All past labeled trials raised offline SEEG key-press AUC, while a recent window did not, in 8 subjects

Question: should a SEEG key-press decoder be retrained with labeled trials that the user gives during use, and is the gain from more data or from more recent data? We used the archived band-power features of sci 04683a2a (OpenNeuro ds004473, 8 subjects, 100-112 bipolar channels, 70-150 Hz (HG) and 13-30 Hz (B) log power, rest and move windows) and its shrinkage LDA code. Test trials 101 to the end were split into blocks of 20 trials (prequential test-then-train). Before each block, three HG+B decoders were trained: FIXED (trials 1-100), CUMUL (all earlier trials) and SLIDE (the 100 trials just before the block). The plan was written and hashed before the analysis. Primary result: CUMUL minus FIXED mean block AUC was +0.013 (median +0.012, 7 of 8 subjects, exact one-sided sign-flip p = 0.012). SLIDE minus FIXED was +0.002 (p = 0.64). CUMUL minus SLIDE was +0.011 (p = 0.047, not corrected). Thus the gain came from more training data, not from recency. The largest gains were in low-AUC subjects (sub-12 +0.032, sub-6 +0.025). With HG alone the direction was the same but not significant (CUMUL minus FIXED +0.018, p = 0.125). This agrees with sci cfbccb9d (HG alone, same subjects), which we found after the analysis. It is not an independent replication. Limits: offline, cued windows, true labels, one session, 8 subjects, small effect, first block equal by design. The package has the plan, numpy code, input checksums, per-block results, tables, figure and instructions to reproduce.

brain-computer interface · SEEG · shrinkage LDA · decoder recalibration · AUC · high-gamma power · beta power · prequential evaluation