An affective learning-based system for diagnosis and personalized management of diabetes mellitus

Abstract

Diabetes Mellitus is a major health problem with high global morbidity and mortality rates. While, conventional diagnosis methods are based on monitoring blood glucose levels, variables such as body mass index and diastolic blood pressure have been reported as having stable correlations with the incidence and prevalence of diabetes. Recently, machine learning approaches are developed for diagnosis of diabetes, but the existing models are mostly trained and validated on single dataset while the apt strategies needed for proper management of diabetes mellitus are omitted. In this study, we develop an affective learning-based system for diagnosis and therapy of diabetes mellitus. The integrated system consists of a multimodal adaptive neuro-fuzzy inference model designed for diagnosis of diabetes, and a knowledge-based diets recommender model designed for personalized management of diabetes. The diagnosis model was trained and validated with 87.5% and 13.5% of Pima Indians diabetes dataset, respectively; and re-validated with Schorling diabetes dataset; both of which are publicly available. Then, the model was applied retrospectively on a private dataset consisting of 14 female patients’ records obtained at Obafemi Awolowo University Teaching Hospital Complex, Ile-Ife, Nigeria. Also, the recommender model combines users’ diagnoses results with their eating formulae to determine users’ food-per-day consumption and generate weekly personalized food-plan from an expert-designed template. Evaluation results from the studies show that both models performed well. Specifically, the multimodal model attained training and validation accuracies of 83.8% and 79.2%, respectively for Pima dataset, and prediction accuracies of 72.9% and 94.3% for the Schorling and private dataset cases, respectively. In addition, the model shows the best performance when compared with individual baseline models, and ten existing machine learning methods used in related studies. Similarly, the proposed recommender model received the highest average score when compared with several existing diet recommender systems used for chronic diseases therapy. With these promising features, the proposed affective learning-based system could effectively reduce the morbidity and mortality rates of diabetes mellitus in the world.