Generalizing ECG-based fatigue detection across fatigue-induction paradigms using deep learning

Introduction Fatigue and sleepiness can arise under different experimental and behavioral conditions, yet ECG-based detection models are often evaluated only within the dataset used for development. This study investigated the participant-independent and cross-dataset performance of two ECG-derived approaches: a 1D CNN based on six HRV features and a dual-input CNN-BiLSTM combining RR interval sequences with statistical HRV descriptors. Methods Models were developed using the DROZY database, in which sleepiness was induced by prolonged wakefulness. Internal performance was assessed using five-fold subject-wise evaluation on 25 recordings from 14 participants. The models were then applied without retraining or parameter adaptation to the independent FatiQData dataset comprising 20 participants performing a prolonged monotonous driving task. Results On DROZY, the HRV-based 1D CNN achieved 60.0% accuracy, 61.5% F1-score, and a ROC-AUC of 0.593, whereas the dual-input CNN-BiLSTM achieved 68.0% accuracy, 73.3% F1-score, and a ROC-AUC of 0.753. In FatiQData, Stanford Sleepiness Scale scores increased significantly from 2.40 ± 0.82 to 4.15 ± 0.81. For binary external discrimination, the HRV-based model achieved 60.0% accuracy, 69.2% F1-score, and a ROC-AUC of 0.766, while the dual-input model achieved 65.0% accuracy, 77.4% F1-score, and a ROC-AUC of 0.750. Discussion The results indicate that ECG-derived HRV and RR dynamics retain information related to subjective sleepiness, while participant-independent and cross-dataset performance remains moderate. These findings emphasize the importance of subject-wise validation, independent external testing, and cautious interpretation across heterogeneous experimental conditions.