Abstract
In affective computing, machine learning models are frequently employed to detect stress using wearable sensor data. However, a critical question persists: are these models identifying unique physiological signatures of stress, or are they merely detecting general emotional arousal? This study performs a cross-corpus evaluation using three datasets—APD (anxiety), WESAD (stress/amusement), and CASE (general emotion)—to investigate this ambiguity. Our results demonstrate that models trained on general arousal generalize effectively to stress detection tasks, suggesting that current "stress detectors" primarily capture broad physiological intensity rather than stress-specific markers.
Introduction
The ability to detect stress through wearable sensors like heart rate monitors and skin conductance sensors has significant implications for mental health monitoring. However, stress is psychologically categorized as a high-arousal state, a characteristic it shares with non-stressful emotions like excitement or amusement. If AI models cannot distinguish between a panic attack and an exciting movie, their utility in clinical settings is limited.
This research addresses the "generalizability gap" by testing whether models trained on general emotional arousal can detect stress in entirely different datasets, and vice versa. By mapping emotions onto the circumplex model of affect (Figure 1), we can better understand the overlap between activation (arousal) and valence (pleasantness).
Methodology
The researchers utilized a multi-stage pipeline to evaluate physiological generalizability:
- Signal Processing: Focused on Electrocardiogram (ECG) and Electrodermal Activity (EDA), extracting features such as heart rate variability (RMSSD, SDNN) and skin conductance levels (SCL).
- Machine Learning Models: Evaluated five classification approaches: Support Vector Machines (SVM), LightGBM (LGBM), Random Forest (RF), XGBoost (XGB), and an ensemble model.
- Validation Strategies: Employed within-corpus, cross-corpus (train on one, test on another), and leave-one-corpus-out (LOCO) testing to ensure robust results.
Datasets
Three distinct datasets were used to bridge the gap between stress and general emotion:
- APD (Anxiety Phases Dataset): Contains physiological responses to specific anxiety triggers like public speaking and phobias.
- WESAD (Wearable Stress and Affect Detection): A benchmark dataset featuring social stress tests and amusement.
- CASE (Continuously Annotated Signals of Emotion): Focuses on general emotional arousal and valence elicited through video clips.
Key Results
The study's most significant finding is the high degree of overlap between stress and general arousal. Models trained on general arousal (CASE) performed remarkably well when identifying stress in the WESAD and APD datasets.
As shown in Figure 5, models trained on the CASE (arousal) dataset and tested on the WESAD (stress) dataset achieved up to 77.4% accuracy (SVM). Conversely, models trained on stress (APD) and tested on general arousal (CASE) reached 70.8% accuracy (Ensemble). While within-dataset performance reached as high as 99% (on WESAD), the cross-corpus results provide a more realistic measure of how these models might perform in diverse, real-world scenarios.
Discussion
The analysis revealed that heart rate (BPM) was the most consistent and significant feature across all three datasets, regardless of the emotional label. This suggests that physiological "red flags" for stress are largely identical to those of general high arousal. The difficulty in finding stress-specific physiological features highlights a major hurdle for the field: without additional context, AI struggles to differentiate between various high-energy states. Interestingly, "arousal" appears to be a more stable and robust label for machine learning than "stress," as evidenced by the success of cross-label training.
Conclusion
This research underscores the challenge of specificity in affective computing. While current models are adept at detecting physiological activation, they often fail to distinguish the underlying cause of that activation. Future work should leverage Explainable AI (XAI) techniques to isolate features unique to stress versus general arousal. Additionally, moving beyond controlled lab settings to "in-the-wild" data will be crucial for developing diagnostic tools that can account for physical activity and other real-world confounders.
Summary updated