P.J. van Diemen, Willem O. Elzinga, Anil Tarachandani, Ajay Verma, Geert J. Groeneveld, Robert J. Doll
Abstract: Neuromuscular diseases (NMDs) are often progressive and characterised by impaired muscle function leading to physical disability. As a result of muscle weakness, daily physical activities can become more complex for these patients. Quantification of their upper limb function in phase two clinical trials is currently done by using handheld dynamometers. However, the clinical relevance of such devices is limited as it is relatively unrelated to daily physical activities. Opening a jar is one of many tasks that patients with an NMD may find difficult to perform. A device which could potentially quantify the upper limb function during such a task is the PowerJar (Usin’Life LLC), a standing bottle with a rotating lid. It measures pressure on the side of the bottle, and the rotation of the lid. Additionally, the torque required to turn the lid can be controlled. Here, we assess the feasibility of using this device in the context of a clinical study. During this exploratory study, 60 healthy subjects and 18 patients with an NMD (Parkinson’s Disease (8), Myasthenia Gravis (5), and Inclusion Body Myositis (5)), performed a series of tasks related to torque and grip strength using the PowerJar. Furthermore, the maximum grip force was measured using a handheld dynamometer (Jamar). Here, we focus on part of the tasks carried out with the PowerJar (e.g., continuous grip force and repeating grip release cycles), and compare the results between healthy subjects and patients. Both healthy subjects and patients were able to perform the tasks. Performance of both groups are compared in this study. Additionally, the maximum grip force obtained using the novel device is compared to the gold standard, the Jamar dynamometer. Finally, the repeatability of the tasks is presented. In this study, we assessed the feasibility of using the PowerJar in a clinical trial including both healthy subjects and patients with NMDs. For the next steps, we are going to 1) combine the results of the surface electromyography (sEMG) measured during the assessments with those of the PowerJar, and 2) study the sensitivity of the PowerJar in drug studies.
Towards an energy-efficient current-mode neural stimulator with non-rectangular waveforms
Konstantina Kolovou-Kouri, Amin Rashidi, Wouter Serdijn, Vasiliki Giagka
Abstract: Electrical neural stimulation has been proven to be a successful treatment for many neurological disorders by activating/inhibiting neural activity through the delivery of electric charge. Electrical neurostimulators typically control the injected charge through either voltage or current pulses; referred to as Voltage Mode Stimulation (VMS) and Current Mode Stimulation (CMS), respectively. Even though VMS benefits from more energy-efficient architectures, CMS is preferred in many applications, due to its increased safety and control over the injected charge. Furthermore, recent studies have exhibited increased energy efficiency as well as stimulation efficacy by employing non-rectangular waveforms (e.g. triangular, Gaussian, exponential) for CMS, instead of the conventional rectangular pulses¹. However, conventional CMS comes at the cost of higher circuit complexity, and less energy-efficient output stages. The low efficiency is mainly due to the constant supply voltage that needs to accommodate the highest required voltage compliance. Consequently, should an output stage demand a lower voltage compliance, the voltage drop over its associated current driver is dissipated as heat. This issue is addressed by adapting the supply voltage level(s) based on the instant voltage/current requirements of output stages²,³. Similarly, the CMS’s efficiency can also degrade over the course of a non-rectangular waveform, as the required voltage compliance varies according to the time-variant stimulation current. Nonetheless, the importance of employing adaptive supply levels for non-rectangular waveforms is overlooked in the literature. The current work highlights the importance of supply level scaling for an output stage with non-rectangular waveform stimulation, and accordingly, proposes a system-level architecture for multi-channel stimulators. A multi-output DC/DC Converter (DDC) allows each channel to choose among the available supply levels (i.e. DDC outputs) independently and based on its instant voltage/current requirement. A system-level analysis is carried out in Matlab to calculate the possible energy savings of this solution, compared to the conventional approach with a fixed supply. For this, a maximum supply voltage of 9V, a purely resistive load of 1kΩ, and different stimulation waveforms are assumed. The energy savings have been simulated for a different number of supply levels and different waveform amplitudes. The simulations illustrate maximum energy savings of 83% by employing 6 supply levels.
Validation and comparison of various finger tapping tasks in Parkinson’s disease patients: A placebo-controlled randomized clinical trial
Eva Thijssen, Emilie M. J. van Brummelen, Geert J. Groeneveld, Robert-J. Doll
Abstract: Various finger tapping tasks are used to quantify the acute effects of symptomatic medication in patients with Parkinson’s Disease (PD). However, direct comparisons between different methods are lacking. Here, we assessed three finger tapping tasks and their sensitivity to detect changes in medication states. The first two tasks were touchscreen-based: Finger-Tapping (FT) and Side-Tapping (ST). These were developed and validated in-house in a previous study including healthy volunteers. The third task, Wide-Tapping (WT), involved a goniometer continuously measuring the angle of the index finger. A total of 20 PD patients were recruited in a randomized, double-blind, placebo-controlled, two-way crossover study. An ‘OFF’ state was induced by withholding subjects’ medication. The following day, the regular morning dose of levodopa/carbidopa or placebo was administered as over encapsulated tablets. Finger tapping tasks were performed pre-dose, as well as 10-, 25-, 45-, 60-, 75-, 90-, 105-, and 210-minutes post-dose. During FT, the active treatment was associated with an increased ability to correctly perform the task, increased speed, and improved tapping rhythm. During ST, the active treatment resulted in increased speed, lower accuracy, more taps inside the target area, and improved rhythm, compared to placebo. During WT, active treatment resulted in increased tapping frequency, improved rhythm, increased opening and closing velocity, increased mean tapping amplitude, and decreased fatigue. The results are partly in line with the clinical literature. First, our previous validation study predicted that the total number of taps, rhythm, and accuracy parameters can be key in detecting medication effects. Second, the ST and WT can be effectively used to detect medication effects. Both devices capture the improvement in tapping speed and rhythm. ST is sensitive to captures effects in motor accuracy compared to WT. In comparison, WT captures fatigue effects more accurately and it allows the capture of continuous finger motion. Lastly, the FT task showed poor sensitivity, probably due to subjects facing difficulties to perform the task. Based on our findings, we recommend using the ST task for clinical research to demonstrate acute pharmacodynamic effects of symptomatic (dopaminergic) treatments in PD.
