Date created: | Last Updated:

Category: Project

Description: The clinical viability of deep learning in neurology is critically dependent on the trustworthiness of its outputs. While deep learning models for EEG denoising can effectively reduce noise, they often fail to preserve the underlying neurophysiological structure, leading to outputs that are clean but diagnostically misleading. This paper argues that this failure is twofold: a failure of naive loss functions that ignore the physics of the signal, and a failure of simplistic evaluation metrics like RMSE that are blind to structural integrity. We address this by introducing and validating a \textbf{Physics-Informed Regularizer} we call Spatio-Temporal Physiological Consistency (STPC). STPC augments a standard L1 loss with a Temporal Gradient term to preserve event sharpness and a Spatial Laplacian term to enforce electrodynamic plausibility. We trained a U-Net with both a baseline L1 loss and our STPC loss, evaluating them on a completely held-out epileptic seizure segment. The results reveal a fascinating divergence in metrics: while both models achieved a near-identical, low RMSE ($\approx$1.8e-5) and a similar Mean SSIM ($\approx$0.74), the STPC model demonstrated superior preservation of the signal's frequency content, evidenced by a higher Mean Spectral Coherence. Most critically, qualitative analysis revealed that only the STPC model produced a reconstruction that was visually and structurally faithful to the ground truth, while the baseline model collapsed into non-physiological artifacts. Our findings demonstrate that for AI to be trustworthy in clinical neuroscience, it must be guided by physics-informed principles and evaluated with a holistic combination of visual analysis and targeted, structure-aware metrics.