A Machine Learning Approach to Predicting Diabetes Complications

Abstract

Machine learning and data mining techniques have been widely used over the years to extract knowledge from data. The goal of this thesis is to study several diabetes complications. Diabetes Mellitus (DM) is a chronic disease that is considered to be life threatening. It can affect any part of the body over time resulting in more serious complications such as impacts on eyesight, perception, motor control and more. To study diabetes complications, a dataset collected by the Rashid Centre for Diabetes and Research (RCDR) located in Ajman, UAE was utilized. The dataset consists of 884 records with 79 features and 8 complications. The complications’ set contains metabolic syndrome, dyslipidemia, neuropathy, nephropathy, diabetic foot, hypertension, obesity, and retinopathy. Some essential preprocessing steps were needed to handle the missing values and imbalanced data problems. Moreover, several techniques were used to study the problem in hand. The first part of this thesis focused on generating association rules from the dataset using unsupervised learning techniques. This step was essential to extract valuable knowledge and relations between several attributes in the dataset and helped to develop a better understanding of DM and its complications. For instance, we extracted several rules indicating some possible relations between metabolic syndrome, hypertension and dyslipidemia. Further preprocessing steps were needed such as data discretization. For the second part of the research, different supervised classification algorithms were utilized to build several models to predict and diagnose eight diabetes complications. Furthermore, feature selection was performed to select the top 5 and 10 features for each complication. Repeated stratified k-fold cross validation was employed for a better estimation of the performance with a k=10 and a total of 10 repetitions. Accuracy and F1-score were used to evaluate the models’ performance reaching a maximum of 97.8% and 97.7% for accuracy and F1-scores, respectively.