Longitudinal Structural MRI Change Improves Alzheimer’s Disease Classification in a Large ADNI Cohort

Introduction: A single structural MRI examination captures anatomy at one time point, whereas longitudinal change may more directly reflect neurodegeneration. We tested whether MRI-derived change rates improve Alzheimer’s disease classification beyond baseline anatomy.

Methods: We assembled 2,037 Alzheimer’s Disease Neuroimaging Initiative participants with at least two FreeSurfer examinations after an MRI aligned to the earliest diagnosis (731 were cognitively normal, 984 had mild cognitive impairment, and 322 had Alzheimer’s disease/dementia). The median MRI-to-diagnosis offset was 2 days; participants had a median of 4 scans over 3.0 years. Baseline bilateral hippocampal and ventricular volumes were normalized by intracranial volume, and mean entorhinal thickness was retained in native units. Annualized change was estimated by participant-specific linear slopes. Baseline MRI, change, and their combination were compared using L2-regularized logistic regression and extremely randomized trees. Participant-level five-fold cross-validation was repeated twice, with preprocessing restricted to training folds. Paired bootstrap resampling quantified AUC differences.

Results: For Alzheimer’s disease versus cognitively normal classification (n = 1,053), adding change to baseline MRI increased AUC from 0.930 to 0.957 with logistic regression (difference, 0.028; 95% CI, 0.015–0.040; p < 0.001) and from 0.929 to 0.964 with extremely randomized trees (difference, 0.035; 0.025–0.045; p < 0.001). Balanced accuracy increased from 0.867 to 0.905 and from 0.855 to 0.906, respectively. For three-class classification, macro-AUC increased from 0.711 to 0.733 and from 0.704 to 0.754, respectively (both p < 0.001). The direction of improvement was retained across temporal-alignment, follow-up-duration, scan-count, and acquisition-schedule sensitivity analyses.

Conclusions: Longitudinal structural MRI change provided reproducible incremental information beyond baseline anatomy. Because follow-up imaging was used, these findings support retrospective longitudinal disease-state profiling rather than baseline-only prognostic prediction.