13:30   Lung and Respiration
An Adversarial Learning Approach to Generate Mechanical Ventilation Waveforms for Asynchrony Detection
L. Hao, T.H.G.F Bakkes, A. van Diepen, A.J.R de Bie Dekker, P.H. Woerlee, M. Mischi, S. Turco
Abstract: Patient-ventilator asynchrony (PVA) during mechanical ventilation can lead to pulmonary damage, complications and even increase mortality. A machine learning method is proposed in [1] for detection and diagnosis of PVAs. Proper training of these machine learning detection models requires a large amount of labelled data. However, the availability of clinical data with annotations is limited and simulated data is not ventilator specific. The influence of the hardware, servo control algorithms, and filtered noise is not included in the simulations. In our research, a new framework based on a generative adversarial network, called VentGAN, is developed to improve simulated data by learning the ventilator fingerprints from clinical unlabelled data, with the goal of providing a more powerful training dataset for algorithms as in [1]. The input of the network is the simulated data obtained by the patient-ventilator model developed in [2]. The VentGAN model consists of a generator and a discriminator. The generator reproduces the characteristics of the clinical data while preserving the annotations of the simulated ones. The role of the discriminator is to distinguish clinical data from generated data, thus pushing the generator to produce more realistic data. For this purpose, we construct the structure of the generator based on a U-Net architecture and develop the discriminator from a convolutional neural network. The loss function of the generator is modified to enforce similarity between the generated and simulated data. Qualitative validation was performed visually by the clinical expert. We compared the generated and simulated data to check whether the inspiratory and expiratory timings in the simulated data were altered when passing through VentGAN. Moreover, we verified that the characteristics of the clinical data, normally not present in the simulated data, were learned by the network and were visible in the generated data. Our qualitative analysis suggests that VentGAN is able to produce more realistic ventilatory waveforms, while maintaining the annotations unchanged. For a quantitative validation of our method, we will compare the PVA detection accuracy of the model in [1] by using as a training dataset the simulated data before and after passing it through VentGAN. The ability of VentGAN to produce more realistic simulated data was demonstrated qualitatively. This framework can be integrated with the PVAs detection models to further improve the accuracy of PVA detection, possibly decreasing the risk of pulmonary damage during mechanical ventilation. [1] Bakkes T H G F, Montree R J H, Mischi M, et al. A machine learning method for automatic detection and classification of patient-ventilator asynchrony[C]//2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE, 2020: 150-153. [2] van Diepen A, Bakkes T H G F, De Bie A J R, et al. A Model-Based Approach to Synthetic Data Set Generation for Patient-Ventilator Waveforms for Machine Learning and Educational Use[J]. arXiv preprint arXiv:2103.15684, 2021.
The use of simulated data can improve models for automated detection of patient-ventilator asynchrony
Tom Bakkes, Anouk van Diepen, Ashley De Bie, Massimo Mischi, Pierre Woerlee, Simona Turco
Abstract: Patient-ventilator asynchrony (PVA) is a hard-to-catch adverse event in mechanically ventilated critically ill patients and is associated with increased mortality [1]. There is a need for automated detection of asynchronies since current solutions are suboptimal. In previous research, a method for the automated detection of patient-ventilator asynchrony (PVA) was developed [2]. Proper validation of these methods is difficult because this requires independent datasets containing accurately labeled PVAs. However, these datasets are difficult to obtain, since labeling the data is labor-intensive, time-consuming, and prone to human errors. In contrast, simulating ventilation data allows for rapid generation of accurately labeled data and comprehensive datasets, as it does not require measurements on actual patients, and the timings needed to detect PVAs are inherently known. In this study, we investigated the feasibility of obtaining accurately labeled ventilation data from simulations, and the generalizability of the previously developed method to this data. A simulator developed in [3] was used to obtain the simulated data. In total, the simulation generated 64898 breathing cycles. The clinical data was the same data utilized in previous research, which consisted of 15 patients and contained 4275 breathing cycles [2]. The model was trained and tested in three different approaches. First, we performed cross-validation as in [2] with only clinical data. Second, the model was trained on clinical data and tested on simulated data. Finally, the model was trained on simulated data and tested on clinical data. The resulting precision and recall were respectively 97.5% and 98.0% for the first approach, 94.3% and 93.5% for the second approach, and 94.5% and 97.8% for the third approach. The overall performance was excellent, with the second and third approaches performing remarkably high, almost similar to the first approach in which the algorithm was both trained and tested on clinically labeled data. The second and third approaches confirm that the simulations generate accurate ventilation data for which the detection can be easily generalized to different datasets. This approach with simulated datasets will therefore help to accurately detect PVAs in real-time at the bedside of the patients. [1] C. de Haro et. Al., Intensive care Med Exp. (2019) [2] T. Bakkes et. Al., EMBC. (2020) [3] A. van Diepen et. Al., arXiv (2021)
