Background: Accurate prediction of dementia onset is critical for clinical trial design but often fails to capture dynamic disease progression. Methods: We benchmarked four survival architectures using latent neurodegenerative signatures ( \(z\) -scores) and genetic data (APOE4) over a 15-year horizon, evaluating integrated predictive performance focused on a 10-year study period. Results: Parametric Accelerated Failure Time (AFT) models, specifically the Log-Normal architecture, achieved the highest discriminative accuracy ( \(C\text {-index}=0.765\) , \(iAUC=0.791\) ). Sensitivity analysis revealed that while APOE4 dominates late risk, latent biomarkers drive accuracy at early stages of the disease, especially the \(z_2\) in the 2.5–10 years, eventually subsuming traditional demographic signals. Conclusion: Latent phenotypic markers are essential for short-term prognosis, requiring model architectures matched to specific clinical objectives.