Multimodal Stress Detection: Feature Selection and Segmentation Strategies for Improved Machine Learning

Accurate stress detection using machine learning (ML) relies heavily on optimal feature selection and signal segment length, especially when dealing with multimodal physiological signals. This study explores the impact of feature categorization and segment duration on ML model performance in a 4-class stress detection task. Physiological data, including Electrodermal Activity (EDA), Heart Rate (HR), Temperature, Oxygen Saturation (SpO2), and Accelerometer signals, were used to extract features, which were then categorized into time-domain linear, time-domain nonlinear, and frequency-domain groups. The effectiveness of these features was evaluated over 1-minute and 2-minute signal segments.Experimental results indicate that time-domain nonlinear features are highly discriminative, enabling even a simple Logistic Regression (LR) model to achieve 100% accuracy within the first minute of stress onset. Conversely, time-domain linear and frequency-domain features also contribute to high accuracy (up to 99%) but require nonlinear models such as Random Forest (RF) or Support Vector Machine (SVM) for optimal performance. Additionally, while 2-minute segments necessitate more complex models like SVM for accurate classification, 1-minute segments provide sufficient discriminatory power for stress detection with simpler models. These findings emphasize the importance of feature engineering, segment selection, and multimodal signal integration in optimizing ML-based stress detection systems.