Study of Dry Electrode ECG Signal Characteristics towards Explainable AI

Single-lead electrocardiograms (ECG) are a reliable modality for long-term remote monitoring applications.The development of wearable ECG devices requires understanding of the inherent characteristics of the acquired signals to create ECG centric signal processing and machine learning (ML) algorithms.However, most single-lead ECG wearable signal analysis blindly follows algorithms that were built for clinical ECG.This approach yields inappropriate choice of filtering and ML models that create inaccuracies and increase unreliability.A framework to examine signal characteristics to guide development of explainable artificial intelligence (xAI) for newborn ECG monitoring applications is proposed.3D printed dry electrode newborn ECG signals (n = 19) are examined under different testing paradigms.The signals undergo linearity, stationarity, and Gaussian testing.Also, the signals undergo cubic Hermite spline interpolation to show appropriateness of representing signals through spline polynomial functions.Across all newborn ECG signals non-linear, non-stationary, and hyper Gaussian behaviors are observed.By understanding the signal characteristics, adaptive algorithmic development is realizable especially for remote ECG monitoring applications.