Machine learning based ECG analysis with data augmentation for fatigue status detection in elderly individuals

Abstract

Fatigue detection in elderly individuals is critical for preventing health complications and enhancing quality of life. Electrocardiogram (ECG) signals offer a non-invasive way to capture fatigue-related physiological changes, but few solutions are lightweight, interpretable, and deployable in real-world settings. This study proposes a portable, low-cost ECG-based system for real-time fatigue detection using an extended AD8232 module and ESP32 microcontroller. A dataset of 6304 ECG segments was collected from 100 elderly participants. Data augmentation (SMOTE and Gaussian noise) was applied. GRU, LSTM, and RNN models were trained and validated with subject-wise splitting. The GRU model achieved 97.96% accuracy, 0.99 AUC, and 0.98 F1-score, significantly outperforming LSTM, RNN, Random Forest, and Logistic Regression baselines. This work demonstrates the feasibility of low-cost, embedded ECG-based fatigue detection and provides interpretable, generalizable results suitable for elderly care applications. Further clinical validation is planned.