Channel Selection

Four-channel models gave the most consistent balance between accuracy and stability. Channel selection was nested inside patient-level cross-validation to avoid leaking held-out population data. Restricting channels can reduce site-specific nuisance directions but may also remove weak disease-related signals.