Fairness and External Generalizability Auditing of a Multimodal Alzheimer's Disease Prediction Framework

To audit a previously developed hybrid multimodal Alzheimer's disease (AD) predictionframework for (i) subgroup performance disparities and (ii) external generalizability, two gapscommonly left unreported in AD prediction studies despite their direct relevance to equitableand safe clinical deployment. Using the identical test-set predictions (n = 233) underlying everyother reported metric, subgroup accuracy and Equalized-Odds metrics were computed for sexand for data-driven age and education tertiles, with 95% bootstrap confidence intervals. Fourclassical baselines, trained on ADNI, were evaluated without refitting on an independentOASIS cohort (n = 235) using five harmonized features. Point-estimate accuracy gaps were4.0 percentage points by sex, 8.6 points across age tertiles, and 7.6 points across educationtertiles, none statistically distinguishable from sampling variability at this sample size. A finergrained Equalized-Odds analysis surfaced a more concrete concern: the youngest age tertile'sAD recall was only 20.0% versus 84.6% in the middle tertile, though based on only 5 ADpositive cases, with a wide confidence interval. False Positive Rates were uniformly low. Allfour classical baselines generalized reasonably well to OASIS (AUC 0.78–0.86); the hybridmodel itself has not yet been externally validated. Point-estimate fairness gaps exist but requirea larger, purpose-built sample to confirm; external validation evidence to date covers only theclassical-baseline components of this line of work, not the hybrid architecture. Fairnessauditing with proper uncertainty quantification and external validation should be standard,mandatory components of AD prediction research, not optional extensions.