Effective Feature Selection and Soft Voting Classifier based Diabetes Detection Using Machine Learning Approaches

An unnecessarily increased blood glucose level is a symptom of the metabolic disorder diabetes mellitus (DM). Early diabetes detection is reduce the risk factor of kidney disease, heart failure, liver cirrhosis, it might be hazardous. Our motivation is to developed a machine learning models to forecast diabetes in the future. We applied efficient data cleaning, label encoding, normalization and pre-processing methods for better models accuracy. Prediction with synthetic minority over-sampling (SMOTE) implies the equilibrium of an unstable dataset. Different feature selection techniques such as Chi-square test, Information gain attribute evaluator, Extra trees classifier, SelectKbest, and correlation-based approach to finding the key variables. Machine learning algorithms including Random Forest(RF), K-nearest neighbor(KNN), Naive Bayes (NB), Support Vector Machine (SVM), Gradient Boosting (GB), Decision Tree (DT), and Logistic Regression (LR), have been applied in the initial phase. Random Forest accuracy of 95.03% after normalizing the dataset with synthetic minority over-sampling (SMOTE) techniques. We using a voting classifier for ensemble all machine learning models in the second phase. The models achieved soft voting highest accuracy 96.69%. The findings of our models comparing with some current research indicate that it can offer greater accuracy. Now, we can accurately forecast diabetes.