Conference Agenda

Mathematical patient models play a profound role in the managing and understanding of anaesthesia. This article proposes a patient model for simulating anaesthetic uptake, i.e. the storing of anaesthetic gases in the human body, thus extending a well known four compartment model from literature. Real time data is used for the simulation, from which the sevoflurane concentration is extracted. The exhaled concentrations from the model and the data are compared with a cost functional sum of least squares and an error bound around the measured exhaled concentration. The literature model is extended with ratios of compartmental perfusion, the introduction of coefficients representing diffusion and the change of the solubility coefficient. The extended and the literature model are applied to five training data sets, with prior optimization of the parameters that have the most impact. Afterwards, both models are validated on five additional data sets. The results of the evaluation parameters of the extended model are mostly better than the values of the literature model and thus show, that the model is able to represent anaesthetic uptake on adults. This work lays a basis for validating a model with patient data.

Deep learning models for the classification of electrocardiograms (ECGs) are able to learn disease-specific patterns, but they are rarely implemented in medical practice due to their "black box" nature. Post-hoc explainable artificial intelligence (XAI) methods compute regions of interest (ROI) which are of importance for a model's decision making. However, it needs to be further analyzed whether a model focuses on the morphological or rhythmical information within the ROIs.

We evaluate a pre-trained ResNet for sinus bradycardia (SB) and sinus tachycardia (ST) classification on the PTB-XL dataset using the XAI method Integrated Gradients. We compare the confidence of the model predictions to ECG features used by clinicians using correlation analysis.

Correlation is highest for RR intervals and atrial as well as ventricular heart rates, with the majority exceeding clinical thresholds for both disorders, indicating that the model learned rhythmical features. Except for QT intervals in ST classification, morphological features such as duration and amplitudes of P-/T-waves do not show any correlation.

Background:

Heart rate variability (HRV) is a physiological variation of intervals between consecutive heartbeats and a prognostic marker in numerous cardiovascular and non-cardiovascular conditions and it could be used to differentiate between healthy subjects and patients affected with cardiovascular conditions.

Objective:

The purpose of this study was to evaluate a method for differentiation between healthy volunteers and patients with valvular heart diseases based on time domain and frequency domain heart rate variability indices calculated from electrocardiographic (ECG) signals and three linear models (Logistic Regression, Ridge Regression, Support Vector Machine with a linear kernel).

Material and Methods:

The study was carried out on two publicly available data sets with concurrent ECG, SCG, and GCG signals (Mechanocardiograms with ECG reference and An Open-access Database for the Evaluation of Cardio-mechanical Signals from Patients with Valvular Heart Diseases) that consist of 29 recordings and 30 recordings, respectively. The HRV analysis was carried out according to the current recommendations and included the following indices: AVNN, SDNN, RMSSD, pNN50 (in time domain), VLF, LF, HF, and LF/HF (in frequency domain). We set the maximum number of iterations to 1000 and a constant random seed of 20240205 with 5-fold stratified cross-validation.

Results:

The highest sensitivity, PPV, accuracy and F1 score were observed for Logistic Regression (0.8810, 0.8819, 0.8814, 0.8812, respectively), followed by Ridge Regression (0.8805, 0.8858, 0.8814, 0.8808, respectively), and the lowest were observed for the linear SVM (0.8310, 0.8318, 0.8305, 0.8305, respectively).

Discussion and Conclusion:

The results showed that it is possible to distinguish healthy volunteers and patients with linear classifiers and time domain and frequency domain HRV indices obtained from ECG signals with a decent performance. However, a higher number of false negatives (6) than of false positives (3) in logistic regression and ridge regression indicates similar values of HRV indices in a limited number of healthy subject and VHD patients.

Introduction:

Wearable devices integrating photoplethysmography (PPG) sensors enable noninvasive monitoring of vital parameters such as heart rate and blood oxygen saturation. However, the reliability of the measurements is compromised by motion artifacts. Deep learning approaches, particularly Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), have shown potential in detecting corrupted segments and recognizing elusive patterns but are often evaluated on different non-public datasets, leading to incomparable results. This study assesses the fair comparative performance of these algorithms on the same public dataset for real-time signal quality classification.

Methods:

We employed the PPG-DaLiA dataset, comprising 36 hours of data from 15 subjects performing various activities. Five algorithms were adapted from the literature and implemented for real-time operation: two one-dimensional CNNs (custom by Zargari et al. and InceptionNet) and two two-dimensional CNNs (custom by Roh and Shin, and VGG-19), as well as a Long Short-Term Memory (LSTM) RNN. The models were trained using a binary cross entropy loss function with the Adam optimizer, oversampling for class-balancing, an 80/20 dataset split for training and testing, and Optuna for hyperparameter tuning targeting accuracy.

