Improving cognitive impairment classification in Alzheimer's disease using tabular diffusion augmentation and late fusion deep learning

Alzheimer's disease is the most common cause of dementia and is often detected after substantial neurodegeneration has occurred. Distinguishing cognitively impaired individuals from cognitively normal ones using genetic and demographic data can help characterize the genetic contributions to impairment. Deep learning can capture complex, nonlinear genetic relationships among these features, but its use in clinical genomics is limited by the small size of patient cohorts.

We evaluated whether TabDDPM-based synthetic augmentation could improve cross-sectional classification of cognitive impairment using demographic and APOE-region genotype features. From 542 ADNI participants, we created a stratified 433-participant development cohort (80%) and a 109-participant held-out cohort (20%). Eight predictive models were evaluated using repeated stratified 5-fold cross-validation with 10 repeats. The 346 or 347 participants in the outer training partition were divided into three internal TabDDPM subsets: 80% for generator training, 10% for hyperparameter selection, and 10% reserved for internal generator evaluation. Run-specific synthetic samples were added to the outer training partition at 3×, 5×, 7×, and 10× augmentation. The selected predictive model configuration and augmentation ratio were then evaluated on the 109-participant held-out cohort for assessment.

The synthetic data showed similar demographic structure and common-variant genotype distribution to the real cohort. Synthetic augmentation improved mean AUC across nonlinear models. The best-performing model, a categorical late-fusion neural network, achieved a mean cross-validated AUC of 0.781 with 7× augmentation, compared with 0.738 using real data alone. For the 7× versus 0× comparison, the internal cross-validation analysis yielded a raw paired DeLong p-value of 0.0012 and a Holm-adjusted p-value of 0.03. This improvement was observed in the 109-participant held-out cohort: AUC increased from 0.721 to 0.769, with gains in both sensitivity (0.68-0.77) and specificity (0.58-0.66).

These results suggest that synthetic data augmentation can improve deep learning performance in data-constrained clinical genomic classification. The approach does not replace real patient data, but it can help models learn more stable patterns from cohorts with limited sample size. Because the framework uses standard tabular genomic and demographic features, it may extend to other genomic classification tasks where individual-level patient data are limited.