Predicting Alzheimer’s Trajectory: A Multi-PRS Machine Learning Approach for Early Diagnosis and Progression Forecasting

AbstractINTRODUCTIONPredicting the early onset of dementia due to Alzheimer’s Disease (AD) has major implications for timely clinical management and outcomes. Current diagnostic methods, reliant on invasive and costly procedures, underscore the need for scalable and innovative approaches. To date, considerable effort has been dedicated to developing machine learning (ML) based approaches using different combinations of medical, demographic, cognitive, and clinical data, achieving varying levels of accuracy. However, they often lack the scalability required for large-scale screening and fail to identify underlying risk factors for AD progression. Polygenic risk scores (PRS) have shown promise in predicting disease risk from genetic data. Here, we aim to leverage ML techniques to develop a multi-PRS model that captures both genetic and non-genetic risk factors to diagnose and predict the progression of AD in different stages in older adults.METHODSWe trained and tested ML-based multi-PRS models, integrating genetically predicted clinical, behavioral, psychiatric, and lifestyle risk factors to predict the diagnosis of AD as well as the progression between different cognitive stages. We developed an automatic feature selection pipeline that identifies the relevant traits that predict AD. We also analyzed the interpretability of our pro-posed ML models and the selected features. Leveraging data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), Religious Orders Study and Memory and Aging Project (ROSMAP), and the IEU OpenGWAS Project, our study presents the first known end-to-end ML-based multi-PRS model for AD.RESULTSRelevant features were selected from an initial set of 53 polygenic risk scores computed for 1567 patients in the ADNI and 1642 patients in the ROSMAP dataset. The proposed multi-PRS ML method produced AUROC scores of 77% on ADNI and 72% on ROSMAP for predicting the diagnosis of AD, substantially surpassing the performance of the uni-variate PRS models. Our models also showed promise in predicting transitions between various cognitive stages (65%-75% AUROC scores). Moreover, the features identified by our automated feature selection pipeline are closely aligned with the widely recognized potentially modifiable risk factors for AD.DISCUSSIONMulti-PRS-based machine learning models can identify risk factors and construct predictive models for early Alzheimer’s disease (AD) diagnosis. This approach offers an automated mechanism to harness genetic data for AD diagnosis and prognosis, enhancing our understanding of the role of various traits in AD development and progression. It will facilitate the implementation of preventive measures at an early stage, thereby contributing to more effective interventions and improved patient outcomes.