Results:

The accuracy and sensitivity values obtained for all algorithms on the public dataset were significantly lower than those presented by their authors for their own data. Besides, the biggest models (VGG-19 and InceptionNet) failed to converge. Nevertheless, the RNN LSTM model demonstrated the highest accuracy at 89.28%, followed by the 2D CNN at 87.5%. The RNN also showed the highest sensitivity (92.5%) and the fastest classification time of 0.0039 seconds, making it feasible for real-time applications.

Conclusion:

In conclusion, the performance of deep learning algorithms for motion artifact detection in PPG varies across datasets. Still, the RNN LSTM showed the highest accuracy and fastest classification on the PPG-DaLiA dataset, making it a promising option for real-time signal quality classification.

Introduction

Modern brain-computer interfaces (BCI) allow precise and reproducible user-environment interactions using electroencephalography (EEG) recordings. In gaming applications, Perri Karyal uses the EMOTIV EPOC X EEG-headset (14-channel, wireless) to control game avatars. The online data stream processing takes place on comput-er. Our research is concerned with achieving a high level of user comfort with real-time processing on a mobile solution. This requires local signal processing and machine learning on an embedded platform.

Methods

Our objective is to achieve an end-to-end signal processing for controlling games with EEG recordings using comfortable dry Flower electrodes. Data acquisition is conducted using MentaLab Explore+ (12 channels, 1 kHz, wireless). The online data processing includes digital filtering and action recognition using AI accelerators on a Field Programmable Gate Array (FPGA) within the ElasticAI ecosystem. Therefore, AI models are trained using PyTorch, and transferred to the FPGA. To improve the robustness of our algorithm in gaming scenarios, we pro-vide an extended dataset by combining EEG recordings and 9-axis acceleration data for artifact detection.

Results

We successfully acquired our first datasets using EEG recordings with dry Flower electrodes and MentaLab device. This dataset includes eight recording sessions from six candidates to imagine the movement of the left or right hand. We assessed data quality by employing EEG-Net to classify the two classes, which achieves an accu-racy of 97%. It is worth noting that the impact of pre-processing (digital filtering, pruning channels, reducing sampling rate) has been explored and shown to maintain the same level of accuracy.

Conclusion

Our first approach of building an EEG-based gaming controller allows to generate first datasets and to perform online data processing and ML inference on a workstation. In future, the dataset is extended to enable motion de-tection for gaming scenarios like 'Pac-Man'. Moreover, the AI models used are trained and transferred to the local hardware for controlling more complex video games.

Extended sitting in office can lead to health issues. To alleviate tension, it’s recommended to change positions every 30 minutes. In this study, we monitor the pressure pattern on a pressure sensor mat (PSM) and classify sitting postures comparing machine learning (ML) algorithms including kNN, SVM and Random Forest (RF) with and without feature extraction.

Two datasets were collected with the PSM: 1) ten participants taking five sitting positions; 2) five participants assuming eight postures on 7-second videos with 10 images per second. Dataset 1 underwent 10-fold cross-validation with a) ran-dom train-test-split; b) training on data from nine participants and testing on the remaining participant. For Dataset 2 images were divided into 4 quadrants, and white pixel distribution and distances from quadrant centres of mass (COM) and absolute COM were extracted, employing also a 10-fold cross-validation strategy. Finally, the model trained on Da-taset 2 was employed to classify the unknown data within Dataset 1.

Accuracies in Dataset 1 decreased from maximum 93.3 % (SVM) for method a) to 72.0 % (SVM) for method b). In Dataset 2 all classifiers delivered 99.9 % accuracy for raw data and RF the best accuracy (99.6 %) for the extracted data. When training the model with the data from Dataset 2 and testing with Dataset 1, best accuracies decreased to 52.3 % (kNN) for raw data and 56.0 % (RF) for the extracted data.

Different evaluation scenarios yield varying accuracies for posture classification using ML algorithms. Feature extraction, like 4-quadrants approach, improves performance. Results decreased for unknown participants in both datasets. Indeed, pressure distribution images on the PSM are influenced by individual characteristics such as gender, age, body structure, and habits. Thus, a calibration process for new users is essential for accurate posture estimation, requiring further research to determine the minimum calibration data needed for each individual.