Title : Adaptive multimodal artificial intelligence for early Alzheimer’s disease risk assessment and progression forecasting using plasma biomarkers, neuroimaging, and clinical data
Abstract:
Accurate risk stratification in Alzheimer's disease is complicated by a fundamental issue: the most informative biological signals span blood, brain imaging, and clinical history — data types that are rarely interpreted together in routine practice. I present an adaptive multimodal AI platform, developed as a computational framework for Alzheimer's disease risk modeling, that integrates plasma biomarkers including the amyloid-beta 42/40 ratio, phosphorylated tau-217, neurofilament light chain, GFAP, soluble TREM2, and YKL-40; structural MRI features such as hippocampal volume, cortical thickness, ventricular size, and regional patterns of atrophy; FDG-PET measures of cerebral glucose metabolism where available; and routine clinical variables including age, sex, and cognitive assessment scores into a unified framework for risk prediction, disease trajectory modeling, and clinical decision support. Predictions are generated using a confidence-weighted ensemble of logistic regression, random forests, gradient boosting, support vector machines, and neural networks, with weights adjusting dynamically by disease stage and robustness maintained across heterogeneous, longitudinal, and partially missing data — conditions that characterize real clinical cohorts. Each output includes a normalized 0–100 risk score with an associated confidence estimate, feature-level attribution identifying which biomarkers or imaging measures most influenced the prediction, longitudinal projections of evolving risk, and plain-language reports generated separately for clinicians and patients. Validation on publicly available longitudinal cohorts will assess predictive performance, calibration, temporal stability, and bias across diverse patient populations.
