Abstract
Background/Aims: Research and surveillance work addressing ectopic pregnancy (EP) rely largely on ICD diagnostic codes and CPT procedure codes available from automated data sources. However, cases identified in this way may not be true EP cases, as these codes may also be used to rule out, follow up or note a history of EP. Through the use of additional automated data on treatment, procedures, and other aspects of care, we developed a classification algorithm that could improve the accuracy of EP case identification.
Methods: Using Group Health automated data files, we initially identified 2,632 potential EP episodes occurring in women aged 15–44 years during 1988–2007 using ICD and CPT codes for EP and for surgical or laparoscopic treatment of EP. Chart reviews to verify EP status were conducted on 130 potential EP cases in the algorithm development dataset and on 150 in the algorithm validation dataset. Using additional information on demographic factors, other diagnosis and procedure codes, treatment modalities, site of care (ambulatory vs. inpatient), and laboratory data available with each EP episode, we conducted a classification and regression tree (CART) analysis to create a case finding algorithm for EP.
Results: From the CART analysis, the case-finding algorithm for EP contained three main predictors: at least two encounter dates with an EP diagnosis or procedure code during an episode; treatment with methotrexate; and presence of an ICD -9 code of 633.1, 633.10 or 633.11 for tubal pregnancy. Sensitivity for the development and validation sets, respectively, was 95% and 91% while specificity was 78% and 83%. The EP misclassification rate using the new algorithm was 10.7% in the development set and 11.5% in the validation set compared to 32% when EP cases were originally defined using EP diagnosis and procedure codes.
Conclusions: The CART-derived algorithm for identifying EP cases was highly sensitive, had good specificity, and misclassification rates were notably improved over the original case identification techniques. Additional pharmacy and encounter data available in many health plan databases can markedly improve the accuracy of EP identification. In particular, the identified predictors in the algorithm are available in the CRN VDW, and it would be of interest to test this algorithm at other HMORN sites.
