Machine learning workflow for edge computed arrhythmia detection in exploration class missions

Long-duration deep-space missions present significant healthcare challenges due to limited access to ground-based medical assistance. The extended communication delays and the detrimental effects of microgravity on the cardiovascular system, including increased risk of arrhythmias, necessitate on-board diagnostic capabilities. This study addresses this need by developing a machine learning (ML) model for real-time arrhythmia detection deployable on wearable edge devices using the Open Neural Network Exchange (ONNX) format. The focus is on AFIB and AFL, prevalent tachycardias treatable with medications manageable by medically trained astronauts. Early detection of arrhythmias is crucial to prevent mission-threatening complications. The use of edge computing, facilitated by ONNX, ensures the system's independence from ground resources and enables prompt diagnosis and treatment even with significant communication delays. The study prioritizes the interpretability of the ML model to ensure its usability by both medically trained and untrained crew members, balancing speed and transparency in decision-making.

The literature highlights the detrimental effects of spaceflight on the cardiovascular system, including decreased left ventricular mass, atrial structural changes, and loss of cardiac contractility, leading to an increased risk of arrhythmias. Studies on previous space missions show arrhythmias as a prevalent medical issue. Existing wearable systems and telemedicine solutions are limited in deep-space scenarios due to communication delays. This necessitates the development of autonomous on-board diagnostic tools, emphasizing the need for edge computing and interpretable ML models. While various biometric models exist, their adaptability to resource-constrained edge computing architectures in deep space remains a challenge. This study leverages the ONNX format to address this limitation, focusing on interpretable ML for increased trust and usability by medical personnel during deep-space missions.

The study utilized a modular seven-component pipeline: Databases, Pre-processing, Denoising, Feature Extraction, Model Development, Evaluation, and Edge Inference. Three noise databases (MIT-BIH Noise Stress Test, MIT-BIH ST Change, and EPHNCORD) and four arrhythmia databases (MIT-BIH Arrhythmia, MIT-BIH Atrial Fibrillation, MIT-BIH Normal Sinus Rhythm, and Long Term Atrial Fibrillation) were employed. ECG recordings were pre-processed into 30-second normalized samples. Denoising involved Fast Fourier Transform (FFT) and Variational Frequency Mode Decomposition (VFMD). Seventeen features were extracted: five heart rate variability (HRV) features and twelve morphological features. A decision tree classifier was selected and optimized using a triple-nested cross-validation strategy, balancing hyperparameter tuning, parameter optimization, and feature selection. The resulting model was converted to ONNX format for edge inference on an Android device. Model performance was evaluated using metrics including accuracy, recall, precision, F1-score, and Area Under the ROC Curve (AUROC).

The optimized decision tree classifier achieved a macro F1-score of 0.899, with individual F1-scores of 0.993 for NSR, 0.938 for AFIB, and 0.767 for AFL. The AUROC values were 0.988 for NSR, 0.902 for AFIB, and 0.912 for AFL, surpassing a logistic regression baseline model. The model's inference time was 9.2 seconds per sample. Key features included median P-wave amplitudes, PRR40, and mean heart rates. The ONNX-translated model demonstrated robust performance even in noisy conditions, showing high accuracy, precision, and recall for detecting the three studied rhythms. The confusion matrices for both training and test sets illustrated the high accuracy of the model in classifying NSR, AFIB, and AFL, with only minor misclassifications.

The study's findings demonstrate the feasibility and effectiveness of using an ONNX-based ML pipeline for real-time arrhythmia detection on edge devices during deep-space missions. The high accuracy, speed, and interpretability of the model make it suitable for use by both medically trained and untrained crew members. The modular design of the pipeline allows for easy integration of new features and models, improving adaptability and potential for future expansions. The model's performance surpasses that of existing approaches, particularly regarding speed and edge computing compatibility. The reliance on interpretable features familiar to cardiologists enhances confidence and trust in the system's output.

This research presents a novel and effective approach to arrhythmia detection in the challenging context of deep-space exploration. The developed ML pipeline, optimized for edge computing through ONNX, offers a robust and reliable solution for real-time diagnosis, supporting medical autonomy and mission success. Future work could focus on expanding the model to include other arrhythmias, integrating additional biometric data, and validating its performance using in-space ECG data.

The study's reliance on Earth-based ECG data may not perfectly capture the unique conditions of the spaceflight environment. The limited sample size for AFL could affect the model's generalizability to this specific rhythm. Further research is needed to validate the model's performance in actual space missions.