Optimizing algorithms to identify Parkinson's disease cases within an administrative database.

Patients assigned the diagnostic ICD-9-CM code for Parkinson's disease (PD) in an administrative database may not truly carry that diagnosis because of the various error sources. Improved ability to identify PD cases within databases may facilitate specific research goals. Experienced chart reviewers abstracted the working diagnosis of all 577 patients assigned diagnostic code 332.0 (PD) during 1 year at a VA Healthcare System. We then tested the ability of various algorithms making use of PD and non-PD diagnostic codes, specialty of clinics visited, and medication prescription data to predict the abstracted working diagnosis. Chart review determined 436 (75.6%) patients to be PD or Possibly PD, and 141 (24.4%) to be Not PD. Our tiered consensus algorithm preferentially used data from specialists over nonspecialists improved PPV to 83.2% (P = 0.003 vs. baseline). When presence of a PD prescription was an additional criterion, PPV increased further to 88.2% (P = 0.04 vs. without medication criterion), but sensitivity decreased from 87.4 to 77.1% (P = 0.0001). We demonstrate that algorithms provide better identification of PD cases than using a single occurrence of the diagnostic code for PD, and modifications of such algorithms can be tuned to maximize parameters that best meet the goals of a particular database query.