Integrated statistical and machine learning analysis provides insight into key influencing symptoms for distinguishing early-onset type 2 diabetes

Abstract

Background: Being able to predict with confidence the early onset of type 2 diabetes from a suite of signs and symptoms (features) displayed by potential sufferers is desirable to commence treatment promptly. Late or inconclusive diagnosis can result in more serious health consequences for sufferers and higher costs for health care services in the long run. Methods: A novel integrated methodology is proposed involving correlation, statistical analysis, machine learning, multi-K-fold cross-validation, and confusion matrices to provide a reliable classification of diabetes-positive and -negative individuals from a substantial suite of features. The method also identifies the relative influence of each feature on the diabetes diagnosis and highlights the most important ones. Ten statistical and machine learning methods are utilized to conduct the analysis. Results: A published data set involving 520 individuals (Sylthet Diabetes Hospital, Bangladesh) is modeled revealing that a support vector classifier generates the most accurate early-onset type 2 diabetes status predictions with just 11 misclassifications (2.1% error). Polydipsia and polyuria are among the most influential features, whereas obesity and age are assigned low weights by the prediction models. Conclusion: The proposed methodology can rapidly predict early-onset type 2 diabetes with high confidence while providing valuable insight into the key influential features involved in such predictions.