Simulating the effects of mechanical ventilation in realistic lung models
Sjeng Quicken, Milou van Mil, Eline van Engelen, Frans van de Vosse
Abstract: Background: Intensive care patients often require mechanical ventilation (MV). During MV an external ventilator applies a pulsatile pressure to the trachea to ventilate the patient. While MV is crucial for patient survival, it typically induces ventilator-induced lung injury (VILI) which can potentially be life-threatening. VILI is primarily caused by regional alveolar collapse or overdistension due to pressure inhomogeneities in the ventilated lung. During the COVID-19 pandemic a surge in patients requiring MV was observed, resulting in an increased incidence of VILI. Hence, there is a clear need for strategies that minimize VILI. A widely adopted approach to combat VILI during the COVID pandemic is prone position ventilation, as opposed to regular supine position ventilation. In this research we develop a computational framework to simulate flow and alveolar mechanics during MV in realistic lung geometries, in order to investigate the effectiveness of strategies to reduce VILI. Here, we introduce the simulation framework and apply it to investigate the potential benefit of prone position ventilation. Methods: A realistic lung geometry was generated using data from the LIDC-IDRI dataset (1). First, the lung lobes and large airways were segmented from CT data, after which smaller airways were automatically generated using a modified implementation of the lobe-filling algorithm by Tawhai et al. (2). Individual airways were modelled using a non-linear resistor. Terminal airways were truncated using a lumped-parameter alveolar model. A typical MV pressure curve was prescribed at the trachea. Pressure distributions representative of either prone or supine ventilation were prescribed at the alveolar elements. Results: The generated respiratory tree filled the segmented lobes and consisted of 24·10³ airways and 12·10³ alveolar elements. On average, alveolar elements in the prone ventilation simulation experienced considerably lower strains and less strain heterogeneity when compared to supine ventilation results, indicating less alveolar overdistension. Conclusion: The framework allows for generating realistic lung models of MV. Our results present a potential mechanism for the benefit of prone ventilation to combat VILI. References 1. Armato et al., Med. Phys. 38, 915–931 (2011). 2. Tawhai et al., Ann. Biomed. Eng. 28, 793–802 (2000).
EMG-based monitoring of respiratory effort using LSTM networks
Hongji Xu, Elisabetta Peri, Xi Long, Sebastiaan Overeem, Johannes P. van Dijk, Massimo Mischi
Abstract: BACKGROUND AND AIM: Obstructive sleep apnea (OSA) is a common disorder in the population, affecting between 2-15% of middle-aged adults and more than 20% of the elderly. The diagnosis of OSA is ideally based on direct measurement of respiratory effort, but the gold standard to assess respiratory effort is invasive, namely esophageal pressure (Pes). Therefore, the clinical practice mostly relies on surrogate measures, such as thoracic and abdominal movement. Respiratory muscle activity measured by electromyogram (EMG) could be a non-invasive alternative to rather directly monitor respiratory effort. However, interpretation of EMG signals with respect to the Pes gold standard is lacking as of yet. The aim of this work is to estimate respiratory effort from diaphragmatic EMG (dEMG) and sternocleidomastoid EMG (scmEMG) based on Long Short-Term Memory (LSTM) neural networks. METHODS: Thirteen healthy volunteers participated in a study carried out at Kempenhaeghe Sleep Medical Center, Heeze, the Netherlands. To simulate different types of respiratory effort, participants were asked to breathe through an obstacle that increased the resistance to the flow while Pes, dEMG and scmEMG signals were acquired (sampling frequencies = 128 Hz). After preprocessing the signals (with a notch filter at 50 Hz and a 4th-order Butterworth band-pass filter at 5-250 Hz for EMG and at 0.05-10 Hz for Pes), we extracted a set of features to train a bidirectional LSTM model. Features were extracted separately from dEMG and scmEMG using various methods including variational mode decomposition, area under the curve, envelope (rectified signals), sample entropy, and Mel-frequency cepstral coefficients of signals. The LSTM consists of three consecutive recurrent layers with the tanh activation function and mean square error loss function. Dropout regularization was used after each layer to reduce overfitting. The data set was divided into a training set (80%) and a test set (20%). Respiratory effort was also estimated using the conventional envelop computation by 200 ms moving average filter. The Spearman correlation coefficient (R) and the normalized mean error (NME) between the estimated respiratory effort and the Pes were computed to assess the performance of the LSTM model. The NME range was [0,2] according to the normalization. A paired t-test was used to compare the results obtained by the LSTM model with respect to the conventional envelope. RESULTS: In relation to the Pes signal, the respiratory effort estimated with the conventional envelope has an NME = 0.44±0.12 and R = 0.68±0.09. Using the LSTM model the error and correlation improved significantly (p-value<0.05) to NME = 0.36±0.07 and R = 0.74±0.08. DISCUSSION AND CONCLUSIONS: The proposed method based on LSTM has the potential of estimating the respiratory effort from EMG signals. The error is significantly lower than the error obtained with the envelope method. However, this work mainly used a lab-test data set and healthy subjects, therefore, further work is needed to verify the method in OSA patients.
