Interpretable Machine Learning for Risk Stratification of Hippocampal Atrophy in Alzheimer’s Disease Using CSF Erythrocyte Load and Clinical Data

Background/Objectives: Hippocampal atrophy indicates Alzheimer’s disease (AD) progression and guides follow-up and trial enrichment. Identifying high-risk patients is crucial for optimizing care, but accessible, interpretable machine-learning models (ML) are limited. We developed an explainable ML model using clinical data and CSF erythrocyte load (CTRED) to classify adults with AD as high- or low-risk based on hippocampal volume decline. Methods: Included ADNI participants with ≥2 MRIs, baseline lumbar puncture, and vital signs within 6 months of MRI (n = 26). The outcome was the Annual Percentage Change (APC) in hippocampal volume, classified as low or high risk. Predictors were standardized; models included SVM, logistic regression, and Ridge Classifier, tuned and tested on a set (n = 6). Thresholds were based on out-of-fold predictions under a 10–90% positive rate. Explainability used PFI and SHAP for per-patient contributions. Results: All models gave identical classifications, but discrimination varied: Ridge AUC = 1.00, logistic = 0.889, and SVM = 0.667. PFI highlighted MAPres and sex as main signals; CTRED contributed, and age had a minor impact. Conclusions: The explainable ML model with clinical data and CTRED can stratify AD patients by hippocampal atrophy risk, aiding follow-up and vascular assessment planning rather than treatment decisions. Validation in larger cohorts is needed. This is the first ML study to use CSF erythrocyte load to predict hippocampal atrophy risk in AD.