#3 · ML / AI methods

Cross-cohort tests expose limits of EEG, vocal, and transcriptomic biomarkers

The common thread is whether biomarkers and representations hold beyond the cohorts used to derive them. A population-aware EEG framework for Parkinson's detection enumerates 75 cross-cohort train/test configurations, making evaluation design the object of study rather than a reporting detail. A published study on resting EEG complexity predicts single-session iTBS responses with 81% accuracy internally, but accuracy drops on external testing and the approach fails to generalize. A blood transcriptomic study introduces temporal entropy to flag unstable dynamics and improve psychiatric classification across public cohorts. Entropy-based vocal features improve depression detection over pooled acoustic baselines (AUC 0.646 vs 0.593), though on DAIC-WOZ alone. Independent-component encoding of fMRI from LLM representations extends the representational line in a different direction.