AN INTELLIGENT SLEEP-BASED MACHINE LEARNING SYSTEM FOR CONTINUOUS AND NON-INTRUSIVE HUMAN STRESS LEVEL DETECTION

Abstract

Machine learning-based human stress detection from sleep patterns leverages wearable- generated physiological and behavioural sleep indicators, including sleep duration, heart rate variability, nocturnal body movement, and sleep stage distribution, to enable objective, continuous evaluation of stress state. Nevertheless, the actual wearable sleep data are usually noisy, incomplete, and highly heterogeneous across individuals, limiting the reliability of features and hindering generalisation. To overcome these problems, this paper proposes a combined sleep-based stress detection model that integrates multiple smart learning modules. The first step is a Robust Wearable Sleep Data Preprocessing Network (RWSD- PN) that performs adaptive noise reduction, missing-value processing, and signal standardisation to improve data quality. This is followed by the Sleep Attribute Impact Quantification Model (SAIQ-M), which is used to assess the relative importance of sleep variables to prioritize informative physiological indicators. It is then processed into a Deep Physiological Sleep Representation Network (DPSR-Net) that produces high-level multivariate representations of intricate sleep-stress interaction patterns, both nonlinear and temporal. In order to enhance further efficiency, the most discriminative and non-redundant subset of features is selected based on the Hybrid Swarm Learning (MSFO-HSL), which results in the diminishing of dimensionality and the threat of overfitting. Lastly, a Transformer-Guided Sleep Stress Classification Network (TG-SSCN) is a model based on attention-guided learning that estimates feature dependencies and correctly classifies stress levels. Through experimental analysis, the framework can be shown to have high predictive performance, including Accuracy, Precision, Recall, F1-Score, and AUC, thereby enhancing robustness, scalability, and reliability. The framework offers an unobtrusive, data-driven, continuous human stress monitor in real-world conditions.