POSC301 Optimization of a Pre-Defined Algorithm Using Cluster Analysis - Application to a Cohort of Diabetic Patients from the French National Health Data System

Abstract

Due to lack of clinical data in the French National Health Data System (SNDS), proxies of clinical variables must be defined. To date there is no consensus on algorithms to be used to identify diabetes types within SNDS. The objective was to optimize a pre-defined-algorithm with innovative methods to identify types of treated diabetes (type 1 diabetes (T1DM), type 2 diabetes (T2DM), gestational diabetes (GDM)) using claims data. Diabetic patients were identified in the Permanent Representative Sample (Echantillon Généraliste des Bénéficiaires (EGB)) from SNDS. Patients with at least two dispensations of an antidiabetic treatment during 2017 were included and analysed on 2012-2017 period. A pre-defined algorithm has been developed to characterize diabetes type based on hospitalizations, treatments and long duration disease and has been applied to study population. Its validity was assessed by: i) comparison of results with distribution of diabetes types in France, ii) cluster analysis on antidiabetic treatment sequences (insulin, oral glucose-lowering drug (OGLD), glucagon-like peptide-1 (GLP-1) analogues). The population consisted of 29,288 diabetic patients. Application of the pre-defined algorithm resulted in categorization of 4,169 (14.2%) T1DM, 25,037 (85.5%) T2DM and 82 (0.3%) GDM, and led to an overrepresentation of T1DM patients. Clustering of treatment sequences on T1DM patients resulted in three clusters: 2,493 patients with insulin alone, 910 with OGLD alone, 766 with insulin and OGLD. The cluster analysis highlighted that the pre-defined algorithm grouped under the category T1DM patients on OGLD or GLP-1 who should have been categorized as T2DM. Based on the results, the pre-defined algorithm has been refined and led to identification of diabetes types consistent with the literature (1,964 (6.7%) T1DM, 27,243 (93.0%) T2DM, 81 (0.3%) GDM). When using claims databases where clinical data are limited, clustering analysis on treatment sequences is a relevant method to refine pre-defined algorithms.