Study Objectives
Heart rate variability (HRV)-based machine learning models hold promise for real-world vigilance evaluation, yet their real-time applicability is limited by lengthy feature extraction times and reliance on subjective benchmarks. This study aimed to improve the objectivity and efficiency of HRV-based vigilance evaluation by associating HRV and behavior metrics through a sliding-window approach.
Methods
Forty-four healthy adults underwent psychomotor vigilance tasks under both well-rested and sleep-deprived conditions, with simultaneous electrocardiogram recording. A sliding-window approach (30s length, 10s step) was used for HRV feature extraction and behavior assessment. Repeated-measures ANOVA was used to examine how HRV related to objective vigilance levels. Stability selection technique was applied for feature selection, and the vigilance ground truth—high (fastest 40%), intermediate (middle 20%), and low (slowest 40%)—were determined based on each participant's range of performance. Four machine-learning classifiers—k-nearest neighbours, support vector machine (SVM), AdaBoost, and random forest—were trained and tested using cross-validation.
Results
Fluctuated vigilance performance indicated pronounced state instability, particularly after sleep deprivation. Temporary decrements in performance were associated with a decrease in heart rate and an increase in time-domain heart rate variability. SVM achieved the best performance, with a cross-validated accuracy of 89% for binary classification of high versus low vigilance epochs. Overall accuracy dropped to 72% for three-class classification in leave-one-participant-out cross-validation, but SVM maintained a precision of 84% in identifying low-vigilance epochs.
Conclusions
Sliding-window-based HRV metrics would effectively capture the fluctuations in vigilance during task execution, enabling more timely and accurate detection of performance decrement.
