Enhancing Stress Classification on the WESAD Dataset Through Regularized Ensemble and Optimized Deep Learning Techniques

Dr.B.Purushotham1 , Redyam Haritha2 , M.Uday Kiran3 , B.Ramya4 , V.Prasanth5
1 Associate Professor, HOD,Dept of Information Technology,SV College of Engineering, Tirupati, India.
2 B.Tech, Dept of Information Technology,SV College of Engineering, Tirupati, India.
3 B.Tech, Dept of Information Technology,SV College of Engineering, Tirupati, India.
4 B.Tech, Dept of Information Technology,SV College of Engineering, Tirupati, India.
5 B.Tech, Dept of Information Technology,SV College of Engineering, Tirupati, India

Abstract-Existing systems for stress detection utilize machine learning (Logistic Regression, Gaussian Naive Bayes, AdaBoost, XGBoost, Decision Trees, Extra Trees, Random Forest) and deep learning (DNN, CNN, RNN) models on the WESAD dataset's multimodal physiological signals from wearable sensors, including ACC, ECG, BVP, TEMP, RESP, EMG, and EDA. These systems classify four states—aseline,stress, amusement, and meditation—via two-phase evaluation:  Phase 1 (cross-subject training/testing) where RNN achieves F1-scores of 80.8% (chest) and 93.6% (wrist),and Phase 2 (intra-subject 80-20 split) where XGBoost  reaches 99.8% F1-scores on both chest and wrist data.Despite high accuracies, limitations include poor cross subject generalization (e.g., XGBoost's overfitting to intra-subject patterns, dropping from 99.8% to lowercross-subject performance), extended training/testing times for deep models (RNN up to 614 seconds), and high computational demands unsuitable for real-time wearable applications. Chest data excels intra-subject, while wrist data performs better cross-subject, but overall efficiencyremains constrained by resource-intensive processing. The proposed system addresses these by applyingregularization, ensemble methods, and hyperparametertuning to enhance machine learning generalizability,alongside deep learningoptimizations like early stopping,learning rate schedulers, and distributed training to reduce computation time. Benefits include improved crosssubject F1-scores for diverse wearables, faster inference for real-time monitoring and robust deployment on low compute devices, enabling scalable stress intervention with balanced accuracy and efficiency.Keywords: Logistic Regression, Gaussian Naive Bayes,AdaBoost, XGBoost, Decision Trees, Extra Trees,Random Forest, multimodal physiological signals,hyperparameter, deep learning.