Diabetes needs continious monitoring to prevent complications. Periodic glucose checks can miss suden flactuations. Regular fingerstick tests and periodic lab draws aren’t built to capture minute by minute shifts. We propose an IoT enabled, real time monitoring system integrated with explainable AI. The core model is TabTransformer, designed for structured medical data, while connected devices continuously stream patient measurements. XAI tools such as SHAP and LIME clarify why a prediction was made, helping clinicians and patients understand each decision. We first evaluate the approach on the Pima Indians Diabetes dataset. Against Random Forest, Logistic Regression, and TabNet, TabTransformer trained with the Adam optimizer reaches 94.50% accuracy. It shows consistent gains in accuracy, precision, and transparency, and compares favorably with prior work on the same dataset. We use standard train/validation splits and cross-validation to reduce overfitting, and results remain stable across repeated runs. To evaluate generalization across populations, we use the UCI Early Stage Diabetes Risk Prediction dataset and to assess real-time adaptability we test the model on raw, unprocessed data. Model performance is measured across both datasets. Beside forecasting risk the system can alert patients when glucose readings change suddenly, enabling timely intervention. Clinicians can help customize thresholds, and if necessary, they can refer patients to other experts. Future work will consider additional health factors and broaden the mix of IoT sensors for richer, real-time data. A reliable and interpretable AI tool for diabetes management with storng predictive perfromance is main goal. • Developed a real-time diabetes monitoring system using IoT and AI integration. • TabTransformer model achieved 94.5% accuracy with ADAM optimizer on medical data. • Incorporated Explainable AI (SHAP, LIME) for transparent and interpretable predictions. • Real-time alerts generated for abnormal glucose levels using implantable sensors. • Outperformed traditional ML models in accuracy, precision, recall, and explainability.
Mukherjee et al. (Sun,) studied this question.