Predicting Progression from Mild Cognitive Impairment to Alzheimer’s Disease using Survival Analysis

Identifying which individuals with mild cognitive impairment (MCI) will progress to Alzheimer’s disease (AD) is critical for therapeutic planning and clinical trial enrollment. We used Random Survival Forest models to evaluate the prognostic value of structured blood biomarkers (p-Tau217, A$\beta$42/40, NfL, GFAP), amyloid-$\beta$ PET imaging, APOE ~$\varepsilon$4 genotype, and demographic data to predict MCI-to-AD conversion in the Alzheimer’s Disease Neuroimaging Initiative (n=201). At near 4 years, all biomarker-based models outperformed a demographics-and-APOE ~$\varepsilon$4 baseline, with a multi-blood biomarker panel achieving the highest AUROC (0.818, 95\% CI [0.748--0.888]). However, p-Tau217 alone (AUROC = 0.787 [0.705--0.868]) performed comparably to both amyloid-$\beta$ PET (0.758 [0.669--0.848]) and the multi-marker panel. Partial dependence analysis revealed a critical window in which increasing p-Tau217 levels from 0.08 to 0.56 pg/mL corresponded to a rise in predicted 4-year conversion probability from approximately 10\% to 45\%. A two-cutoff thresholding strategy to stratify high (20\%), intermediate (31.4\%), and low (48.6\%) risk groups based on 90\% specificity and 90\% sensitivity had a negative predictive value of 83.9\% and positive predictive value of 44.4\%. These findings support p-Tau217 as a minimally invasive and scalable prognostic marker and demonstrate the value of survival modeling in structured clinical data for risk stratification in prodromal AD.