Interpretable multi-horizon prediction of progression from mild cognitive impairment to Alzheimer’s disease: cross-cohort transportability and web-based risk calculator development

Abstract Background Progression from mild cognitive impairment (MCI) to Alzheimer’s disease (AD) dementia is heterogeneous, highlighting the need for baseline prognostic models that provide individualized risk estimates across clinically relevant horizons. However, existing models often focus on single prediction horizons, incompletely quantify the incremental contribution of individual modalities, or lack external evidence of transportability beyond the development cohort. We aimed to quantify the incremental value of structural magnetic resonance imaging (MRI) across prediction horizons and evaluate model transportability under restricted predictor availability. Methods We developed a baseline fixed-horizon framework to estimate the risk of first clinical conversion from MCI to dementia at 12, 24, and 36 months using cohort-specific diagnostic definitions. Predictors were obtained at, or aligned to, the baseline MCI visit, whereas longitudinal follow-up was used only to ascertain conversion outcomes. Models were developed in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) using demographic, cognitive-functional, and structural magnetic resonance imaging (MRI) features, with apolipoprotein E ε4 allele (APOE ε4) assessed as an extension. Transportability was evaluated using a prespecified common-clinical model developed in ADNI and externally validated in the National Alzheimer’s Coordinating Center (NACC) using harmonized variables available across cohorts. Model performance was assessed using discrimination, calibration, Brier score, decision curve analysis, and risk stratification. Model interpretation was performed using SHapley Additive exPlanations (SHAP). Results The selected Model 2 logistic regression framework demonstrated consistent discrimination across prediction horizons, with AUCs ranging from 0.797 to 0.859 in ADNI. The common-clinical model maintained moderate discrimination in NACC using harmonized clinical variables, with AUCs ranging from 0.720 to 0.758. Risk stratification identified distinct conversion-risk groups across horizons. Structural MRI provided consistent incremental prognostic information beyond clinical variables, whereas APOE ε4 yielded limited additional value after incorporation of imaging features. Conclusions A baseline fixed-horizon framework provided interpretable estimates of MCI-to-AD dementia conversion risk across multiple clinically relevant horizons. Structural MRI contributed consistent incremental prognostic information, and a common-clinical model demonstrated moderate transportability under restricted predictor availability. Prospective validation, local recalibration, and evaluation in independent cohorts with broader feature availability are required before clinical implementation.