Validation of knee osteoarthritis case identification algorithms in a large electronic health record database

PURPOSE: To facilitate studies of knee osteoarthritis (OA) in large databases, case finding algorithms with high levels of diagnostic performance are needed. METHODS: From a UK general practitioner (GP) practice derived database, we selected adults ages 40–90 years meeting algorithms that included various combinations of codes for knee OA or knee pain and imaging. The GP for each patient was mailed a questionnaire to assess the cause of knee pain and provide knee x-ray and/or MRI findings. We considered knee pain with x-ray and/or MRI findings consistent with OA the gold standard. We calculated positive predictive values (PPV) and sensitivity for case identification algorithms. RESULTS: Of 100 questionnaires sent, 93 were returned; we excluded 8 subjects who had other rheumatic disorders or total knee replacements. Among those with one code for OA, the PPV was 64% (95% CI ​= ​49%–79%) and it increased to 92% (95% CI ​= ​76%–100%) when two or more OA codes over six months were required. The increase in PPV was accompanied by a drop in sensitivity from 44% (95% CI ​= ​31%–57%) to 19% (95% CI ​= ​9%–30%). Use of one pain code yielded similar results to use of one OA code. Requiring two or more knee pain codes over six months yielded a PPV of 68% (95% CI ​= ​49%–88%) and sensitivity of 26% (95% CI ​= ​15%–38%). DISCUSSION: A case identification algorithm requiring two or more knee OA codes yielded the highest PPV at the cost of reduced sensitivity. Tradeoffs between PPV and sensitivity will need to be weighed alongside study goals when selecting a case identification algorithm.