The E79 Alzheimer index: a spectral multimodal biomarker tool for stratification of rapid MCI-to-AD conversion

Identifying which patients with mild cognitive impairment (MCI) will progress rapidly to Alzheimer’s disease (AD) within 24 months has direct implications for anti-amyloid therapy candidacy, monitoring intensity, and prognosis. Current machine learning approaches report high area under the curve (AUC) values for any-conversion prediction but conflate slow and rapid progressors, limiting clinical utility. A fundamentally different approach is needed: one that targets the rapid-conversion subphenotype with interpretable, routine biomarker inputs. We applied spectral decomposition to a \(10 \times 10\) Pearson correlation matrix of 10 routine multimodal biomarkers (cognitive, functional, structural MRI, and CSF) from 607 MCI participants in the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Eigendecomposition identified two disease-sensitive eigenmodes: E7 (cognitive-functional instability) and E9 (structural-pathological burden). Subject-level E7 and E9 scores were derived by projecting individual biomarker profiles onto MCI-derived eigenvectors and served as inputs to logistic regression. Performance was evaluated on a stratified hold-out test set ( \(n = 122\) ) and validated by 5-fold cross-validation. Models were benchmarked against raw biomarker and PCA-based comparators using DeLong’s method with Bonferroni correction. E79 achieved AUC 0.891 (bootstrap 95% CI: 0.863-0.916), AUC 0.862 on the held-out test set, and 5-fold CV AUC 0.880. At 90% sensitivity, E79 yielded 95.4% negative predictive value (NPV), 67.4% specificity, and 80-fold risk enrichment between high-risk (47.9% conversion) and low-risk (0.6% conversion) groups. E79 was statistically equivalent to raw biomarker logistic regression (AUC 0.862 vs. 0.859, p = 1.000) and significantly superior to PCA-based compression of identical dimensionality (AUC 0.862 vs. 0.840, p = 0.004). The primary limitation is single-cohort ADNI derivation; external validation in ADNI4 is underway as follow-up matures through 2026-2027. Spectral decomposition of routine multimodal biomarkers identifies two interpretable eigenmodes that stratify rapid MCI-to-AD conversion with 80-fold risk enrichment and 95.4% NPV using only baseline clinical data, without genetic testing or advanced imaging. The organizing principle of compression determines clinical utility: disease-sensitive spectral decomposition outperforms variance-based PCA of identical dimensionality while providing one-to-one biomarker attribution.