A Comparative Framework for Diabetes Mellitus Diagnosis Using Artificial Neural Network Models

Introduction: The growing global concern for the high prevalence of diabetes mellitus has resulted in an increased focus on measures for its early and accurate diagnosis and strategies for its control. Various machine learning and deep learning models can be utilized for accurate diabetes diagnosis. This study focuses on designing a comparative framework for Diabetes Mellitus detection using the deep learning model of Artificial Neural Networks (ANN). The study examines the behaviour of the altering hyperparameters, learning rate, epochs, and the number of hidden layers in an ANN, on the accuracy of diabetes mellitus prediction. Materials and Methods: This study developed ANN models on a novel diabetes dataset for diabetes mellitus prediction, and compared them on key performance metrics. Results: The results demonstrate that an ANN model with two and three hidden layers and learning rates of 0.005 and 0.0005, respectively, achieved 88.0% accuracy and an AUC score above 0.8. Discussion: Interpreting the ROC Curve, the study found that small or very large epochs showed lower accuracy in the dataset. Learning rates of 0.005 and 0.0005 demonstrated a higher performance than the learning rate of 0.05. Conclusion: A holistic plan for diabetes management requires integrating deep learning-based ANN models for prediction, coupling them into clinical practice. Future studies would apply the hybrid and ensemble algorithms to the novel dataset and compare the performance of the prediction. Our research would contribute to the global effort in diabetes mellitus management, thus reducing the overall cost of this life-long disease.