Definition and validation of a predictive model to identify patients with chronic obstructive pulmonary disease (COPD) from administrative databases

OBJECTIVE: To develop and validate a predictive model for the identification of patients with Chronic Obstructive Pulmonary Disease (COPD) among the resident population of the Lazio region, using information available in the regional administrative systems (SIS) as well as clinical data of a panel of COPD patients. SETTING AND PARTICIPANTS: All residents in the Lazio region over 40 years of age in 2007 (2,625,102 inhabitants) MAIN OUTCOME MEASURES: The predictive model was developed through record linkage of health care related consumption patterns among 428 panel patients with confirmed COPD diagnosis in 2006 and a control group of patients without COPD (selection from outpatients specialized health care registry, 1:4). Hospital admission for COPD was defined a priori to be sufficient to identify a COPD patient. For all other panel patients and controls, specific drug use (minimum 2 prescriptions during 12 months) and hospitalization for respiratory causes during the past 9 years were retrieved and compared between panel and control patients. COPD associated factors were selected through a Bootstrap- Stepwise (BS) procedure. The predictive model was validated through internal (cross-validation-bootstrap) and external validation (comparison with external COPD patients with confirmed diagnosis), and through comparison with other COPD identification approaches. RESULTS: The BS procedure identified the following predictors of COPD: consumption of beta 2 agonists, anticholinergics, corticosteroids, oxygen, and previous hospitalization for respiratory failure. For each patient, the expected probability of being affected by COPD was estimated. Depending on the cut-point of expected probability, sensibility ranged from 74.5% to 99.6% and specificity from 37.8% to 86.2%. Using the 0.30 cut-point, the model succeeded in identifying 67% of patients with diagnosis of COPD confirmed with spirometry. The predictive performance increased with increasing COPD severity. Prevalence of COPD turned out to be 7.8 %. The age-specific estimation was similar to results from other approaches. CONCLUSION: The predictive model shows good performance to identify COPD patients, even if it does not allow to identify those patients who have not been registered in the regional health care service or do not request any public health care service.