Comparison and validation of data-mining indices for signal detection: using the Korean national health insurance claims database
Version of Record online: 5 OCT 2011
Copyright © 2011 John Wiley & Sons, Ltd.
Pharmacoepidemiology and Drug Safety
Volume 20, Issue 12, pages 1278–1286, December 2011
How to Cite
Choi, N.-K., Chang, Y., Kim, J.-Y., Choi, Y.-K. and Park, B.-J. (2011), Comparison and validation of data-mining indices for signal detection: using the Korean national health insurance claims database. Pharmacoepidem. Drug Safe., 20: 1278–1286. doi: 10.1002/pds.2237
- Issue online: 22 NOV 2011
- Version of Record online: 5 OCT 2011
- Manuscript Accepted: 19 JUL 2011
- Manuscript Revised: 29 JUN 2011
- Manuscript Received: 3 MAY 2010
- health insurance claims database;
- relative risk;
To detect the signals of celecoxib compared with other analgesics and anti-inflammatory drugs (AAIDs) by proportional claims ratio (PCR), claims odds ratio (COR), information component (IC), and relative risk (RR) using the Korean claims database. In addition, the concordance of the identified signals by the data-mining indices (DMIs) and the validity of the DMIs were evaluated.
The Korean Health Insurance Review and Assessment Service claims database was used. The study population consisted of elderly ambulatory care patients with osteoarthritis who were prescribed AAIDs in Seoul from 1 January 2005 to 30 September 2005. A short-term serious adverse event (SAE) was defined as a hospital admission within 12 weeks from each AAID prescription. Among the screened SAEs, signals were identified by the DMIs. The sensitivity, specificity, and predictability were estimated with reference to known adverse events associated with celecoxib.
A total of 135 232 elderly patients with osteoarthritis were prescribed AAIDs. There were 309 717 drug–SAE pairs and 481 different SAEs. The PCR, COR, IC, and RR detected were as follows: 56 (11.6%), 57 (11.9%), 129 (26.8%), and 123 (25.6%) signals for celecoxib, respectively. The RR detected signals had a relatively high sensitivity (23.4%) compared with the other indices (PCR 9.9%, COR 10.8%, and IC 18.9%). The specificity of RR (73.8%) was higher than that of IC (70.8%). The positive and negative predictive values of the RR were 21.1% and 76.3%, respectively.
This study suggested that the RR was the most accurate of the DMIs for detecting signals in the claims database. Copyright © 2011 John Wiley & Sons, Ltd.