Improving type 2 diabetes mellitus risk prediction using classification

Diabetes is a chronic disease that contributes to a significant portion of the healthcare expenditure for a nation as individuals with diabetes need continuous medical care. In order to prevent or delay the onset of type 2 diabetes, it is necessary to identify high risk populations and introduce behavior modifications as early as possible. Screening the population to identify high risk individuals is an important task. One of the most accurate tests of diabetes is through the analysis of fasting blood sugar, but it is invasive and costly. Furthermore, it is only useful when the individual is already displaying symptoms i.e., making a diagnosis, which is considered too late to be an effective screening mechanism. Therefore, a reliable non-invasive inexpensive test to predict high risk individuals in advance is needed. In this paper, we use classification to mine an extensive dataset gathered from 12 hospitals in Thailand during 2011-2012 with 22,094 records of screened population who are females age 15 years or older. We use RapidMiner Studio 7.0 with Naive Bayes and CHAID (Chi-squared Automatic Interaction Detector) Decision Tree classifiers to predict high risk individuals and compared our results to existing hand-computed diabetes risk scoring mechanisms. We define the goal of risk prediction as coverage which is the ability to use screening data to identify individuals that are eventually diagnosed with diabetes. Our results indicate that the use of classification introduced in this paper instead of hand-computed scoring can improve the prediction performance with an increase in coverage.