Partial Domain Adaptation for Stable Neural Decoding in Disentangled Latent Subspaces

Objective: Brain-Computer Interfaces (BCI) have demonstrated significant potential in neural rehabilitation. However, the variability of non-stationary neural signals often leads to instabilities of behavioral decoding, posing critical obstacles to chronic applications. Domain adaptation technique offers a promising solution by obtaining the invariant neural representation against non-stationary signals through distribution alignment. Here, we demonstrate domain adaptation that directly applied to neural data may lead to unstable performance, mostly due to the common presence of task-irrelevant components within neural signals. To address this, we aim to identify task-relevant components to achieve more stable neural alignment. Methods: In this work, we propose a novel partial domain adaptation (PDA) framework that performs neural alignment within the task-relevant latent subspace. With pre-aligned short-time windows as input, the proposed latent space is constructed based on a causal dynamical system, enabling more flexible neural decoding. Within this latent space, task-relevant dynamical features are disentangled from task-irrelevant components through VAE-based representation learning and adversarial alignment. The aligned task-relevant features are then employed for neural decoding across domains. Results: Using Lyapunov theory, we analytically validated the improved stability of late our neural representations through alignment. Experiments with various neural datasets verified that PDA significantly enhanced the cross-session decoding performance. Conclusion: PDA successfully achieved stable neural representations across different experimental days, enabling reliable long-term decoding. Significance: Our approach provides a novel aspect for addressing the challenge of chronic reliability in real-world BCI deployments.