An Enhanced Source-Free Unsupervised Domain Adaptation Framework for Cross-Dataset EEG Emotion Recognition via Predictive Coding and Test-Time Training

Technical Summary: An Enhanced Source-Free Unsupervised Domain Adaptation Framework for Cross-Dataset EEG Emotion Recognition

Problem Statement
Electroencephalogram (EEG)-based emotion recognition faces significant challenges in generalizing across different datasets due to domain shifts caused by inter-subject variability, differences in recording devices, electrode configurations, and experimental protocols. While Unsupervised Domain Adaptation (UDA) has been explored to mitigate these shifts, conventional methods require access to source-domain data during adaptation, which is often impractical in real-world healthcare and brain-computer interface (BCI) applications due to privacy, storage, and ethical constraints. Source-Free Unsupervised Domain Adaptation (SF-UDA) addresses this by adapting a pre-trained source model to an unlabeled target domain without source data. However, existing SF-UDA methods struggle with large domain gaps, the accumulation of errors from noisy pseudo-labels, and unstable adaptation, particularly in the context of cross-dataset EEG emotion recognition.

Methodology
The authors propose an enhanced SF-UDA framework comprising four distinct stages: pre-training, computation, target adaptation, and inference.

1. Non-Contrastive Predictive Coding Pre-training:
   Unlike previous work relying on supervised source training, this framework introduces a self-supervised pre-training strategy. A predictive coding model, utilizing an encoder-decoder architecture, is trained to reconstruct the Power Spectral Density (PSD) of future EEG segments based on past activity. This non-contrastive approach learns robust, temporally predictive representations without requiring negative pairs or contrastive objectives. The pre-training objective combines classification loss, classifier discrepancy loss, and predictive coding loss to ensure discriminative and temporally consistent feature learning.

2. Target Adaptation (Dual-Stage Strategy):
   Once the source model is pre-trained, the framework adapts to the unlabeled target domain using a two-step process:

   • Multi-Loss Adaptive Regularization (MLAR): This stage estimates target-domain structure via class-wise clustering. It generates pseudo-labels by assigning samples to the nearest cluster centroids. To stabilize training and reduce noise, MLAR employs a reliable sample mask based on prediction confidence and agreement between cluster-based pseudo-labels and model predictions. The optimization minimizes pseudo-label agreement loss, classifier discrepancy loss (Jensen-Shannon divergence between two parallel classifiers), prediction entropy, and a target distribution balancing term.
   • Localized Consistency Learning (LCL): This stage enforces prediction consistency among reliable neighbors. Neighbors are selected based on the intersection of feature-space similarity and classifier-output similarity, ensuring that the model aligns predictions within local neighborhoods while filtering out unreliable neighbors.

3. Lightweight Test-Time Training (TTT):
   During inference, the framework implements an uncertainty-based online update mechanism. Samples are evaluated using prediction entropy and predictive reconstruction error. Only uncertain samples (those exceeding quantile-based thresholds for both metrics) are selected for adaptation. For these samples, a lightweight update is performed on normalization layers and attention gate modules using a combination of entropy minimization and predictive reconstruction loss. This selective update enhances robustness against distribution shifts while maintaining computational efficiency by avoiding updates for confident predictions.

Key Contributions

• Enhanced Framework: The authors propose a comprehensive SF-UDA framework specifically designed for cross-dataset EEG emotion recognition, addressing the absence of source data and the challenges of domain shifts.
• Self-Supervised Pre-training: Introduction of a non-contrastive predictive coding strategy that learns robust temporal EEG representations without relying on labeled source data or contrastive objectives.
• Robust Adaptation Mechanisms: The design of a dual-stage adaptation strategy (MLAR and LCL) that improves stability and neighborhood consistency under noisy pseudo-labels.
• Efficient Test-Time Refinement: A lightweight TTT mechanism that selectively updates the model for uncertain samples, balancing robustness with computational efficiency.

Experimental Results
The framework was evaluated on three public datasets: DEAP, SEED, and DREAMER, using cross-dataset settings (e.g., training on DEAP and testing on SEED).

• Performance: The proposed method consistently outperformed state-of-the-art SF-UDA baselines (including methods by Ahmed et al., Du et al., Ragab et al., and Zhao et al.) and the authors' previous SF-UDA work.
  • On the SEED dataset (trained on DEAP), the method achieved 69.56% accuracy for binary classification and 57.08% for multi-class classification.
  • On the DREAMER dataset (trained on DEAP), it achieved 63.03% accuracy for binary classification.
  • When trained on SEED and tested on DEAP, it achieved 61.38% (binary) and 53.87% (multi-class).
• Statistical Significance: Paired-sample t-tests confirmed that improvements over competing methods were statistically significant.
• Ablation Studies: Experiments demonstrated that removing the predictive coding pre-training, the MLAR module, or the test-time training resulted in significant performance drops, validating the necessity of each component.
• Interpretability: SHAP analysis revealed that the model relies heavily on gamma and beta band features from occipital, prefrontal, and frontal-central regions, aligning with known neurophysiological markers of emotion.

Significance and Claims
The paper claims that the proposed framework effectively combines predictive self-supervised learning, source-free domain adaptation, and test-time refinement to achieve robust EEG emotion recognition under severe domain shifts. By eliminating the need for source data and addressing the instability of pseudo-labels through a multi-stage, consistency-driven approach, the method offers a practical solution for privacy-sensitive affective computing applications. The authors position this work as a significant step forward in bridging the gap between controlled experimental datasets and real-world deployment, where data sharing is restricted and domain variability is high. Future directions identified by the authors include exploring continual adaptation for evolving target distributions and extending the framework to other physiological signals and multimodal fusion.