Tensor-based source separation of functional ultrasound data
Sofia-Eirini Kotti, Borbala Hunyadi
Abstract: Functional ultrasound (fUS) is an emerging technique that provides high sensitivity imaging of cerebral blood volume (CBV) changes in the whole brain without the use of contrast agents. In other words, fUS detects blood flow, or the number of moving red blood cells, in the voxels. As increased metabolic demand of active tissue induces changes in CVB, these changes reflect neuronal activity in the corresponding brain area. The main advantages of this technique are that it can image the whole depth of the brain with unprecedented spatial (50-500um) and temporal resolution (10-100ms), and that it constitutes a (potentially) portable solution, as opposed to fMRI. Additionally, it leads to a large increase in raw ultrasound data available per acquisition. The fundamental challenge that comes with this technique is that it only provides an indirect measure of brain activity through neurovascular coupling (NVC), which is the link between local neural activity and the resulting changes in the cerebral blood flow. This is a system with not entirely known dynamic and non-linear characteristics. Moreover, besides the activity of interest, fUS records a mixture of other ongoing brain activity, physiological artifacts and noise. The goal of this research is to develop tensor-based source separation techniques in order to estimate the activity of interest and the brain’s hemodynamic response function (HRF) to stimuli by learning its non-linear coupling with the fUS signal. Indeed, tensors (multidimensional matrices) are the natural representation of the obtained fUS data: the data can be three-dimensional, for time-varying 2D slices (images), or of higher order when, for instance, time-varying 3D voxels or measurements across different subjects are used. The tensor decomposition techniques should be able to handle the resulting large datasets. Some specific challenges to solve on the way are the following: 1. Improve the pre-processing stage: current SVD-based clutter filtering techniques act on unfolded (vectorised) data, destroying spatial dependencies in the images. This is not optimal, since neighbouring voxels in the brain volume are often highly correlated. 2. Adapt spatial independent component analysis (ICA), a successful tool for functional image processing, such that it can (a) act on unfolded data, exploiting the fact that the images are low-rank, (b) incorporate a parametrized physical model of the neurovascular coupling on the time courses and (c) perform efficient joint decomposition of many (±50) slices that are not acquired simultaneously but pairwise share spatial characteristics.
Interpreting softmax predictions of sleep staging models
Iris Huijben, Lieke Hermans, Alessandro Rossi, Sebastiaan Overeem, Merel van Gilst, Ruud van Sloun
Abstract: Sleep staging is the process of assigning sleep stage annotations to windows of sleep recordings, typically recorded using a polysomnography (PSG). The American Academy of Sleep Medicine (AASM) distinguishes five stages: rapid eye movement (REM) sleep, non-REM sleep (divided into N1-N3), and Wakefulness. Sleep staging is typically done manually by a sleep technician, but thanks to the advent of machine learning, automation of this labor-intensive job has attracted considerable attention. Commonly used machine learning models are convolutional classification networks that - by means of a softmax function - provide a probability for each of the stages, from which the stage with the largest probability is selected. Recently, it was proposed to use these softmax predictions as a means to get more insights into the continuous dynamics of the sleeping brain [1]. Plotted over time, these probabilities are called a hypnodensity, as opposed to the conventional hypnogram that displays only one stage per data window. A hypnodensity has great potential to provide more information about the sleeping brain, since it facilitates visualization of non-abrupt transitions, and co-occurrence of characteristics belonging to multiple stages at the same time. The question now arises how to interpret this hypnodensity. Does, e.g., a 80-20% division between N2 and N3 reflect a 80% certainty for N2 classification, or does it mean that 80% of the characteristics are typical for N2, while 20% of the characteristics belong to N3? To investigate this, we first create a synthetic data set inspired by PSG recordings to enable controlled experiments, in which the ground-truth signal model is known. We model the selection of a sleep stage as a function of the contribution of five underlying (abstract) signals aggregating the characteristics belonging to one sleep stage. We conclude that a hypnodensity displays the distribution from which annotations were (implicitly) drawn during annotation of the training set [2]. Moreover, we found that, despite the additional information available in a hypnodensity, potentially relevant information is still being lost due to the non-linear normalizing softmax layer, and the label-dependency of supervised training. As a solution, we propose to consider pre-softmax predictions and unsupervised training.

end %-->