Performance Assessment of Different Machine Learning Algorithms in Predicting Diabetes Mellitus

Abstract

Diabetes Mellitus (DM) is a heretical metabolic disorder and a commonly distributed long-term sluggish poison that poses a serious threat to human health. Faster and more precise diabetes diagnosis is critical, and Machine Learning (ML) will help improve medical health technologies and build an e-healthcare infrastructure. Thirteen machine learning algorithms have been thoroughly researched in this respect, and Jupyter Notebook has been used to apply them. Hence, The ML models are trained with dataset of Kaggle machine learning data repository of Frankfurt hospital, Germany. To achieve stable accuracy, a 5-fold cross validation approach is proposed as an effective data processing method. However, in order to improve the efficiency of the ML models, the hyper-parameter tuning technique is used. Gaussian Process (GP) emerged as the highest performing algorithm and is proposed as the most effective classifier with an accuracy of 98.25% after comprehensive simulation. However, the accuracy of Random Forest (RF) and Artificial Neural Network (ANN) was 97.25% and 96.5%, respectively, which is very well. As a result, the performance of the ML models is evaluated using various metrics such as Accuracy, Sensitivity, Precision, F1-score, Specificity, and ROCAUC, and a graphic comparison of all the ML models is shown. Effective diabetes prediction using machine learning algorithms can help to reduce annual mortality rates, particularly in developing countries like Bangladesh. As a result, this research will help healthcare providers manage Diabetes Mellitus more effectively and efficiently, paving the way for a more robust e-healthcare system in the future