Hybrid Prediction Model For Type-2 Diabetes With Class Imbalance

Diabetes mellitus is a fast-growing disease affecting millions of people around the globe. According to the statistics obtained from the International Diabetes Federation, the growth of diabetic cases around the world is estimated to be around 10.2% (578 million) by 2030 and 10.9% (700 million) by 2045. Machine learning techniques can be used for early diagnosis of diabetes so that individuals can adapt to better lifestyle habits and diet plan to prevent further complications of symptoms. In this paper we have implemented an efficient model to classify type-2 diabetes using a hybrid approach. The dataset which is relatively imbalanced with respect to the number of diabetic or non-diabetic individuals was obtained from the Biostatistics program of Vanderbilt University, USA which contains parameters like Glucose, Age, Cholesterol, Weight, Waist/Hip ratio of 390 African-American individuals. The features from the dataset were obtained by filter and wrapper methods. The dataset was balanced by applying oversampling and undersampling methods. A voting ensemble model consisting of 5 algorithms was then used as a classifier with model validation done using stratified K-fold technique. The model's performance was assessed using recall, precision, F1 score and accuracy. The results obtained by comparison of the models with and without sampling have shown significant improvements in recall. The main motive is to improve recall and have reasonable precision at the same time.