Machine Learning Embedded Smartphone Application for Early-Stage Diabetes Risk Assessment

Abstract

Diabetes Mellitus (DM) is a metabolic disease that hampers the function of glucose in the human body. DM screening remains challenging due to its initial asymptotic be-havior, invasive nature of risk factor measurement, and complex physiological relation with other diseases. An efficient early-stage diagnosis tool that provides a diabetes risk assessment via non-invasive means is imperative to mitigate the lingering risk as well as to take preventive measures. In this paper, a highly accurate diabetes risk assessment pipeline using medical health records is presented. The dataset obtained from the University of California Irvine (UCI) Machine Learning Repository consists of 520 samples of both male and female subjects, each having 16 attributes related to signs and symptoms of diabetes. Different su-pervised machine learning (ML) models have been implemented to predict the probabilities of the diabetic and non-diabetic classes for a given record. A comparative analysis of different ML methods confirms that the Random Forest is best suited for this problem with an accuracy of 95%. The best performing pre-trained machine learning predictive model is embedded into a smartphone application to generate a recommendation on the risk of DM based on simple patient-related questions and answers. Therefore, the outcome of this research will help improve early diagnosis, monitoring, and clinical screening of DM.