E79 index from routine biomarkers stratifies rapid MCI-to-Alzheimer's conversion: AUC 0.891 in 607 ADNI patients

A team led by John D. Mayfield, on behalf of the Alzheimer's Disease Neuroimaging Initiative (ADNI), described the E79 Alzheimer index, a tool that stratifies patients with mild cognitive impairment (MCI) by risk of rapid progression to Alzheimer's disease within 24 months. The work was published in Alzheimer's Research & Therapy and used data from 607 MCI participants. The index reached an AUC of 0.891 (bootstrap 95% CI 0.863–0.916) and uses only baseline clinical data, without genetic testing or advanced imaging biomedcentral.comAlzheimer s Research & Therapybiomedcentral.comOpen source ↗.

Identifying rapid progressors matters for anti-amyloid therapy candidacy, monitoring intensity and prognosis. According to the authors, existing machine learning approaches report high AUC for any-conversion prediction but conflate slow and rapid progressors, which limits their clinical use.

Spectral decomposition of ten biomarkers

The input panel contained ten routine biomarkers: cognitive, functional, structural MRI and cerebrospinal fluid measures. The authors computed a 10×10 Pearson correlation matrix and applied eigendecomposition to it. This revealed two disease-sensitive eigenmodes: E7, reflecting cognitive-functional instability, and E9, reflecting structural-pathological burden.

Individual biomarker profiles were projected onto the MCI-derived eigenvectors to obtain subject-level E7 and E9 scores. These scores were the inputs to a logistic regression model biomedcentral.comAlzheimer s Research & Therapybiomedcentral.comOpen source ↗.

Validation and comparison

On a stratified hold-out test set of 122 participants the AUC was 0.862; 5-fold cross-validation gave 0.880. At a fixed 90% sensitivity, the model showed 95.4% negative predictive value and 67.4% specificity.

Between the high-risk and low-risk groups the authors report an 80-fold risk enrichment: 47.9% of the high-risk group converted within 24 months versus 0.6% of the low-risk group.

Using DeLong's method with Bonferroni correction, E79 was statistically equivalent to logistic regression on all raw biomarkers (AUC 0.862 vs. 0.859, p = 1.000). It was significantly better than PCA compression of the same dimensionality (0.862 vs. 0.840, p = 0.004). Unlike PCA, the approach keeps one-to-one attribution to individual biomarkers biomedcentral.comAlzheimer s Research & Therapybiomedcentral.comOpen source ↗.

The main limitation is derivation in a single ADNI cohort. External validation in ADNI4 is underway as follow-up matures through 2026–2027.