Authors - Kashish Goyal, Parteek Kumar, Karun Verma Abstract - The clinical deployment of continuous epileptic seizure forecasting systems is severely hindered by the cold-start problem. Current state-of-the-art deep learning models require patient-specific fine-tuning, necessitating the recording of multiple seizures from a newly admitted patient before the system becomes operational. To achieve immediate clinical utility, forecasting models must operate in a zero-shot capacity. This paper presents a Zero-Shot Cross-Patient Transfer Framework, leveraging the Horizon-Aware Graph Transformer as a universal feature extractor, coupled with the Strict Discipline Protocol as a rigid domain adaptation layer. By anchoring the batch normalization layers to a global source distribution and utilizing a brief interictal calibration phase, the framework mitigates the severe covariate shift inherent in cross-patient electroencephalogram signals. Experimental validation on the CHB-MIT dataset demonstrates a sensitivity of 87.3% with a false alarm rate of 0.28 per hour, achieving a Time-to-Utility of exactly 10 minutes, a 99.9% reduction compared to conventional patient-specific approaches requiring 5-14 days of monitoring. The framework successfully bypasses patientspecific training, offering immediate clinical interoperability while minimizing alarm fatigue through disciplined feature scaling.
Open Zoom