AI/MLEarlier work · 2023 – 2026
Diabetes Risk Prediction
A 2023 notebook that estimates Type 2 diabetes risk from eight clinical features on a ~100,000-row public dataset, later packaged as a Streamlit app with an explicit 13-column feature encoding and a model loaded once per container.
Technology stack
Pythonscikit-learnSMOTEpandasStreamlit
01
What it is
- Five model families compared (Random Forest, Gradient Boosting, Logistic Regression, Decision Tree and Gaussian Naive Bayes); Random Forest is served.
- A Streamlit app with an explicit feature-encoding helper, scikit-learn pinned for pickle compatibility, and the model downloaded from Google Drive on cold start, cached per container, and mapped from probability to a risk band.
02
Why no headline metric
The ROC-AUC of 0.996 reported earlier came from oversampling (SMOTE) the whole dataset before the train/test split, which leaks synthetic neighbours of test rows into training. The code now applies SMOTE to the training fold only, but the notebook has not been re-run, so there is no trustworthy metric to show yet.
Sources
- Source code
- Streamlit appsleeps when idle
Checked against the repositories on 22 September 2026.