DiabetesCare AI

DiabetesCare AI predicts a patient’s risk of diabetes from their health metrics and then uses GenAI to turn that prediction into something genuinely useful — personalised dietary and lifestyle advice.

On the prediction side, the model is a Random Forest classifier, tuned with GridSearchCV and trained on a dataset of around 100,000 records. It reaches about 0.94 accuracy on the test set. The user enters indicators like age, gender, hypertension, heart disease, smoking history, height, weight, HbA1c level and blood glucose, with BMI computed automatically, and the app returns a risk prediction along with a simple visualisation of the result.

Where it goes beyond a plain prediction is the Gemini-1.5-Flash integration. For patients flagged at risk, the model generates tailored lifestyle and dietary suggestions, including pointers to relevant resources and hospitals in India. There’s also a chatbot built on the same model, so patients can ask follow-up questions and get advice in a conversational way, with their query history kept through the session.

The whole thing is built as a Streamlit web app, using scikit-learn for the model and Seaborn/Matplotlib for the visualisations.

website now live at DiabetesCare AI

code available at repo