Optimized machine learning based predictive diagnosis approach for diabetes mellitus

Abstract

Aims: Diabetes mellitus is a metabolic disease caused by elevated blood sugar. If this disease is not diagnosed on time, it has the potential to pose a risk to other organs and tissues. Machine learning algorithms have started to preferred day by day in the detection of this disease, as in many other diseases. This study suggests a diabetes prediction approach incorporating optimized machine learning (ML) algorithms.
 Methods: The framework presented in this study starts with the application of different data pre-processing processes. Random forest (RF), support vector machine (SVM), K-nearest neighbor (K-NN) and decision tree (DT) algorithms are used for classification. Grid search is utilized for hyperparameter optimization of algorithms. Different performance evaluation measures are used to find the algorithm that best predicts diabetes. PIMA Indian dataset (PID) is chosen for testing the experiments. In addition, it is investigated to what extent the attributes in the data set affect the result using Shapley additive explanations (SHAP) analysis.
 Results: As a result of the experiments, the RF algorithm achieved the highest success rate with 89.06%, 84.33%, 84.33%, 84.33% and 0.88% accuracy, precision, sensitivity, F1-score and AUC scores. As a result of the SHAP analysis, it is found that the “Insulin”, “Age” and “Glucose” attributes contributed the most to the prediction model in identifying patients with diabetes.
 Conclusion: The hyperparameter optimized RF approach proposed in the framework of the study provided a good result in the prediction and diagnosis of diabetes mellitus when compared with similar studies in the literature. As a result, an expert system can be designed to detect diabetes early in real time using the proposed method.