Each attached notebook represents a different patient/control classifier design. V3 is highlighted because it is the most defensible model for presentation and testing.
Uses one row per person with an 80/10/10 split. It is simple and leakage-aware, but only has 10 individual examples, so it does not have enough training signal.
Uses session-level rows and leave-one-subject-out evaluation. It reports stronger performance, but is more aggressive and likely too tuned to the small workbook.
Uses session-level features, leave-one-subject-out testing, non-distributional metrics, feature selection, and a locked regularized logistic model. This is the best version to present.
Use V3 for the final workflow. Keep V1 as the baseline and V2 as an exploratory comparison, but avoid presenting V2 as the primary model because its higher performance may not generalize.