P4‐554: An Ehr‐based Cohort Discovery Tool for Identifying Probable Ad

Electronic health records (EHRs) contain decades of longitudinal clinical data on hundreds of thousands of potentially at-risk individuals for Alzheimer's disease (AD). The ability to automatically identify probable AD patients within EHRs would facilitate downstream computational analyses on such large-scale datasets, by eliminating the need for labor intensive chart review. To this end, we developed and validated a cohort discovery tool that can be applied to EHR data for automatic classification of individuals with AD. We extracted EHR data from Michigan Medicine's Research Data Warehouse (RDW) pertaining to Michigan Alzheimer's Disease Center (MADC) participants with a consensus-based diagnosis ranging from cognitively normal to probable AD. We investigated the accuracy of different EHR-based rules for identifying patients with AD. Rules were based on combinations of criteria pertaining to ICD diagnoses, medications, laboratory results and encounter types. Applied to data from the RDW, these rules were evaluated against MADC diagnoses (Figure 1), in terms of sensitivity, specificity, and positive predicted value (PPV). To optimize for the probability that patients identified by the rule have AD, we prioritized PPV when ranking different rules. MADC and RDW records overlapped in 624 patients 65 years and older. Though a diagnostic code for AD alone resulted in relatively low specificity, combination with an encounter involving medium/high complexity medical decision making resulted in increased specificity and the highest PPV (Figure 2). This rule yielded a PPV, specificity, and sensitivity of 0.82 (95% confidence interval (CI) 0.75-0.87), 0.95 (95%CI 0.93-0.97), and 0.65 (95%CI 0.60-0.68) respectively. For true positives, the first RDW diagnosis of probable AD occurred on average three years before the first MADC diagnosis (95% CI 2.3−3.7 years). Applied to the entire RDW, this rule identified 4,152 patients with probable AD.