Consensus Risk Modeling and Uncertainty Quantification of Alzheimer’s Disease Using 5ADCSI Plasma Biomarkers and Multiple External Machine-Learning Frameworks

Background

Blood-based biomarkers are increasingly used to identify Alzheimer’s disease (AD)-related pathology, but differences in p217tau assay methodology, training cohorts, and model-development context can substantially influence machine-learning (ML) predictions. Whether emerging biomarker platforms preserve biologically meaningful AD-related information across independently developed ML frameworks remains incompletely understood.

Objective

To evaluate the biological coherence and translational consistency of plasma biomarker measurements generated using the 5ADCSI platform by applying multiple externally trained ML frameworks and developing a consensus-risk approach that integrates framework predictions while quantifying prediction uncertainty.

Methods

Plasma biomarker measurements from 472 participants in the Louisville Twins Study were analyzed using three independently trained ML frameworks: an A4- derived model using the Lilly p217tau MSD assay and two ADNI-derived models using Quanterix Simoa p217tau measured with either the AlzPath or Janssen antibody.

Framework-specific predictions of amyloid positivity probability and predicted centiloid burden were integrated into consensus amyloid risk, consensus centiloid burden, and composite consensus AD-risk scores. Prediction uncertainty and rank instability were used to characterize framework agreement and participant-level classification stability.

Results

All three frameworks recognized biologically coherent AD-related signal despite differences in training cohort and assay methodology. Agreement was strongest between the A4-MSD and ADNI-AlzPath frameworks, whereas agreement involving the ADNI-Jan framework was weaker. Consensus-risk modeling identified a reproducibly high-risk subgroup characterized by elevated consensus-risk scores, low prediction uncertainty, and low rank instability. Participants prioritized by the consensus framework were enriched for APOE ε4 burden, p-tau217, p-tau217/Aβ42, and GFAP, while discordant high-risk participants exhibited substantially greater framework disagreement.

Conclusions

Plasma biomarker measurements generated using the 5ADCSI platform preserve biologically meaningful AD-related information that is consistently recognized across multiple independent ML frameworks. Consensus-risk modeling provides a practical strategy for integrating complementary information from external biological reference models while explicitly characterizing prediction uncertainty, thereby supporting evaluation of emerging blood-based biomarker platforms when direct pathological validation is unavailable.