The 95% bootstrap interval for cohort prediction gain from normalized thickness differences included zero in 145 AOMIC test participants

We tested whether normalized left-right cortical thickness differences improved prediction of the source cohort. We used the supplied AOMIC PIOP1 and PIOP2 version 2.0.0 tables. Fixed rules retained 433 people: 288 for training and 145 for test. The reference model used age, recorded sex, head size, and mean thickness. Adding 34 indices, 2*(L-R)/(L+R), reduced test log loss from 0.73685 to 0.68136 nats. The gain was 0.05549 nats. A 1,000-draw bootstrap sampled training and test people and fitted both models again. Its 95% interval was -0.06072 to 0.13054 nats and included zero. The interval assumes independent people and fixed source cohorts, split, and folds. This result does not establish absence of a cohort difference. A planned comparison with 34 paired means gave test loss 0.23770 nats and gain 0.49915 nats. All six planned checks are reported. These data cannot identify scanner or neural causes. The package preserves unchanged source files, fixed URLs and hashes, code, the local plan, predictions, bootstrap records, figures, and instructions. A separate parser and optimizer passed. A full analysis from fresh public downloads reproduced all 32 result files byte for byte.

cortical thickness · brain asymmetry · cohort prediction · bootstrap · AOMIC · log loss