Nearly 6 million Americans ≥65 years suffer from Alzheimer’s disease (AD) or AD-related dementias (ADRD). AD/ADRD poses significant emotional, physical, and financial burdens on patients, families, and societies. There is no cure for AD/ADRD, and apart from the June 2021 controversial “accelerated approval” of aducanumab, no new symptom-modifying drug has been approved since 2003, highlighting the need for AD/ADRD prevention. Currently, no drug is available to delay the onset of AD/ADRD. The prohibitive cost of developing new drugs or repositioning partially developed drugs for AD/ADRD treatment would be even more prohibitive for AD/ADRD prevention as the latter would require larger sample size and longer follow-up. An alternative cost-effective and efficient approach is to repurpose from >20,000 FDA-approved drugs for AD/ADRD prevention. However, repurposing of drugs is often accidental. A timely and purposeful discovery of new clinical benefits of old drugs requires a systematic examination of large comprehensive clinical databases with longitudinal records and long follow-up, using innovative, sophisticated mixed machine learning and statistical tools. This application has been prepared in response to the NIA PAR-20-156 entitled “Translational Bioinformatics Approaches to Advance Drug Repositioning and Combination Therapy Development for Alzheimer’s Disease”. We propose a 3-Step Medication-Wide Association Study Plus (MWAS+) approach. Our MWAS+ will employ innovative explainable deep (machine) learning, a powerful artificial intelligence tool for noisy, nonlinear data. We will use Veterans Affairs (VA) electronic health record (EHR) data of >3 million Veterans ≥65 years (54,411 women; 202,000 African American), ~600 prescription drugs (each used by ≥10,000 Veterans), ≥10 years of history and ~200,000 AD/ADRD cases. In Step 1 (Aim 1), we will conduct a hypothesis-free exploratory case-control MWAS (akin to GWAS) to identify drugs associated with AD/ADRD in the VA EHR data. Drugs identified in Aim 1 will be reviewed by a panel of experts for plausible mechanistic pathways and 10 drugs will be recommended for hypothesis testing in Step 2 using VA EHR data (Aim 2) and external validation in Step 3 using Medicare data (Aim 3). In Aims 2 and 3, we will conduct outcome-blinded cohort studies using new user design. Marginal structural models and other causal inference methods, including doubly-robust inference procedures, will be used to estimate time- fixed (“intent-to-treat”) and time-varying (“as-treated”) effects of those drugs on incident AD/ADRD. The proposed project is highly significant because it will rigorously accelerate the identification of already approved drugs that have a high potential to be repurposed to delay and prevent AD/ADRD, a rapidly growing public health crisis. The project is innovative as it combines state-of-the-art deep learning and statistical methods to conduct an MWAS+ study that has never been used before for AD/ADRD ...