Convolutional Neural Network-Based Human Stress Detection Using Multivariate Sensor Data and Cross-Validation

Abstract

Early identification of human stress levels plays a crucial role in promoting mental well-being and preventing related health issues. However, conventional stress assessment methods often involve multi-step procedures or subjective evaluations, making them inefficient and impractical for continuous monitoring. This study introduces a Convolutional Neural Network (CNN)-based approach to automatically detect human stress using multivariate sensor data, such as heart rate, oxygen saturation, body temperature, and movement signals. Unlike traditional machine learning methods that rely on handcrafted features and shallow classifiers, the proposed deep learning model leverages raw sensor data to learn hierarchical representations of physiological patterns associated with various stress levels. The dataset utilized in this research is the SaYoPillow dataset obtained from Kaggle, which includes labeled physiological signals based on subjective stress assessments. Input features are normalized and reshaped into one-dimensional sequences compatible with the CNN architecture. A stratified 5-fold cross-validation strategy is used to ensure robust and generalizable model performance. The proposed CNN model achieved an outstanding accuracy of 0.999, with a precision of 0.998, recall of 0.991, and F1 score of 0.994, outperforming baseline models such as Decision Tree with accuracy of 0.987 and Random Forest with accuracy of 0.981. These results highlight the CNN model’s strong potential for real-time, reliable stress monitoring using wearable sensors, making it a promising solution for digital health and well-being applications.