Exploring data-driven algorithms to identify Lyme disease cases in electronic health records

A tool capable of quickly and accurately identifying Lyme disease (LD) cases in endemic areas would have powerful applications in both research and public health. Using data from a single health system in an area with high LD incidence, we evaluated the feasibility and accuracy of data-driven algorithms to accurately identify LD cases based on easily extracted, structured data elements that are available in electronic health records (EHR). We used random forest and decision tree algorithms to explore the importance of various data elements and to investigate how different combinations of data elements accurately classified clinical events as either LD cases or non-cases. The data elements explored included patient demographics, International Classification of Diseases, 10th Revision, (ICD) codes, LD laboratory tests and their results, and antibiotic prescription orders. Our best performing algorithm was a manually constructed decision tree that incorporated information from three presence/absence variables: an LD ICD code (A69.2x), a positive LD serology test, and a doxycycline order occurring within 30 days of either of the other two elements, which achieved 82% accuracy, 93% sensitivity, 60% adjusted positive predictive value, 68% specificity, and 91% adjusted negative predictive value. Findings from this exploratory effort suggest that data-driven algorithms based on EHR elements hold promise to identify LD cases when detailed chart reviews are impractical. Further efforts may differentiate algorithms between confirmed and probable LD cases, pediatric and adult populations, and utilize text searching to improve accuracy.