[Validation of algorithms for the identification of subjects with chronic disease using health information systems]

to test the validity of algorithms to identify diabetes, chronic obstructive pulmonary disease (COPD), hypertension, and hypothyroidism from routinely collected health data using information from self-reported diagnosis and laboratory or functional test.

clinical or self-reported diagnosis from three surveys conducted in Lazio Region (Central Italy) between year 2010 and 2014 were assumed as gold standard and compared to the results of the algorithms application to administrative data.

prevalence resulted from administrative data and from information available in the surveys were compared. Sensitivity, specificity, positive predictive value, and positive likelihood ratio of algorithms with respect to self-reported diagnosis, laboratory or functional test, assumed as gold standards, were calculated.

we analyzed data of 7,318 subjects (1,545 for diabetes, 1,783 for COPD, 2,448 for hypertension, and 1,542 for hypothyroidism). For hypertension and hypothyroidism, we observed a higher prevalence from laboratory or functional test compared to self-reported diagnosis (54.5% vs. 44.9% and 7.5% vs. 1.5%). Sensitivity of administrative data with respect to self-reported diagnosis resulted 90.9%, 38.5%, 88.3%, and 47.8%, respectively, for diabetes, COPD, hypertension, and hypothyroidism. Respectively, specificity was 97.4%, 91.7%, 84.8% and 91.8%; positive predictive value was 70,9%, 38.1%, 82.6% and 8.1%. All values of positive likelihood ratio resulted moderate (about 5), with exception of the diabetes algorithm and the disease-specific payment exemptions register for hypertension (respectively 35.5 and 17.4).

hypertension and hypothyroidism resulted markedly underdiagnosed from self-reported data. Case identification algorithms are highly specific, allowing their utilization for selection of cohort of subject affected by chronic diseases. The sub-optimal sensitivity observed for COPD and hypothyroidism could limit the utilization of the algorithms for prevalence estimation.