Domain adaptation for physiological signal classification

Physiological signals, such as Electroencephalograms (EEG) and Electrocardiograms (ECG), provide valuable insights into brain activity, cardiovascular health, and emotional states. However, the scarcity of labeled data and variations in signal morphology across individuals limit cross-domain model applicability, especially in cross-dataset scenarios with differences in demographics, sensor technologies, and recording conditions. Unsupervised Domain Adaptation (UDA) addresses this by adapting models from a labeled source domain to an unlabeled target domain, but performance often suffers due to large domain shifts and noisy pseudo-labels. Additionally, UDA typically requires access to source data, which may not be feasible due to privacy or computational constraints. This thesis proposes novel models for both UDA and Source-Free UDA (SF-UDA) tailored to physiological signal adaptation. Our UDA method is a multi-stage framework that begins with source-domain pre-training and forms clusters using source labels and confident target predictions. We introduce a robust target sample selection strategy that filters predictions based on model confidence, reducing misclassification. The model then aligns feature distributions by jointly optimizing four loss terms. To prevent negative transfer, we propose Gradual Proximity-guided Target Data Selection (GPTDS), which incrementally incorporates target samples based on proximity to source clusters and prediction certainty. For improved inference, we introduce Prediction Confidence-aware Test-Time Augmentation (PC-TTA), which activates augmentations selectively based on prediction confidence—enhancing robustness without the overhead of traditional full-scope TTA. In the source-free setting, we present a multi-stage SF-UDA method that does not require access to source data. It includes Dual-Loss Adaptive Regularization (DLAR), which minimizes prediction inconsistency and aligns predictions with reliable pseudo-labels. Localized Consistency Learning (LCL) further reinforces neighborhood-level label stability, reducing the impact of noisy pseudo-labels. Experiments on cross-dataset EEG-based emotion recognition and ECG-based arrhythmia classification show that our UDA model outperforms state-of-the-art methods. On three arrhythmia datasets, it achieves accuracy gains of 9.90%, 6.96%, and 9.42%, averaging 14.29% over baseline. For emotion recognition, UDA yields 6.01% and 2.47% gains, while SF-UDA improves accuracy by 5.44% and 3.86%. PC-TTA reduces inference time by a factor of 15.