Do More Expressive Probes Improve EEG Age Prediction? A Cross-Cohort Study with Frozen REVE

Do more expressive prediction heads improve age prediction from frozen EEG representations? We compared four heads using a fixed REVE encoder, the same Healthy Brain Network development split, and 10 training seeds. The primary external evaluation included 75 MIPDB participants. The mean-pooled linear baseline achieved a mean Pearson correlation of 0.644. The three alternatives increased correlation by 0.008–0.032 on average, but all 95% confidence intervals included zero when both seeds and participants were resampled. None met our prespecified criterion for a reliable improvement. The head with the highest correlation differed from the head with the lowest prediction error. Exploratory analyses varied training-set size and representation layer on the same MIPDB participants. A separate evaluation on 126 participants from OpenNeuro ds006780 also found a positive mean difference at 800 training subjects, but its interval included zero. These results illustrate why repeated training runs alone cannot assess uncertainty in external performance. Comparisons of frozen EEG representations should report the prediction head, training-set size, evaluation metric, and uncertainty across participants. This study does not establish equivalence between heads; its findings apply to the tested REVE setup.