Train five model families on the same tabular features, and compare them across three train/test splits rather than one.
Preprocessing dropped identifiers, imputed missing values (median for numeric fields, mode for categorical), label-encoded gender and handedness, standardized the features, and binned the Clinical Dementia Rating into four classes: none, very mild, mild and moderate.
The models: a 1D CNN (three convolution layers, 32 to 128 filters) to learn interactions between features; a two-layer LSTM over sliding windows of three records; an attention-augmented ResNet-style network, for explainability; a 100-tree Random Forest as the interpretable baseline; and a soft-voting ensemble that weights neural and Random Forest predictions 60:40.
Each was evaluated on stratified 70:30, 80:20 and 90:10 splits, to see which models hold up as the amount of training and test data shifts, not just which scores highest once.