The challenge
A longitudinal Alzheimer’s research dataset contains hundreds of variables per visit. The task was to build a classifier for three diagnostic groups at the first visit and understand which variables actually drive the result.
The solution
A set of clinically motivated baseline variables: demographics, genetic risk (APOE4), cognitive test scores and MRI-derived volume measures normalised to intracranial volume. Random Forest and SVM were compared in a pipeline with missing-value imputation, and permutation importance showed which variables the model used.
How it was built
The project was carried out in a group; I contributed across data processing, modelling and evaluation. The data comprised 1,658 participants at their first visit (508 cognitively normal, 836 with mild cognitive impairment and 314 with Alzheimer’s), with transition diagnoses excluded. The data were split in a stratified way into training, validation and test (60/20/20); hyperparameters were chosen by cross-validation on the training data, and the test set was used once at the end. An ablation compared models with and without cognitive test scores, and an additional experiment tried to predict conversion from mild cognitive impairment to Alzheimer’s.
Result
The final Random Forest reached a macro-F1 of 0.93 on the test set (332 participants). The ablation showed, however, that the result is driven almost entirely by the cognitive test scores: without them, validation accuracy fell to about 0.53, while the MRI volume measures added little. Predicting future conversion from mild cognitive impairment to Alzheimer’s was far harder, with a balanced accuracy of about 0.58.
Context
Cognitive test scores are part of the basis for the diagnosis itself, so the high accuracy is partly circular and says little about early detection. This is a classification experiment on a retrospective research dataset, not a clinical tool.