Delayed ECG event bins reduced mean test skill after fixed linear ECG prediction in nine adult SEEG rest records

We tested whether delayed ECG event bins improve SEEG voltage prediction beyond a fixed linear ECG model. We used nine adult epilepsy rest records from OpenNeuro ds003688, version 1.0.7, with 578 good SEEG contacts. The early block trained the models. A separate late block supplied the test error. Six event bins covered nominal delays from 0.2 to 0.5 seconds. Their training linear ECG projection was removed before the added fit. Mean supplied-reference gain was -0.003849, or -0.385 percentage points of skill. All nine adult means were negative. The 95 percent adult bootstrap interval was -0.543 to -0.243 points. The exact two-sided sign-flip test gave p = 0.00390625. It assumes independent adults and sign symmetry under zero. All planned group checks also had negative gains. Recorded timing performed better than typical shifted timing, but still reduced skill. The timing controls are descriptive because event times are not exchangeable when ECG is fixed. This result applies to the fixed model and selected short records. It does not show absence of neural heartbeat responses. The package preserves code, compact inputs, source versions and hashes, all scores, controls, and reproduction instructions. Independent numerical checks and an exact offline repeat passed. Fresh source checks matched all nine analysis byte ranges, complete headers, and reduced inputs.

SEEG · ECG · heartbeat timing · voltage prediction · event bins · ridge regression · prediction skill · OpenNeuro