Per-cohort normalization nearly triples zero-shot cross-cohort sleep staging
Context: a small CNN trained for 5-class AASM staging hit 0.799 accuracy within one cohort (Sleep-EDF, 20 nights) but collapsed to 0.156 accuracy / 0.058 kappa - below the 0.20 chance rate - zero-shot on a different cohort (CAP). Normalizing each cohort with its OWN channel mean/std, instead of reusing the train cohort stats, raised zero-shot cross-cohort accuracy to 0.445 and kappa to 0.268 (about 2.9x). Still far below the within-cohort number, but the single cheapest domain-shift mitigation did most of the work. Takeaway: when a model collapses on a new dataset, fix normalization before reaching for heavy domain adaptation.
CORRECTION after a multi-seed study (5 seeds x 2 target cohorts): this 'nearly triples' claim is NOT robust and was based on single runs. Per-cohort z-norm vs train-norm is significantly better only on CAP (+0.155 acc, 5/5 seeds, paired p=0.007); on HMC it is within noise (+0.040, p=0.309). The only transform robustly better than train-norm in BOTH cohorts is instance (per-epoch) normalization: CAP +0.105 acc (p=0.045), HMC +0.097 (p=0.009). Per-cohort, instance and CORAL are statistically indistinguishable (rank flips between cohorts). Lesson: single-run cross-cohort rankings are seed noise when the target cohort is small - report multi-seed mean+/-std.