A Hybrid Deep Learning Approach to Stress Detection: Integrating CNN-LSTM with Reinforcement Learning and Active Learning

Abstract

Stress has become a critical mental health issue in modern society and detection methods need to be appropriate and timely. In recent years, advances in sensor technology, particularly wearable devices, have significantly improved the collection of physiological biomarkers to monitor stress. Furthermore, numerous studies have been conducted on stress recognition using machine learning techniques. In this work, we introduce an improved approach to enhance stress recognition. A combined model of Convolutional Neural Networks and Long Short-Term Memory (CNN-LSTM) was first implemented, achieving an enhanced classification accuracy of 94.10%. To further improve performance, we present a new framework that integrates Double Deep Q-Networks (DDQN) with Active Learning (AL) to optimise the selection of informative data samples. This DDQN-AL-driven strategy significantly improved classification accuracy, demonstrating the model’s capability to generalise across individual variations effectively. These results were evaluated in Leave-One-Subject-Out (LOSO) cross-validation, revealing that the performance of the enhanced CNN-LSTM architecture, supported by DDQN and AL, offers greater accuracy (95.10%) in stress recognition.