Abstract / Summary
Alzheimer's diagnostic work-up often moves quickly toward expensive investigations such as MRI and PET. This work presents a cost-aware, four-stage cascading machine learning framework (cognitive screening, fluid biomarkers, structural MRI, PET) that escalates a patient to the next stage only when model-estimated risk crosses a threshold. Using 2,419 baseline ADNIMERGE records, Random Forest, XGBoost and MLP classifiers were compared at each of the first three stages. A five-feature cognitive model reached a test ROC-AUC of 0.921, with biomarker and MRI stages adding modest gains (0.925 and 0.940). The paper also discusses limitations, including label circularity, imputation strategy and uncalibrated probabilities. The system is a research prototype for triage support and is not a diagnostic tool. Code: https://github.com/sanidhya2506/Alzheimers_Cascading_Multimodal