Physiological Signals as Predictors of Mental Workload: Evaluating Single Classifier and Ensemble Learning Models

Abstract

With a growing emphasis on cognitive processing in occupational tasks and the prevalence of wearable sensing devices, understanding and managing mental workload has broad implications for safety, efficiency, and well-being. This study aims to develop machine learning (ML) models for predicting mental workload using Heart Rate Variability (HRV) as a representation of the Autonomic Nervous System (ANS) physiological signals. A laboratory experiment, involving 34 participants, was conducted to collect datasets. All participants were measured during baseline, two cognitive tests, and recovery, which were further separated into binary classes (rest vs workload). A comprehensive evaluation was conducted on several ML algorithms, including both single (Support Vector Machine/SVM and Naïve Bayes) and ensemble learning (Gradient Boost and AdaBoost) classifiers and incorporating selected features and validation approaches. The findings indicate that most HRV features differ significantly during periods of mental workload compared to rest phases. The SVM classifier with knowledge domain selection and leave-one-out cross-validation technique is the best model (68.385). These findings highlight the potential to predict mental workload through interpretable features and individualized approaches even with a relatively simple model. The study contributes not only to the creation of a new dataset for specific populations (such as Indonesia) but also to the potential implications for maintaining human cognitive capabilities. It represents a further step toward the development of a mental workload recognition system, with the potential to improve decision-making where cognitive readiness is limited and human error is increased.