Interpretable Multi-Task PINN for Emotion Recognition and EDA Prediction

TL;DR

This paper introduces a multi-task physics-informed neural network that simultaneously predicts electrodermal activity and classifies emotions, integrating physiological models with machine learning for improved interpretability and performance.

Contribution

It presents the first multi-task PINN framework for wearable emotion recognition, combining biophysical constraints with deep learning for enhanced accuracy and interpretability.

Findings

Achieves an EDA RMSE of 0.0362 and F1-score of 94.08%

Outperforms classical models like SVR and XGBoost

Learns stable, interpretable physical parameters

Abstract

Understanding and predicting human emotional and physiological states using wearable sensors has important applications in stress monitoring, mental health assessment, and affective computing. This study presents a novel Multi-Task Physics-Informed Neural Network (PINN) that performs Electrodermal Activity (EDA) prediction and emotion classification simultaneously, using the publicly available WESAD dataset. The model integrates psychological self-report features (PANAS and SAM) with a physics-inspired differential equation representing EDA dynamics, enforcing biophysically grounded constraints through a custom loss function. This loss combines EDA regression, emotion classification, and a physics residual term for improved interpretability. The architecture supports dual outputs for both tasks and is trained under a unified multi-task framework. Evaluated using 5-fold cross-validation, the model achieves an average EDA RMSE of 0.0362, Pearson correlation of 0.9919, and F1-score of 94.08 percent. These results outperform classical models such as SVR and XGBoost, as well as ablated variants like emotion-only and EDA-only models. In addition, the learned physical parameters including decay rate (alpha_0), emotional sensitivity (beta), and time scaling (gamma) are interpretable and stable across folds, aligning with known principles of human physiology. This work is the first to introduce a multi-task PINN framework for wearable emotion recognition, offering improved performance, generalizability, and model transparency. The proposed system provides a foundation for future interpretable and multimodal applications in healthcare and human-computer interaction.