An Explainable Deep Learning Approach for Stress Detection in Wearable Sensor Measurements.

Area of Science:

• Computational neuroscience
• Affective computing
• Wearable technology

Background:

• Stress significantly impacts human health.
• Wearable sensors and deep learning offer new avenues for detecting emotional states.
• Accurate, real-time stress detection can mitigate health risks in critical situations.

Purpose of the Study:

• To propose an explainable deep learning methodology for automatic stress detection.
• To utilize physiological sensor data from the Empatica E4 wristband.
• To address challenges of low data regimes in sensor measurements.

Main Methods:

• Developed a Long-Short Term-Memory (LSTM) network enhanced by a Deep Generative Ensemble of conditional GANs (LSTM DGE).
• Employed Integrated Gradients (IG) for model explainability, identifying key predictive features.
• Compared performance against state-of-the-art expert-based stress detection methods.

Main Results:

• The LSTM DGE model achieved superior performance, outperforming the state-of-the-art by 3% in recall and 7.18% in precision.
• Integrated Gradients analysis revealed significant overlap between model-identified stress features (electrodermal activity) and established literature.
• Demonstrated the model's ability to provide interpretable insights into stress detection.

Conclusions:

• The proposed LSTM DGE model offers an effective and explainable approach for stress detection using wearable sensor data.
• Explainability through Integrated Gradients validates the model's reliance on physiologically relevant features.
• This methodology holds promise for advancing stress monitoring in real-world applications.