Physiological state recognition from wearable measurements is inherently nonlinear, with tree-based ensemble models consistently outperforming linear baselines across stress, cognitive load, and exercise detection tasks.
Key Findings
Results
Nonlinear models consistently outperformed linear models across all datasets for physiological state recognition.
Tree-based ensembles achieved 0.89–0.98 accuracy and 0.96–0.99 AUC across all datasets.
Linear models remained below 0.70–0.73 AUC across the same datasets.
Four linear models were tested: logistic regression, linear SVM, linear discriminant analysis, and ridge classifier.
Nonlinear approaches included SVM(RBF), random forest, gradient boosting, XGBoost, and LightGBM.
This pattern held across structured stress induction, real-world exam stress, aerobic and anaerobic exercise, and cognitive load tasks.
LOSO validation was used as the evaluation scheme to assess subject-independent generalization.
Substantial inter-individual variability was observed across all three datasets.
Despite this variability, nonlinear models maintained moderate generalization performance across subjects.
Three publicly available Empatica E4 datasets were used for evaluation.
Results
Multimodal fusion was confirmed as important for classification performance, with electrodermal activity, temperature, and accelerometry being the most critical modalities.
Ablation studies were conducted to assess the contribution of individual signal modalities.
Electrodermal activity (EDA), temperature, and accelerometry were identified as particularly important modalities.
Removal of these modalities degraded classification performance in the ablation analyses.
The framework used standardized preprocessing and window-based feature extraction across all modalities.
Results
SHAP analysis revealed nonlinear and interaction-driven feature effects consistent with known autonomic mechanisms.
Shapley Additive Explanations (SHAP) were used for interpretability analysis across all models and datasets.
Feature effects were found to be nonlinear rather than monotonic.
Interaction effects between features were observed and considered consistent with known autonomic nervous system dynamics.
SHAP results provided mechanistic support for the superiority of nonlinear modeling approaches.
Methods
The study applied a unified signal-processing and evaluation framework across three publicly available Empatica E4 wearable datasets covering multiple physiological states.
The three datasets covered structured stress induction, real-world exam stress, aerobic and anaerobic exercise, and cognitive load tasks.
All datasets were collected using the Empatica E4 wearable device.
The framework included standardized preprocessing, window-based feature extraction, subject-independent evaluation, LOSO validation, multimodal ablation studies, and SHAP-based interpretability analysis.
The study aimed to establish a unified benchmark for physiological state recognition from wearable measurements.
Conclusions
Physiological state recognition from wearable measurements is inherently nonlinear even when individual modalities exhibit monotonic trends.
Individual physiological signals such as heart rate may show simple monotonic trends (e.g., elevated heart rate under stress), yet overall classification is nonlinear.
The nonlinearity was demonstrated across all three task domains: stress, cognitive load, and exercise.
The findings support the necessity of nonlinear modeling for robust, real-time wearable health-monitoring systems.
Linear assumptions about autonomic nervous system dynamics were found to be insufficient for accurate physiological state recognition.