Explainable machine learning aggregates polygenic risk scores and electronic health records for Alzheimer’s disease prediction

In the present study, we constructed PRSs for AD risk and AAO, built ML models for predicting the risk of developing AD, and explored feature importance among PRSs, conventional risk factors, and ICD-10 codes from EHRs. Our results showed that PRSs from risk and AAO tests both substantially improved the discriminatory ability for AD, especially for the age 65 + group, where adding PRSs increased AUC by 16% over the model with only age and sex. Interestingly, PRSs ranked on the top, even higher than age, in feature importance for the age 65 + group. To improve interpretability of the ML technique, we computed SHAP values for feature ranking and visualization. To our knowledge, this is the first report to develop predictive models for AD using genetic, non-genetic information, and ICD-10 codes from EHR in a large-scale cohort study using a modern explainable ML framework.

Our ML…

Read more…