Transformer-Based Decomposition of Electrodermal Activity for Real-World Mental Health Applications

Area of Science:

• Biomedical Engineering
• Physiological Signal Processing
• Machine Learning for Health

Background:

• Electrodermal Activity (EDA) decomposition into phasic and tonic components is crucial for identifying emotional and physiological biomarkers.
• Existing methods face challenges with noisy, in-the-wild data from wearable devices.
• Accurate decomposition is vital for reliable biosignal analysis.

Purpose of the Study:

• To compare knowledge-driven, statistical, and deep learning methods for EDA signal decomposition.
• To introduce and evaluate the Feel Transformer, a novel deep learning model for unsupervised EDA decomposition.
• To assess the model's performance on real-world, noisy data from wearable sensors.

Main Methods:

• Development of the Feel Transformer, a Transformer-based model adapted from Autoformer.
• Implementation of pooling and trend-removal mechanisms for physiologically meaningful decomposition.
• Comparative analysis against established methods like Ledalab, cvxEDA, and conventional detrending.

Main Results:

• The Feel Transformer demonstrated a strong balance between feature fidelity (e.g., SCR frequency, amplitude, tonic slope) and robustness to noise.
• The model effectively separated phasic and tonic components without requiring explicit supervision.
• Performance was validated on in-the-wild data, outperforming or matching traditional methods in key metrics.

Conclusions:

• The Feel Transformer offers a robust and effective approach for decomposing electrodermal activity signals.
• Its ability to handle noisy, real-world data makes it suitable for wearable biosensor applications.
• Potential applications include real-time stress prediction, digital mental health, and physiological forecasting.