Inspiration
Many data scientists and healthcare developers focus heavily on model training while overlooking the quality and cleanliness of the data itself. This often leads to models that look great on paper but fail in the real world due to hidden flaws like impossible values or label leakage. We were inspired to build MedData Doctor to act as an automated "health check-up" for medical datasets, ensuring data integrity before training any AI model.
How we built it
We developed MedData Doctor as a robust web application featuring an intuitive user interface. Users can either upload their own CSV datasets or instantly generate and test a synthetic dataset using our built-in demo tool (Load synthetic demo data). The app automatically audits columns, computes a comprehensive Data Health Score, detects critical issues, and provides actionable, clear suggested fixes.
Challenges we faced
Automating Medical Checks:* Designing a precise auditing engine capable of automatically flagging sensitive flaws like target/label leakage (e.g., columns recorded post-diagnosis) and implausible medical values (e.g., zero values in insulin or skin thickness).
- User Experience (UX): Ensuring the audit results are categorized clearly (using color-coded severity flags) so developers can easily understand and resolve issues instantly.
What we learned
We learned the critical importance of rigorous data preprocessing and validation in healthcare AI. A model is only as good as its underlying data, and catching data risks early saves time and prevents catastrophic real-world deployment failures.
Built With
- data
- healthcare
- mechinelearning
- python
- science
- streamlist
