Integrating MRI Volume and Plasma p-Tau217 for Amyloid Risk Stratification in Early-Stage Alzheimer Disease

Abstract

Background and objectives: Identifying β-amyloid (Aβ) positivity is crucial for selecting candidates for Aβ-targeted therapies in early-stage Alzheimer disease (AD). While Aβ PET is accurate, its high cost limits routine use. Plasma p-tau217 testing offers a less invasive option but also incurs additional costs. Structural brain MRI, routinely used in cognitive assessments, can identify features predictive of Aβ positivity without extra expense. We evaluated a 2-stage workflow integrating MRI-based features and plasma p-tau217 to efficiently predict Aβ PET positivity in early-stage AD.

Methods: This prospective cohort study included participants with mild cognitive impairment (MCI) or early Alzheimer-type dementia (ATD) from the Korea-Registries to Overcome Dementia and Accelerate Dementia Research (K-ROAD; Korea) and Alzheimer's Disease Neuroimaging Initiative (ADNI; US) cohorts. Eligible participants had a Clinical Dementia Rating score of 0.5, along with MRI, plasma p-tau217, and Aβ PET data. A random forest classifier predicting Aβ PET positivity was developed using MRI-based brain atrophy patterns and APOE ε4 status. Participants were stratified into low-risk, intermediate-risk, and high-risk groups; plasma p-tau217 testing was performed only in intermediate-risk individuals. Outcomes included positive predictive value (PPV), negative predictive value (NPV), and overall accuracy.

Results: A total of 807 K-ROAD participants (median age 72.0 years, 58.7% female) and 230 ADNI participants (median age 70.9 years, 49.1% female) were analyzed. Using a 95% sensitivity/specificity strategy, the low-risk group demonstrated NPVs of 94.7% (91.7%-97.7%, K-ROAD) and 99.0% (97.0%-100.0%, ADNI). The high-risk group showed PPVs of 97.6% (95.9%-99.3%, K-ROAD) and 98.8% (96.5%-100.0%, ADNI). Intermediate-risk groups comprised 33.3% (K-ROAD) and 20.9% (ADNI) of participants. Plasma p-tau217 testing in intermediate-risk groups yielded PPVs of 92.5% (88.7%-96.3%, K-ROAD) and 90.0% (79.0%-100.0%, ADNI) and NPVs of 83.1% (75.0%-91.2%, K-ROAD) and 83.3% (66.1%-100.0%, ADNI). The overall workflow accuracy was 94.2% (92.6%-95.8%, K-ROAD) and 96.5% (94.1%-98.9%, ADNI).

Discussion: The 2-stage diagnostic workflow integrating MRI-based risk stratification and plasma p-tau217 testing accurately identified individuals with Aβ PET positivity in early-stage AD, substantially reducing the need for additional biomarker testing. However, the generalizability may be limited by modest incremental improvement over baseline models and limited racial and ethnic diversity.

Figure 1. Proposed Framework

The framework illustrates the results of an AI-powered brain MRI analysis that assesses brain atrophy using quantitative information from MRI scans. The 2-step framework for Aβ PET status classification aims to minimize additional confirmatory tests while ensuring accurate patient stratification. In the first step, a random forest model incorporating APOE ε4 status and regional CSF volume predicts Aβ PET positivity, stratifying participants into 3 risk categories. In the second step, plasma p-tau217 testing is applied exclusively to the intermediate-risk group to refine predictions and facilitate risk-based decision making. *The classification of Aβ-positive and Aβ-negative cases is determined based on Aβ PET status. Aβ = β-amyloid; CU = cognitively unimpaired; LV = lateral ventricle; L/R = left and right; ROI = region of interest.

Figure 2. Distribution and Thresholds of Probability of Aβ (+) Based on MRI-Derived Brain Atrophy Patterns for the K-ROAD Cohort (A) and for the ADNI Cohort (B)

Blue dots represent Aβ PET–negative patients while red dots indicate Aβ PET–positive patients. The right y-axis indicates probability values aligned with evaluated 3 risk thresholds, accompanied by metrics defining Se (90%, 95%, 97.5%) and Sp (90%, 95%, 97.5%). Aβ = β-amyloid; ADNI = Alzheimer's Disease Neuroimaging Initiative; K-ROAD = Korea-Registries to Overcome Dementia and Accelerate Dementia Research; Se = sensitivity; Sp = specificity.

Figure 3. Overall Workflow Accuracy and Cost-Based Approach

(A) The overall accuracy of the 2-step workflow, reflecting the proportion of correct classifications in both low-risk and high-risk groups, along with the accuracy of plasma p-tau217 classifications for the intermediate-risk group, was calculated for the K-ROAD cohort. The error bars correspond to 95% CIs. (B) Bar plots indicating the percentage of the intermediate-risk group by applying the risk stratification strategy, based on each of the risk threshold strategies. (C) The cost estimation for each risk threshold strategy was determined by incorporating additional expenses associated with the plasma p-tau217 test, which was used to assess Aβ PET positivity within the intermediate group. Aβ = β-amyloid; K-ROAD = Korea-Registries to Overcome Dementia and Accelerate Dementia Research.

Figure 4. Two-Stage Diagnostic Workflow for Aβ (+) on PET Prediction Integrating MRI-Based Risk Stratification and Plasma p-Tau217 Testing in K-ROAD (A) and ADNI (B) Cohorts

The 2-stage diagnostic workflow, based on the 95% Se and 95% Sp threshold strategies, is summarized for the K-ROAD (A) and ANDI (B) cohorts. On the right, the results of the first step, MRI-based risk stratification, are displayed, with red, yellow, and blue dots representing individuals in the high-risk, intermediate-risk, and low-risk groups, respectively. The percentage of Aβ PET positivity in the high-risk group and the percentage of Aβ PET negativity in the low-risk group are shown, reflecting the predictive accuracy of Aβ PET status. To the left, the results of the second step, plasma p-tau217 testing conducted exclusively for the intermediate-risk group, are presented. The predictive accuracy for Aβ PET status is represented using the NPV and PPV. Aβ = amyloid-β; ADNI = Alzheimer's Disease Neuroimaging Initiative; K-ROAD = Korea-Registries to Overcome Dementia and Accelerate Dementia Research; NPV = negative predictive value; PPV = positive predictive value; Se = sensitivity; Sp = specificity.

Figure 5. Model Performance and Predictive Value Across Classification Strategies

ROC curves in the K-ROAD cohort (A) and in the ADNI cohort (B) comparing the 1-step (MRI-only) model and the proposed 2-step model. (C) Comparison of PPV and NPV between the 1-step (MRI-only) and 2-step models in the intermediate-risk group across cohorts. ADNI = Alzheimer's Disease Neuroimaging Initiative; K-ROAD = Korea-Registries to Overcome Dementia and Accelerate Dementia Research; NPV = negative predictive value; PPV = positive predictive value; ROC = receiver operating characteristic.