Machine learning and evolutionary computation for increased reliability of wearable devices for monitoring cardiac arrhythmias in the wild.

Abstract

Continuous analysis of electrocardiogram (ECG) signals is essential for diagnosing cardiovascular diseases, particularly arrhythmias, which contribute significantly to global mortality. With the increasing use of wearable devices for non-clinical cardiac monitoring, real-time ECG processing encounters challenges due to memory constraints, particularly when memory is wasted by recording false positives. Noise in real-life ECG monitoring, such as motion artifacts, complicates automatic arrhythmia detection and recording, and calls for efficient methods to selectively filter out noisy signals that may be falsely classified (and recorded) as arrhythmias in devices with limited storage. This study develops and systematically compares machine learning and deep learning approaches for ECG noise detection using real-world wearable data acquired outside controlled clinical environments. By employing various models for classification, including Random Forest, Support Vector Machine, Multi-Layer Perceptron, Gradient Boosting and Convolutional Neural Networks, we investigated four approaches, using as input features: i) raw signal data, ii) Heart Rate Variability-related time-domain features, iii) a combination of Heart Rate Variability-related time-domain and frequency-domain features, and iv) the parameters of a Genetic Programming-based reconstruction algorithm, analyzing both memory and computational requirements and classification performance of the corresponding architectures. Our findings indicate that the Multilayer Perceptron model, combined with Heart Rate Variability-related time- and frequency-domain features, offers an optimal trade-off between accuracy, specificity, sensitivity, and F1-score. This study presents a scalable classification approach specifically designed for real-life arrhythmia detection in wearable devices, demonstrating the potential to enhance continuous cardiac monitoring and to facilitate the integration of such algorithms into wearable devices suitable for clinical and real-world settings.