Adding power variation lowered mean seizure onset ranking AUC in a fixed linear model of two short SEEG records from each of 36 patients

We tested whether variation in low frequency power adds seizure onset contact ranking value above median power and contact position. We used two 120-second public SEEG records from each of 36 patients in OpenNeuro ds004100 version 1.1.3. The matched set had 4,252 contacts and 360 clinical onset labels. Each fixed linear model trained without the test patient. Adding two log-power standard deviation features lowered mean patient AUC from 0.78658 to 0.78090. The mean paired change was -0.00568. Fourteen patients had a higher AUC and 22 had a lower AUC. The conditional 95 percent patient interval was -0.01178 to -0.00026. It holds fitted predictions fixed and omits shared training uncertainty. All 36 additional patient omission checks had negative mean changes after retraining. The common average signal check had a small positive mean change, with an interval that included zero. The result applies to this model, these short records, and the defined power features at one hospital. It does not test a biological cause or treatment benefit. The package preserves the unchanged source measurements, code, all 3,204 fits, predictions, patient results, and retrieval records. Independent raw checks covered four sample rates. A complete offline repeat matched all 12 checked output files byte for byte.

stereoelectroencephalography · seizure onset zone · low frequency power · feature variation · patient-level validation · AUC · epilepsy