Kernel decoders matched L2 logistic regression, while random forests performed worse, in 18 SEEG patients

We compared nonlinear and linear within-patient decoders of later word recall. We used the band-power features of 18 SEEG patients (OpenNeuro ds004789 FR1, session 0; features from sci c4b8c884). All models used the same 5-fold list-grouped folds and fixed default hyperparameters, set in a plan before any fit. On all channels x 8 frequencies, L2 logistic regression (LR) had median AUC 0.670. A Gaussian-kernel and a quadratic-kernel logistic model were equal to LR (median differences -0.003 and +0.003; 95% CIs from -0.010 to +0.012; Holm p = 0.79). A random forest was worse (-0.022, LR higher in 14 of 18 patients, Holm p = 0.017). An LR + forest ensemble did not help. On 8 channel-mean features, kernel models, forest and a small MLP were also not better than LR. The results were the same within serial position. LR on all channels was better than LR on channel means (+0.057, p = 0.003). Thus, for a memory BCI with about 300 training words, a linear decoder on all channels is sufficient; spatial detail helps more than nonlinearity. Limits: one dataset and session, no tuning, numpy implementations, offline AUC only.

SEEG · word recall · memory decoding · logistic regression · kernel methods · random forest · band power · BCI