Towards Early Forecast of Diabetes Mellitus via Machine Learning Systems in Healthcare

Purpose: Diabetes mellitus poses a major challenge to global health, especially in developing nations where early detection and treatment remain difficult. Using a Convolutional Neural Network (CNN) technique, this study seeks to construct an effective early prediction model for diabetes. The focus is on improving diagnostic capacities in settings with limited resources. Materials and Methods: The proposed methodology utilises the PIMA Indians Diabetes dataset. A rigorous data preprocessing pipeline was applied, such as min-max normalization, outlier identification and elimination and organized median imputation of missing values. One possible approach to addressing the issue of class imbalance is the Synthetic Minority Oversampling Technique (SMOTE). The dataset was divided into two parts namely the training set and the testing set by using a ratio of 80:20. Some of the measures adopted to determine the efficiency of a convolutional neural network (CNN) model that was trained on diabetes prediction included accuracy, recall, precision, F1-score, AUC, and Brier score. Findings: The CNN model results were better as compared to the baseline models. It was more accurate than the ANN-based model and the RF-based model, with respective accuracies of 90.29% and 82.35%. With an F1-score of 96.06 per cent and a recall of 96.66 per cent, the CNN model demonstrated considerable faithfulness and predictive power. Unique Contribution to Theory Practice and Policy: The proposed CNN model, with its high accuracy and reliability, has great potential for combining it with a tele-monitoring device to support early diagnosis and routine monitoring of diabetes, especially in underserved areas. Future studies may be of interest in the deployment and validation of the model in real-time within clinics to enhance its practical value.