A Novel Hybrid Machine Learning Model to Predict Diabetes Mellitus

Abstract

Diabetes Mellitus is a chronic condition which is associated with the high level of sugar that increases abnormally in the blood. It is a lifelong disease that can have detrimental effects on the other parts of the human body. The goal of this work is to propose a hybrid machine learning prediction model which can detect Type 2 diabetes more efficiently than previous works. In this work, Pima Indian diabetes dataset was collected from kaggle UCI machine learning repository. Then, we cleaned and detected outliers by considering interquartile ranges from this dataset. After removing outliers, this imbalanced dataset needed to be balanced as well. So, synthetic minority oversampling technique was applied to balance this dataset. Then, we categorized diabetes dataset using simple k-means clustering and removed unclassified data based on ground truth values. However, different classification techniques such as adaboost, na¨ıvena¨ıve bayes, bayes net, multiple layer perception, linear dis-criminant analysis, quadratic discriminant analysis, k-nearest neighbour, sequential minimum optimization, simple logistic, J48 decision tree and random forest were used to analyze this dataset and all classifiers have shown high performance than previous works. Among all of them, random forest shows the best accuracy (99.067%), kappa statistics (98.09%), precision (99.10%), recall(99.10%), f-measure (99.10%), Matthews correlation coefficient (98.10%), area under receiver operating characteristic (99.90%), area under precision recall curve (99.90%) respectively. To evaluate different classifiers performance with random forest, we applied a non-parametric statistical test named Friedman test and found expected significant results for these classifiers.