Pre-driving fatigue screening from short-term heart rate variability with subject-independent validation

Abstract

This study evaluates fatigue screening from 30-second electrocardiogram (ECG) recordings using short-term heart rate variability (HRV) features in a pre-driving context. The dataset comprises 99 participants (one session each) with fatigue labels derived from the Karolinska sleepiness scale (KSS), where the primary label (K1) defines non-fit as KSS ≥ 7. A subject-independent logistic-regression model was trained under a leave-one-subject-out (LOSO) scheme. Probabilities were calibrated using Platt scaling and evaluated through threshold-free metrics (receiver operating characteristic (ROC)-area under the curve (AUC), precision-recall (PR)-AUC) as well as calibration performance using the Brier score. The model achieved ROC-AUC =0.687 (95% confidence interval: 0.591–0.776), PR-AUC =0.621, and a Brier score of 0.200. At the operating threshold t = 0.255, the model achieved sensitivity of 1.000 with no false negatives, while specificity remained 0.091 (95% confidence interval: 0.030–0.140). Reliability analysis indicated reasonable calibration in the operational probability range. These findings support short-term HRV derived from ECG as a screening tool that prioritizes avoiding missed non-fit cases, paired with a triage scheme (fit/review/non-fit) to manage uncertainty near the decision threshold. Future work should incorporate ECG morphology and signal quality cues and aim to improve specificity without sacrificing sensitivity.