Hybrid ensemble framework for real time stress detection using multimodal wearable smart fabrics

A1 IS 83 A1 N n v 103 w 1 11 nw nww•. @SRM INSTITUTE OF SCIENCE & TECHNOLOGY (Deemed to be University u/s 3 of UGC Act, 1956).

MULTIMEDIA UNIVERSITY @SRM INsrrrurE OF SCIENCE & TECHNOLOGY (Deemed to be University u/s 3 of UGC Act, 1956).

Hybrid Ensemble Framework for Real-Time Stress Detection using Multi-modal Wearable Smart Fabrics.

RESEARCH OBJECTIVES. Develop an interpretable hybrid stress detection model by integrating Random Forest and Logistic Regression using a soft voting mechanism to achieve a balance between accuracy, stability, and transparency for wearable-based applications. Design a customized preprocessing and feature extraction pipeline for smart textile ECG signals, incorporating Notch and Butterworth filters to remove noise and extracting HRV biomarkers (e.g., RMSSD) for reliable physiological analysis. Validate and implement a real-time, modular stress monitoring framework using the WESAD dataset, ensuring consistent performance across multiple emotional states (baseline, stress, amusement) with explainable feature importance for clinical usability..

RESEARCH METHODOLOGY.

PROPOSED SYSTEM MODULES. [image] Tier 1: Data Acquisition WLSAD DATASET Resampling Tier 4: Soft Voting Ensemble Confusion ECG/EMG EDA/'1i.np Fortst Butterworth Rzture hasion Vector Tier 2: Signal Refinement Rolling Mean S nu»thing Tier 3: Feature Engineering IIRV Metrics RMSSD & BPM Score Standardisat ion It-Peak.

System Overview. Tier 1: Data Acquisition Multimodal physiological signals (ECG, EDA, EMG, respiration, temperature, and motion) from chest-mounted sensors are acquired at 700 Hz, reformatted, and down sampled to 70 Hz to significantly reduce computational and memory overhead. Tier 2: Signal Refinement A hybrid ensemble model is developed by combining Random Forest and Logistic Regression using a soft voting classifier to leverage both non-linear pattern recognition and linear interpretability. Tier 3: Feature Engineering HRV biomarkers are derived from ECG by detecting cardiac peaks (≥0.6 s interval) to compute RR intervals, Mean Heart Rate, and RMSSD while minimizing artifacts. Tier 4: Proposed Soft Voting Methodology The system combines a Random Forest (for complex, non-linear pattern recognition) with a Logistic Regression (for linear stability and calibrated probability scores..

Signal Preprocessing and Artifact Removal. Temporal Smoothing A 5-sample rolling mean filter (~70 ms) reduces motion-induced noise while preserving key physiological patterns. Z-score normalization standardizes the signal, ensuring consistency across sensors. This preprocessing improves R-peak detection and enhances HRV feature reliability, including RMSSD..

Signal Preprocessing and Artifact Removal Continues….

Feature Extraction and HRV Analysis. [image]. Figure 4: Distribution of physiological features across stress and non-stress states..

Hybrid Ensemble Classification. Random Forest Model Random Forest with 100 trees is employed to handle noisy and unpredictable smart fabric data using bagging and random feature selection. This improves robustness against sensor noise and captures complex relationships among HRV, heart rate, and EDA. Feature importance further ensures that predictions are driven by meaningful physiological signals. Logistic Regression Logistic Regression with L2 regularization is integrated to stabilize predictions and prevent overfitting to noise. Using a one-vs-rest approach, it provides calibrated probability outputs and serves as a computationally efficient, reliable counterbalance to Random Forest. Soft-Voting Mechanism Soft voting combines predicted probabilities from both models, producing smoother and more reliable decisions than hard voting. This approach reduces false alarms by balancing confident nonlinear predictions with conservative linear estimates..

Experimental Results. Classification Performance.

Experimental Results Continues….. [image] SS. Figure 6: Confusion matrices for individual subjects..

Experimental Results Continues….. Probabilistic Decision Analysis While AUC-ROC values indicate strong classification performance, F1-scores vary across subjects due to class imbalance and fixed thresholding. Balanced subjects (S3, S5) achieve higher F1 (>0.79), whereas imbalanced cases show overlapping probability distributions near the 0.5 threshold. Despite this, clear separation in predicted probabilities confirms effective learning of stress-related physiological features. The results suggest that performance can be significantly improved through subject-specific threshold calibration or balanced training data..

Experimental Results Continues….. What the Model Prioritized The Random Forest feature importance analysis, presented in Figure 8, reveals that heart rate (BPM) is the most influential predictor, contributing 36.4% of the total importance, followed by EDA standard deviation (21.4%) and acceleration magnitude (18.7%)..

[image] Sensor Correlation Heatmap (Raw Data) 012.

Conclusion and Future Directions. The proposed hybrid ensemble combining Random Forest and Logistic Regression achieves strong stress detection performance (AUC-ROC ≈ 0.94) using interpretable models. A robust preprocessing pipeline with Butterworth and Notch filters enables reliable extraction of HRV features, particularly BPM and RMSSD. However, variability in F1-scores across subjects highlights the need for personalized calibration. F uture work will focus on improving generalization through transfer learning and real-world validation. Significant inter-subject variability (F1: 0.0–0.80) limits generalization, indicating that physiological stress responses differ widely across individuals. To address this, future work will focus on subject adaptation using transfer learning, temporal modeling via sliding window analysis, comprehensive validation with leave-one-subject-out across all WESAD subjects, and handling class imbalance through synthetic oversampling..

THANK YOU.