Explainable variational autoencoder (E-VAE) model using genome-wide SNPs to predict dementia

Objective: Alzheimer's disease (AD) and AD related dementias (ADRD) are complex multifactorial neurodegenerative diseases. The associations between genetic variants obtained from genome wide association studies (GWAS) are the most widely available and well documented variants associated with ADRD. Application of deep learning methods to analyze large scale GWAS data may be a powerful approach to elucidate the biological mechanisms in ADRD compared to penalized regression models that may lead to over-fitting. Methods: We developed a deep learning frame work explainable variational autoencoder (E-VAE) classifier model using genotype (GWAS SNPs = 5474) data from 2714 study participants in the Health and Retirement Study (HRS) to classify ADRD. We validated the generalizability of this model among 234 participants in the Religious Orders Study and Memory and Aging Project (ROSMAP). Utilizing a linear decoder approach we have extracted the weights associated with latent features for biological interpretation. Results: We obtained a predictive accuracy of 0.71 (95 % CI [0.59, 0.84]) with an AUC of 0.69 in the HRS test dataset and got an accuracy of 0.62 (95 % CI [0.56, 0.68]) with an AUC of 0.63 in the ROSMAP dataset. Conclusion: This is the first study showing the generalizability of a deep learning prediction model for dementia using genetic variants in an independent cohort. The latent features identified using E-VAE can help us understand the biology of AD/ ADRD and better characterize disease status.