Improving automated case finding for ectopic pregnancy using a classification algorithm.

BACKGROUND: Research and surveillance work addressing ectopic pregnancy often rely on diagnosis and procedure codes available from automated data sources. However, the use of these codes may result in misclassification of cases. Our aims were to evaluate the accuracy of standard ectopic pregnancy codes; and, through the use of additional automated data, to develop and validate a classification algorithm that could potentially improve the accuracy of ectopic pregnancy case identification.
METHODS: Using automated databases from two US managed-care plans, Group Health Cooperative (GH) and Kaiser Permanente Colorado (KPCO), we sampled women aged 15-44 with an ectopic pregnancy diagnosis or procedure code from 2001 to 2007 and verified their true case status through medical record review. We calculated positive predictive values (PPV) for code-selected cases compared with true cases at both sites. Using additional variables from the automated databases and classification and regression tree (CART) analysis, we developed a case-finding algorithm at GH (n = 280), which was validated at KPCO (n = 500).
RESULTS: Compared with true cases, the PPV of code-selected cases was 68 and 81% at GH and KPCO, respectively. The case-finding algorithm identified three predictors: ≥ 2 visits with an ectopic pregnancy code within 180 days; International Classification of Diseases, 9th Revision, Clinical Modification codes for tubal pregnancy; and methotrexate treatment. Relative to true cases, performance measures for the development and validation sets, respectively, were: 93 and 95% sensitivity; 81 and 81% specificity; 91 and 96% PPV; 84 and 79% negative predictive value. Misclassification proportions were 32% in the development set and 19% in the validation set when using standard codes; they were 11 and 8%, respectively, when using the algorithm.
CONCLUSIONS: The ectopic pregnancy algorithm improved case-finding accuracy over use of standard codes alone and generalized well to a second site. When using administrative data to select potential ectopic pregnancy cases, additional widely available automated health plan data offer the potential to improve case identification.