A multimodal machine learning model integrating plasma biomarkers and MRI metrics for non-invasive prediction of amyloid-β pathology in mild cognitive impairment

Background Accurate, non-invasive prediction of cerebral amyloid-β (Aβ) pathology in mild cognitive impairment (MCI) remains challenging yet critical for early intervention. Objective To develop a multimodal machine learning model integrating clinical features, plasma biomarkers, and structural MRI metrics for non-invasive Aβ prediction. Methods Data were obtained from the Alzheimer's Disease Neuroimaging Initiative. Participants with concurrent plasma biomarkers, 3D T1-weighted MRI, and amyloid assessments were included. Logistic Regression, Decision Tree, and Support Vector Machine models were constructed using clinical, plasma, MRI, and combined features. Performance was evaluated via internal validation and external testing in a cognitively unimpaired cohort using AUC, calibration curves, and decision curve analysis. The prognostic value of the model-derived Aβ risk probability was assessed using Cox regression in an independent longitudinal MCI cohort. Results The optimal Logistic Regression model incorporated APOE ε4 status, Mini-Mental State Examination score, plasma p-Tau217, Aβ 42 /Aβ 40 ratio, and bilateral hippocampal and left amygdalar volumes. The combined model achieved an AUC of 0.875 in internal validation and maintained robust performance in the external unimpaired cohort (AUC = 0.883), outperforming single-modality models. The predicted Aβ-positive risk probability effectively stratified disease progression risk in MCI patients (C-index = 0.771). Conclusions A multimodal model integrating plasma and MRI features accurately predicts Aβ pathology and progression risk, offering a practical non-invasive tool for early Alzheimer's disease screening and risk stratification.