From Microbleeds to Iron: AI Prediction of Cerebrospinal Fluid Erythrocyte Load in Alzheimer's Disease.

Background/Objectives: Cerebrospinal fluid erythrocyte load (CTRED) reflects occult red-blood-cell ingress into brain/CSF and consequent heme-iron exposure, a toxic pathway relevant to Alzheimer's disease (AD). We aimed to develop explainable machine learning (ML) models that classify high vs. low CTRED from routine, largely non-invasive inputs, and to position a blood-first workflow leveraging contemporary plasma amyloid-tau biomarkers. Methods: Twenty-six ADNI participants were analyzed. Inputs were age, sex, mean arterial pressure (MAPres), amyloid (Aβ42), total tau, phosphorylated tau, and hippocampal atrophy rate (APC) derived from longitudinal MRI. APC was computed from normalized hippocampal volumes. CTRED was binarized at the median (0 vs. >0). Data were split into train (n = 20) and held-out test (n = 6). Five classifiers (linear SVM, ridge, logistic regression, random forests, and MLP) were trained in leakage-safe pipelines with stratified five-fold cross-validation. To provide a comprehensive assessment, we presented the contribution AUC, thresholded performance metrics, summarized model performance, and the permutation feature importance (PFI). Results: On the test set, SVM, ridge, logistic regression, and random forests achieved AUC = 1.00, while the MLP achieved AUC = 0.833. Across models, PFI consistently prioritized p-tau/tau, Aβ42, and MAPres; age, sex, and APC contributed secondarily. The attribution profile aligns with mechanisms linking BBB dysfunction and amyloid-related microvascular fragility with tissue vulnerability to heme-iron. Conclusions: In this proof-of-concept study, explainable ML predicted CTRED from routine variables with biologically coherent drivers. Although ADNI measurements were CSF-based and the sample was small, the framework is non-invasive by adding plasma p-tau217/Aβ1-42 for amyloid, tau inputs, and integrating demographics, hemodynamic context, and MRI. External, plasma-based validation in larger cohorts is warranted, alongside extension to MCI and multimodal correlation (QSM, DCE-MRI) to establish clinically actionable CTRED thresholds.