Predicting hospitalisation of patients with diabetes mellitus. An application of the Bayesian discriminant analysis.

The objective of this study was to develop, and subsequently test, a Bayesian discrimination model for the purpose of identifying both the personal and the healthcare system characteristics predictive of hospitalisation for the treatment of patients with diabetes mellitus or commonly observed cormorbidities associated with the disease. First, a Bayesian classification framework was proposed. The model was then tested by using a logit regression technique in order to estimate the probability of one or more hospitalisation events among patients with diabetes. The study used claims data extracted from the Hawaii Medical Service Association (HMSA) Private Business Claims (PBS) files for the 1995 calendar year. Patients under 65 years were identified by paid claims with ICD-9-CM diagnosis codes of 250.xx which gave a sample size of 6841 patients. Age, gender, various pharmacotherapy variables, presence of hypertension, hyperlipidaemia, congestive heart failure, multiple cardiovascular diseases, any combination of commonly observed comorbidities, dialysis services and annual eye examination are highly predictive of 1 or more hospitalisation events. The model shows a predictive power of almost 90%. This study found that multivariate discriminant analysis using a logit regression model successfully identifies: (i) important explanatory variables predictive of hospitalisation; (ii) assigns patients into 1 of 2 mutually exclusive classes; and (iii) offers a benchmark for a comprehensive disease management strategy for patients with more complicated diabetes.