Case study — Independent
Healthcare Risk Prediction
An independent machine-learning web app that predicts diabetes risk from 21 lifestyle and health inputs.
- Type
- Machine-learning application
- Role
- Sole developer — model integration & interface
- Stack
- Python · Streamlit · XGBoost
- Live
- Streamlit app ↗
Context & problem
A deployed machine-learning web application that exposes a healthcare risk-prediction workflow through an interactive browser interface.
The project demonstrates the full path from machine-learning logic to a usable web product: trained classifier, real-time inference, and results a non-technical user can act on.
My responsibilities
- Train and integrate an XGBoost classifier on 21 lifestyle, health-history, and demographic features.
- Build the interactive Streamlit frontend around the prediction workflow.
- Implement a personalized recommendation engine alongside the prediction output.
- Deploy the application on Streamlit Cloud.
Implementation & technical decisions
- XGBoost classifier reporting approximately 86% accuracy — the share of correct predictions on the evaluation data — with ROC AUC above 0.82, meaning the model discriminates well between risk classes.
- 21-feature input workflow feeding real-time inference with immediate result display.
- Recommendation engine translating model output into personalized guidance.
Challenges & trade-offs
- Machine-learning output is only useful if the interface earns trust: inputs must be clear, results legible, and recommendations framed as guidance, not diagnosis.
Outcomes & lessons
- A live ML application connecting model-driven logic to user inputs and displayed results.
- Working evidence of the React-to-Python range: the same engineer who maintains SaaS dashboards can ship model-backed products.