## STPC: A Physics-Informed Framework for Trustworthy EEG AI ### Abstract The clinical viability of deep learning in neurology is critically dependent on the trustworthiness of its outputs. This project introduces and validates a **Physics-Informed Regularizer**, Spatio-Temporal Physiological Consistency (STPC), designed to ensure deep learning models for EEG analysis produce physically plausible and diagnostically reliable results. We demonstrate that STPC is superior to baseline methods for denoising complex seizure data and, more importantly, can be used in a self-supervised context to learn meaningful neural representations from unlabeled data, successfully differentiating seizure from non-seizure brain states with a high Silhouette Score of **0.6350**. This project provides the full codebase, experimental results, and final paper for this research. --- ### Key Results #### Phase 1: Spatio-Temporal Denoising of an Epileptic Seizure *Our STPC model (right) faithfully reconstructs the seizure's complex topography, while a baseline L1 model (center-right) collapses into non-physiological artifacts.*  --- #### Phase 2: Frequency-Specific Signal Preservation *The Frequency-STPC model (bottom) surgically removes powerful low-frequency noise (top) while preserving the critical Alpha-band signature of the ground truth (middle).*  --- #### Phase 3: Unsupervised Discovery of Neural States *Without any labels, our self-supervised model learned to separate Non-Seizure (purple) and Seizure (yellow) brain states into distinct clusters.*  --- ### Citation If you use this work, please cite the archived software record: > Mohanarangan Desigan. (2025). STPC: A Physics-Informed Framework for Trustworthy EEG AI (v1.0.0). Zenodo. https://doi.org/10.5281/zenodo.17218523