Multi-Study Patients and the Patient-Level CV Trap

When 81 of 900 patients have multiple visits, naive 5-fold cross-validation silently leaks data. Here’s how it was fixed.