Early prediction of diabetes mellitus type II in Oman using Artificial Intelligence

Abstract

The increasing prevalence of Type 2 Diabetes Mellitus (T2DM), particularly in Oman— where cases are projected to rise by 174% by 2050—necessitates the development of accurate, region-specific predictive models for early detection and risk stratification. This study develops an artificial intelligence (AI)-based predictive framework incorporating two Oman-specific datasets—the Oman Prediabetes Dataset and the Oman Screening Dataset—to improve predictive performance beyond widely used datasets such as the Pima Indian Diabetes Dataset (PIDD). To determine an optimal predictive model, this research evaluates traditional machine learning algorithms alongside three deep learning models: a 1D Convolutional Neural Network (1D CNN for Structured Data) for structured medical records, a 7-layer Long Short-Term Memory (LSTM) network for sequential patient data modelling, and a Hybrid CNN-LSTM model, which integrates spatial and temporal learning for clinical risk assessment. The models were trained and validated using preprocessing, feature selection, and hyperparameter tuning, with performance assessed through accuracy, precision, recall, specificity, F1-score, and AUCROC metrics. The Hybrid CNN-LSTM model achieved the highest performance, with 99.58% accuracy, 100% sensitivity, 99.55% precision, 99.50% specificity, an F1-score of 99.78%, and an AUC-ROC of 97.07%, demonstrating reliability in identifying individuals at high risk of developing T2DM. The seven-layer LSTM model achieved 99.40% accuracy, 100% precision, 100% sensitivity, and 99.34% specificity, confirming its effectiveness in sequential health data modelling. The 1D CNN model outperformed traditional machine learning methods, attaining 99.24% accuracy, 100% precision, 90.2% sensitivity, 100% specificity, and an F1-score of 94.85%, highlighting its suitability for structured data analysis. This research also introduces region-specific datasets to address the limitations of widely used datasets, improving prediction accuracy for populations with distinct genetic and lifestyle factors. A Graphical User Interface (GUI) was developed to facilitate real-time risk prediction, batch processing, and secure data handling in healthcare environments. By integrating localised datasets with deep learning techniques, this research establishes a scalable AI-based framework for early T2DM detection, contributing to precision medicine, clinical decision support, and AI-driven healthcare solutions for Oman and other regions with similar healthcare challenges.