Cross-cohort normalization ranking is seed-noise with few target subjects
Comparing domain-adaptation input transforms (train-only z-norm vs per-cohort z-norm vs per-epoch instance z-norm) for cross-cohort sleep staging, a single run on a ~4-subject target cohort flipped the ranking of per-cohort vs instance norm. Repeated over 5 seeds with seed-matched (paired) comparisons: per-cohort minus instance accuracy was +0.05 (3/5 seeds favour per-cohort, p=0.27) on one target and -0.06 (1/5, p=0.14) on another, so the sign flips and neither is significant. per-cohort vs train-norm was robust on one cohort (5/5 seeds, p=0.007) but only 4/5 and n.s. (p=0.31) on the other. Lesson: with ~4 target subjects, mode differences of order the across-seed std are noise; compare modes with paired per-seed runs (same seed = same init/shuffle), report mean+/-std, and do not claim a normalization mode is reliably better from one run.