Classification of temporal trends in healthcare research databases in support of data quality and hypothesis generation
Wednesday, September 13, 2017 — Poster Session I
- YA Sumathipala
- V Huser
Healthcare systems generate massive amounts of data through electronic health records (EHR) and healthcare claims. Leveraging this data to improve patient safety, outcomes, and further the science of care delivery is widely recognized a tenet and goal of 21st century healthcare. To help with data quality assessment and hypothesis generation we analyzed temporal trends in medical events such as prescription medications, medical procedures, diagnoses, and lab results. We developed methodologies to extract 13 statistical features from time-series data to classify trends. Our methods identify and characterize seasonality by week and by month in data that follows the Observational Medical Outcomes Partnership Common Data Model (CDM) We analyzed 7,761 medical procedures; 1,351 prescription drug ingredients; and 10,680 diagnoses, using claims data. Our results showed the following expected seasonal events: preventive care (peaked in August); influenza vaccines (fall); and pulmonary conditions (winter). Surprising seasonal events included adenoidectomies (spring) and visual field exams (August); fluorouracil, atenolol and dextromethorphan (January and winter); and acute pyelonephritis (August) and concussions (September-October). Analysis of weekly seasonality revealed knee arthroplasties are 17% more common on Mondays (p < 0.001) and, surprisingly, vasectomies were 30% higher on Fridays (p < 0.001). Detection of seasonal trends can be used to assess data quality, complimenting conventional rule-based approaches. The R software package developed in this research can be applied to explore seasonality in any CDM dataset to discover geographic and international differences; we expect Southern Hemisphere nations to exhibit seasonality patterns different from the U.S.
Category: Research Support Services