Developing and Evaluating a Machine Learning Based Diagnosis System for Diabetes Mellitus using Interpretable Techniques

Diabetes is a major global health issue that affects multiple bodily components and contributes to millions of deaths each year. Traditional approaches to diabetes diagnosis and treatment are often limited by their lack of accuracy, transparency, and efficiency. This study aims to develop and evaluate a novel machine learning-based diagnosis system for diabetes mellitus using interpretable supervised and neural network techniques. The study used a dataset of 9 features listed in 2000 patient information from The Frankfurt Hospital, Germany, and trained and tested several ML algorithms including logistic regression, gradient boosting, naive Bayes classifier, random forest classifier, and artificial neural network (ANN). The performance of each algorithm was evaluated using precision, recall, and F1-score, and the findings indicate that the ANN model performs best with a larger number of features, achieving 100% accuracy. Interpretable techniques were used to facilitate understanding of the ML model decision-making process. The suggested system offers several implications and potential impacts on healthcare practice, including improved diagnosis accuracy, automation of diabetes testing and referral algorithms, and reduced time, work, and labor in medical services. These findings highlight the potential of machine learning to address the limitations of traditional diabetes diagnosis and treatment, and contribute to better patient outcomes.