A rank-one ECG model gave mixed later prediction gains in nine adult 1–40 Hz rest SEEG records after early penalty selection

This study compared two ECG voltage models in nine adult rest SEEG records from OpenNeuro ds003688 version 1.0.7. The records had 578 good SEEG contacts. Both models used 33 ECG samples across 250 ms and a 1–40 Hz band. Early data selected a ridge penalty for each model. A separate later block supplied the score. The rank-one model used one shared spatial pattern. It lowered error in three adults, raised error in four, and tied in two. The equal-adult mean gain was 0.092 percentage points. The 95% adult bootstrap interval was -0.014 to +0.244 points. The two-sided exact sign flip p value was 0.3281. The study did not establish a group advantage. At a fixed penalty of 0.01, rank one lowered error in all nine adults. Its mean gain was 0.414 points, but the full model had negative mean skill against the early mean. This fixed check was descriptive. Common average and bipolar checks also gave mixed gains. The package preserves all nine compact inputs, pinned source retrieval, code, model coefficients, tuning scores, timing controls, plots, and independent numerical checks. An offline repeat matched all scientific outputs. The same clinical cohort was used in earlier work. This is an exploratory secondary study. ECG prediction cannot separate cardiac artifact, neural heart responses, and shared noise.

SEEG · ECG · reduced-rank regression · ridge regression · model selection · prediction · OpenNeuro