ActDiffNet: Multisensor Affective State Recognition by Actively Synthesizing Minority Patterns via Conditional Diffusion

A key challenge in personalized ubiquitous healthcare is developing efficient wearable platforms that accurately classify biosignals while remaining adaptive to the evolving data patterns, particularly highlighting an individual's personal and other exterior context dynamics. However, several challenges plague machine learning applications involving biomedical signals, including limited data, imbalanced classes, difficulty accessing reliable annotated data, and noisy measurements. To this end, we propose an active learning model ActDiffNet for affective state recognition from multisensor signals that first leverage only a small annotated data collection to build an initial classifier, and later iteratively upgrade via a shortlisted set of synthesized 'hard' signals conditionally diffused by those unique signal patterns, on which the model has not been sufficiently trained yet. The proposed ActDiffNet converges faster, achieving comparable classification performance with 1-2 orders of magnitude fewer labeled samples than fully supervised approaches to attain a state-of-the-art accuracy of 78%. An effective Context Conditioned Synthetic Signal Generation module that employs multiple sensor-specific copies of the conditioned U-Net to facilitate synthesizing signals that closely mimic the sensor and class-specific patterns of shortlisted 'hard' signals within its generated outputs. Extensive evaluation using two public datasets WESAD and CASE reports outperformance (e.g., 1.5 - 3% improved accuracy) of the proposed ActDiffNet against state-of-the-art supervised or self-supervised models while delivering a consistently robust generalization all across.