Diffusion augmentation boosts cognitive-impairment classification in 542-person ADNI study

Synthetic records generated with a tabular diffusion model improved a deep-learning system’s ability to distinguish cognitively impaired participants from cognitively normal participants using demographic and genetic information. The strongest model, a categorical late-fusion neural network, achieved a mean cross-validated area under the receiver operating characteristic curve, or AUC, of 0.781 with sevenfold augmentation, compared with 0.738 using real data alone. In a separate held-out cohort of 109 participants, AUC increased from 0.721 to 0.769.

The cross-sectional study used demographic variables and genotype features from the APOE region of 542 participants in the Alzheimer’s Disease Neuroimaging Initiative, or ADNI. The researchers created a stratified development cohort of 433 participants, or 80% of the sample, and reserved the remaining 109 participants, or 20%, as a held-out cohort. The task was to classify existing cognitive-impairment status rather than predict future decline or evaluate a treatment, and the reported findings concern discrimination in this ADNI sample rather than diagnostic performance in a broader clinical population.

To test whether synthetic augmentation could support clinical-genomic modeling in a relatively small cohort, the team applied TabDDPM, a diffusion-based approach to generating synthetic tabular records. The generated data had a similar demographic structure and common-variant genotype distribution to the real cohort. This distributional similarity was an observed property of the generated samples, not evidence that the synthetic participants were equivalent to real people or could substitute for them in clinical research.

Eight predictive models were evaluated through repeated stratified fivefold cross-validation with 10 repeats. Within each outer training partition, the 346 or 347 participants were divided again: 80% was used to train the generator, 10% for hyperparameter selection, and 10% for internal generator evaluation. Run-specific synthetic samples were then added to the outer training partition at augmentation ratios of three, five, seven and 10 times the real training-set size.

This nested design allowed model selection and synthetic-data generation to remain within the available training data before the selected configuration was assessed in the held-out cohort. Among the nonlinear models, the categorical late-fusion neural network produced the highest average cross-validated AUC under sevenfold augmentation. For the direct comparison between sevenfold and no augmentation, internal cross-validation yielded a raw paired DeLong p-value of 0.0012 and a Holm-adjusted p-value of 0.03.

The statistical comparison therefore supported an improvement in cross-validated discrimination for the selected model and augmentation setting, although the study evaluated a finite collection of models and augmentation ratios rather than every possible configuration. The same sevenfold strategy also improved performance in the 109-participant held-out cohort. AUC rose from 0.721 without synthetic augmentation to 0.769 with augmentation.

Two components of classification performance increased as well: sensitivity moved from 0.68 to 0.77, while specificity moved from 0.58 to 0.66. These figures describe how well the models separated the cognitively impaired and cognitively normal groups in the held-out sample; they do not demonstrate that the system can diagnose Alzheimer’s disease, predict individual disease progression or improve clinical care. The results show that adding generated examples to a small, tabular clinical-genomic dataset can improve the performance of some neural-network classifiers in this study.

Synthetic augmentation did not replace the 542 real ADNI participants or their records. Instead, the observed gains suggest that the generated samples helped the selected late-fusion network learn more stable patterns from demographic and APOE-region genotype features under the study’s validation design. The improvement should not be interpreted as evidence that artificial records are generally superior to real patient data or that the same gain will occur in every cohort.

Several boundaries define the scope of the finding. The analysis was limited to the available ADNI participants, the selected demographic and APOE-region genotype features, and the eight models evaluated. It used cross-sectional cognitive-impairment labels and did not test prospective prediction, treatment response, causality or clinical deployment.

The authors further note that the framework uses standard tabular genomic and demographic variables, which may make it applicable to other genomic classification tasks with limited individual-level data, but such wider use was not demonstrated in this study. The supplied article also states that the final formatted version was still forthcoming, so this report is based on the published abstract and accompanying study information rather than a complete examination of the formatted article.

Subject of Research: Evaluation of synthetic tabular data augmentation for classifying cognitive impairment using APOE-region genotype and demographic data.

Article Title: Improving cognitive impairment classification in Alzheimer’s disease using tabular diffusion augmentation and late fusion deep learning

Article References: Wang, A. Z., & Shen, X. (2026). Improving Cognitive Impairment Classification in Alzheimer's Disease Using Tabular Diffusion Augmentation and Late Fusion Deep Learning. Frontiers in Bioinformatics, 6, 1949507. https://doi.org/10.3389/fbinf.2026.1949507

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1949507

Keywords: Alzheimer’s disease, cognitive impairment, synthetic data, TabDDPM, deep learning, APOE, clinical genomics, neural networks, ADNI, cross-validation, Improving, cognitive