Development of a new portable headset to monitor energetic load without a mouth mask
Charissa Roossien, Bart Verkerke, Michiel Reneman
Abstract: A portable headset has been developed to analyze breathing gases and establish the energetic workload of physically active workers more comfortably in comparison to the current golden standard, a mouth mask. This proof-of-concept study aimed to investigate the validity of the portable headset compared to (1) the medical indirect calorimetry method using a mouth mask and (2) oxygen consumption (VO2) estimation based on heart rate, and to explore the user experience of the developed headset system. Fifteen subjects performed a submaximal cycling test twice, once with the headset, and once with the two references systems, a mouth mask and heartrate monitor. In comparison to indirect calorimetry, good correlations were observed for the VO2, carbon dioxide production (VCO2) and exhaled volume (Ve). The headset tended to underestimate VO2, VCO2 and Ve at low intensities and to overestimate it at higher intensities. The headset was more valid for estimating VO2 than estimates based on heart rate. The subjects preferred the headset over the mouth mask because it was more comfortable, did not hinder communication and had lower breathing resistance. The headset appears to be useable for monitoring the energetic workloads of physically active workers, being more valid than heart rate monitoring and more practical than indirect calorimetry with a mouth mask. The present version is not yet completely valid, but its potential is supported and indicates opportunities for further professionalization. This professionalization step is on-going in collaboration with the company Relitech who is specialized in developing innovative and reliable medical devices, aiming to make this headset commercially available. Another design step and further validation studies are needed before implementation in the workplace. There is increasing interest in this breathing-gas analyzing headset to objectively monitor physiological responses of individuals. This system is likely to be of interest as a low-level, comfortable and easy-to-use device for monitoring the physical fitness of subjects in multiple settings, including working, (remote) (occupational) healthcare, rehabilitation and sports settings. It can be carried out in users’ actual environment over longer periods of time. This headset could fill a gap in the existing range of instruments for measuring energy consumption.
Towards computer aided design of passive Heat and Moisture Exchangers
Maartje Kamphuis-Leemans, Maarten van Alphen, Wim Vallenduuk, Richard Dirven, Michiel van den Brekel, Saar Muller
Abstract: Background: Small passive Heat and Moisture Exchangers (HMEs) are standard treatment for pulmonary rehabilitation after a total laryngectomy. These HMEs consists of a plastic cassette containing a foam material impregnated with hygroscopic salt (wet core), which acts as a condensation and evaporation surface. During each breathing cycle, part of the moisture from the exhaled air is stored on the wet core and used to humidify the inhaled air (water exchange). Currently, the passive HMEs’ performances are not as high as the upper airways before laryngectomy, and the influence of the HME design on its performance is difficult to predict. We developed a numerical HME model to improve the physical understanding and design of these small passive HMEs. Methods: The numerical HME model was implemented in Matlab 2020b (Mathworks Inc. USA). The physical processes inside the HME are described by four discrete physical equations for the conservation of mass and energy in the air and wet core. The HME model’s physical parameters were tuned using a basic experimental data set of the HME’s performance (water exchange and temperature data) at standardized tidal volumes and flows. The tuned HME model was validated by comparing its performance to an additional experimental data set. Finally, the HME model was applied to HME design variations and environmental conditions outside the scope of the experimental data. Results: The tuning step showed a minimum for the physical parameter settings for which the HME model’s water exchange is in equilibrium and has a good correspondence to the experimental data. The HME model’s performance for tidal volume and flow dependency explains the trends as found in volunteer studies. The physical understanding as found using the model leads to the following design recommendations: the wet core’s mass exchange efficiency and heat capacity are the driving factors of the HME’s water exchange. The wet core’s heat conductivity is similar to the heat conductivity of water, and its influence on the HME’s performance is negligible. Conclusions: The numerical HME model improves the physical understanding of small passive HME’s, predicts the HME’s performance and aids the developmental process towards better performing HMEs.

end %-->