Validation of classification algorithms for childhood diabetes identified from administrative data.

ABSTRACT

OBJECTIVE: Type 1 diabetes is the most common form of diabetes among children; however, the proportion of cases of childhood type 2 diabetes is increasing. In Canada, the National Diabetes Surveillance System (NDSS) uses administrative health data to describe trends in the epidemiology of diabetes, but does not specify diabetes type. The objective of this study was to validate algorithms to classify diabetes type in children <20 yr identified using the NDSS methodology. PATIENTS AND METHODS: We applied the NDSS case definition to children living in British Columbia between 1 April 1996 and 31 March 2007. Through an iterative process, four potential classification algorithms were developed based on demographic characteristics and drug-utilization patterns. Each algorithm was then validated against a gold standard clinical database. RESULTS: Algorithms based primarily on an age rule (i.e., age <10 at diagnosis categorized type 1 diabetes) were most sensitive in the identification of type 1 diabetes; algorithms with restrictions on drug utilization (i.e., no prescriptions for insulin ± glucose monitoring strips categorized type 2 diabetes) were most sensitive for identifying type 2 diabetes. One algorithm was identified as having the optimal balance of sensitivity (Sn) and specificity (Sp) for the identification of both type 1 (Sn 98.6%; Sp 78.2%; PPV 97.8%) and type 2 diabetes (Sn 83.2%; Sp 97.5%; PPV 73.7%). CONCLUSIONS: Demographic characteristics in combination with drug-utilization patterns can be used to differentiate diabetes type among cases of pediatric diabetes identified within administrative health databases. Validation of similar algorithms in other regions is warranted.