PNS162 USING AMBULATORY EMR DATA TO COMPLEMENT MEDICAL CLAIMS DATA FOR PREDICTIVE ANALYTICS APPLICATIONS

Medical claims data has widely been used in studies to derive real world evidence (RWE) about patient medical history. However, RWE based entirely on the medical claims may be insufficient in detecting patients with rare diseases or chronic conditions as it is limited to pre-determined codes to describe diagnoses, prescriptions and procedures. We propose an approach to reduce this limitation by consolidating medical claims with Ambulatory Electronic Medical Records (AEMR) data which comprises rich information from outpatient settings including unstructured medical notes and lab test results. When linked, these records can complement medical claims data in cases where there 1) isn’t a claim code available to describe a given medical condition, 2) where claim codes are inconsistent across ICD-9 and ICD-10 standards, or 3) where a medical condition is best identified through one or more specific lab results. We leverage AEMR data to identify patients with a chronic condition which doesn’t have a corresponding ICD-10 claim code, and is also difficult to distinguish through business rules alone. Our positive cohort, comprised of patients with the chronic condition, was generated using the problem description field and appropriate SNOMED-CT (Systematized Nomenclature of Medicine Clinical Terms) codes from AEMR data. Model features were generated using medical history in claims data for positive and negative cohorts and a Gradient Boosted Tree classifier (XGBOOST) was built to predict this chronic condition. Study sample included 825,473 patients, 75,043 of which were positives. Results revealed that the model had 43% precision at 25% recall level and our model outperformed the rule-based prediction of chronic condition patients which only had a 20% precision at the same recall level. We provide an approach for incorporating AEMR with claims data to address challenges introduced by the lack of comprehensive information in claims data for predictive analytics studies.