Procrustes alignment did not improve few-trial SEEG movement decoding in 15 patients

We tested whether orthogonal Procrustes alignment of condition-mean latent trajectories from other patients could improve few-trial calibration for offline SEEG movement-versus-rest decoding. We used archived 70–150 Hz high-gamma activity window and bin features from 15 patients (3,970 trials) in OpenNeuro ds006253. In a pre-registered time-ordered 60/40 split, we compared a Ledoit–Wolf shrinkage LDA trained on 10 target-patient trials across all channels with an LDA trained on a five-dimensional PCA space augmented by the aligned data of 14 other patients. Mean test AUC was 0.724 for the all-channel within-patient decoder and 0.632 for the aligned pool (difference −0.091, 95% CI −0.125 to −0.058; 0/15 patients improved; primary hypothesis not supported). The aligned pool was also not better than the within-patient decoder in the same latent space (0.643; difference −0.011). Alignment did matter: the unrotated pool scored 0.502, and alignment improved AUC by 0.130 (14/15 patients; Holm-adjusted p = 1.8×10⁻⁴). In the exploratory grid of 25 settings (K = 4–40 trials; d = 2–12 dimensions), the aligned pool was worse than the all-channel decoder in every setting and never significantly better than the within-patient latent decoder. Adding target trials to the pooled LDA changed AUC by less than 0.001. For this task, the tested low-dimensional alignment did not replace within-patient trials. The analysis was offline and limited to one dataset, one session per patient, event-locked windows, linear PCA, and orthogonal Procrustes; it does not rule out other alignment methods or combining aligned data with all-channel decoders.

SEEG · brain-computer interface · few-shot calibration · Procrustes alignment · latent-space alignment · movement decoding · high-gamma activity · AUC