codeSignal AI

Inspiration

Hospital readmissions are expensive and often preventable, but clinicians need fast and understandable risk signals at discharge. I wanted to build a healthcare AI tool that predicts risk and explains it in plain language.

What it does

CareSignal AI predicts a patient’s 30-day readmission risk from key clinical features and returns a risk score with a risk tier. It also highlights the top drivers behind the prediction and generates a short clinical summary.

How we built it

I built the project in Zerve using Python, synthetic healthcare data, feature engineering, gradient boosting, calibration metrics, and SHAP explainability. I then packaged the workflow into a GitHub-ready project with FastAPI-based inference logic.

Challenges we ran into

The biggest challenge was deployment reliability, especially endpoint routing and dependency issues in the hosted environment. I solved this by simplifying the API layer, validating requests in Postman, and separating the stable notebook workflow from the deployment-ready codebase.

Accomplishments that we're proud of

I shipped an end-to-end healthcare AI project from analysis to deployable inference, not just a static notebook. The project combines prediction, interpretability, and narrative generation in a way that feels useful for real discharge planning workflows.

What we learned

I learned that calibration and explainability matter as much as raw accuracy in healthcare use cases. I also learned that turning a notebook into a usable product depends heavily on clean interfaces, reproducibility, and deployment discipline.

What's next for codeSignal AI

Next, I want to improve model quality with richer clinical and social-determinant features and compare multiple modeling approaches. I also want to add a clinician-facing app and stronger deployment monitoring for real-world use.

Built With

• fastapi
• github
• numpy
• pandas
• pydantic
• python
• scikit-learn
• shap
• zerve