AI-Powered Healthcare Systems: A Comprehensive Study on Machine Learning-Based Diabetes prognosis and Online therapeutic Assistance

Diabetes is one of the most prevalent chronic diseases worldwide, and early detection is critical for preventing severe complications and improving patient outcomes. This study uses the publicly available Diabetes Prediction Dataset from Kaggle, which contains 100,000 patient records with demographic and clinical features such as age, gender, body mass index (BMI), glycated haemoglobin (HbA1c), blood glucose level, smoking history, hypertension and heart disease. Three machine learning models, a Support Vector Classifier, Gaussian Naïve Bayes and a Random Forest, were compared, with the Random Forest achieving the highest accuracy. After class balancing with resampling, the Random Forest was tuned with a cross-validated grid search and reached an accuracy of 95.20% and a ROC-AUC of 0.969 on a held-out test set of 19,230 patients. Feature-importance analysis identified HbA1c (38%), blood glucose level (29%), age (17%) and BMI (10%) as the most influential risk factors. To support real-world use, we developed a web interface in which users enter their health data and instantly receive a diabetes risk assessment. Overall, the results show that a tuned and interpretable ensemble model can be delivered as an accessible screening tool that supports proactive disease management.