Abstract

Objectives: This study aims to develop an optimized machine learning (ML) framework for predicting stress levels via physiological signals collected during sleep via Internet of Medical Things (IoMT)-enabled biosensors. The primary goal is to increase the accuracy and efficiency of stress prediction by identifying the most significant features that influence stress classification. Method/Analysis: The proposed framework employs two feature selection methods: particle swarm optimization combined with the whale optimization algorithm (PSO-WOA) and an enhanced version incorporating Lévy flight (PSO-WOA with Lévy flight). These methods are designed to reduce feature dimensionality while maximizing classification performance. A set of single (LR, KNN, NB, MLP, and SVM) and ensemble (RF, XGBoost, and Voting) classifiers are evaluated via 10-fold cross-validation. The Sleep-IoMT stress dataset, comprising biosensor-based physiological signals, was used for experimental validation. Findings: The framework achieved high classification accuracy across all the models, with all the classifiers exceeding 0.98 accuracy. Compared with the PSO-WOA, the PSO-WOA with the Lévy flight method demonstrated superior performance in terms of both feature selection quality and training time efficiency. The results confirm that effective feature selection significantly improves model accuracy and interpretability. Novelty/Improvement: This research introduces a hybrid approach for feature selection (PSO-WOA) in the context of stress prediction from sleep-related IoMT data. The integration of Lévy flight into the PSO-WOA enhances exploration capabilities and reduces premature convergence, offering a robust solution for real-world healthcare applications, e.g., mobile stress monitoring and early intervention systems.