A longitudinal machine-learning framework trained on 150 subjects and 373 repeated observations from the OASIS-2 dataset achieved 88.16% accuracy and a weighted F1-score of 0.860 in classifying patients as Non-demented, Demented, or Converted. The model engineered slope-based temporal features from Clinical Dementia Rating (CDR), Mini-Mental State Examination scores, normalized whole-brain volume, and intracranial volume across successive visits. SHAP explainability analysis identified CDR slope as the dominant longitudinal predictor, outperforming single-visit CDR alone.
This preprint — not yet peer-reviewed — tackles a genuine gap in dementia research: most predictive models treat visits as independent snapshots, discarding the trajectory information that clinicians actually rely on. By constructing rate-of-change features between visits, the framework better captures the dynamic nature of neurodegeneration. However, critical limitations temper enthusiasm. The cohort of just 150 subjects is exceptionally small for a machine-learning claim of this magnitude, raising serious concerns about generalizability. The "Converted" subgroup — arguably the most clinically valuable class to detect early — remained poorly classified, with macro F1 of only 0.776 revealing substantial class-level weakness behind the headline accuracy. Patient-level data splitting reduces but does not eliminate leakage concerns. CDR itself incorporates cognitive and functional assessments that partially encode the outcome, creating circularity risk. External validation in diverse, larger longitudinal cohorts is essential before any clinical translation. Incremental rather than paradigm-shifting, this work nonetheless demonstrates that temporal feature engineering meaningfully augments static biomarker models.
