Machine Learning Classification of Physiological Dynamics During Standardized Task-Demand Transitions.

A major barrier to developing adaptive human-machine systems is the lack of interpretable physiological markers for characterizing physiological responses associated with changes in task demands. The aim of this study was to evaluate the feasibility of machine-learning classification of physiological responses during simple and complex task stages and two standardized task-demand transitions-baseline/rest-to-simple and simple-to-complex-in cognitive and motor-cognitive protocols. Sixty-nine healthy volunteers completed cognitive tasks involving normal and 180° inverted-text reading and motor-cognitive tasks involving simple and complex movement sequences performed with the non-dominant hand. Each protocol comprised a 1 min baseline followed by three consecutive blocks, each including a 1 min simple task, a 1 min complex task, and a 1 min rest period. Photoplethysmography (PPG), skin conductance (SC), and abdominal respiration were recorded continuously. Pulse-to-pulse (PP) intervals were extracted from the PPG signal, and conventional heart rate variability (HRV) indices, including mean RR interval, HR range, SDNN, RMSSD, Total Power, and SD2, were calculated in Kubios HRV from the PPG-derived interval series. Logistic Regression, Random Forest, and Support Vector Machine models were trained using either absolute physiological values or dynamic features (Δ%) calculated for the baseline/rest-to-simple and simple-to-complex transitions. Models based on dynamic features consistently demonstrated higher classification performance than those based on absolute physiological values. For the motor-cognitive protocol, Random Forest achieved an accuracy of 0.726 and a ROC-AUC of 0.730, whereas for the cognitive protocol, Support Vector Machine achieved an accuracy of 0.655 and a ROC-AUC of 0.722 on the held-out participant-level test set. The most informative features included mean RR interval (derived from PPG), Total Power, RMSSD, SDNN, SD2, and HR range in both protocols. Comparable feature-importance patterns across protocols suggest overlap in the physiological variables contributing to classification under the standardized experimental conditions. These findings are limited to the fixed transition sequence used in the present study and require confirmation in randomized or counterbalanced designs.