AI-based novel hybrid approach for stress severity detection using cascaded CNN and threshold-optimized random forests

Stress onset helps prevent chronic illnesses, strengthens individuals intellectually, and improves quality of life. Numerous studies have examined computerised stress prediction. However, current research does not provide any comprehensive or universal techniques for early intervention and tailored wellbeing. This research presents a unique stress detection approach using ML and DL. This approach integrates facial image processing with physiological data including blood pressure, heart rate, glucose, and oxygen using a novel dataset. We employ the Multitask Cascaded Convolutional Neural Network to detect facial landmarks and calculate their Euclidean distance to enhance feature representation. To complete the stress dataset, physiological data are included. We propose a new pipeline to address class imbalance and broaden models. Layered stratified cross-validation, synthetic minority oversampling, and extensive tuning of SVM, KNN, DT, RF, and XGBoost are included in this process. Using the highest F1-macro score, threshold tuning optimises performance across all classes. The best model results from MTCNN-based facial landmark distances with relevant physiological parameters. We achieve 94.81% test F1-macro and 96.13% accuracy. This work represents a promising framework of stress detection, which underlines the importance of a novel set of facial features, synthetic class balancing techniques, model selection, and a validation strategy for developing generalized and accurate predictive models in health monitoring tasks.

Перевод пока недоступен