Predicting MCI-to-Alzheimer's Conversion

Full-cycle machine learning on the OASIS-3 longitudinal multimodal neuroimaging dataset (1,378 participants, 277 MCI patients, a working cohort of 92 after strict cohort definition), forecasting conversion from Mild Cognitive Impairment to Alzheimer's Disease. Integrated 10+ clinical, cognitive, neuroimaging and biomarker files into a unified session- and patient-level dataset. Led exploratory analysis and QA across demographics, CDR and 107 cognitive-assessment variables, building a validated data dictionary against official clinical documentation and resolving encoding errors, out-of-range values and structured missingness. Developed an automated clustering pipeline using exact patient-matching to identify cognitive-test sub-batteries, reducing dimensionality and driving reproducible feature selection. Built and evaluated classification and survival models with patient-level GroupKFold cross-validation and Optuna tuning, and applied SHAP to identify the strongest predictive biomarkers: hippocampal volume, CDR, MMSE. Prototyped a Streamlit clinical decision-support interface returning conversion-risk scores, feature-importance visualisations and downloadable PDF reports, informed by a structured requirements questionnaire run with medical professionals.