Spaces:
Running
Develop a medical-symptom MCP server that takes patient-entered symptoms, maps them to ICD-10 codes via a local knowledge base, and returns a JSON of probable diagnoses with confidence scores.
Verdict: Your medical symptom → ICD-10 MCP server idea appears to be a strong candidate for winning. It aligns with a high-impact, real-world problem, demonstrates a structured use of AI (which highlights technical skill), and can be made user-friendly as an API/tool. It’s also something you’re motivated by (you even have a PhD contact interested in it), which is important for execution. Furthermore, since another team is working on clinical triage, it validates the domain – but you can differentiate your project by focusing on ICD-10 coding and confidence scoring, giving it a unique edge.
To win: Highlight the Use of Credits/Tech in Submission: Explicitly mention in your README or presentation how you utilized the provided resources. For example: “This project was built using Modal’s cloud GPUs to preprocess data, OpenAI’s GPT-4 for code validation (using hackathon API credits), and LlamaIndex for efficient retrieval.” This not only gives credit to the sponsors (good etiquette) but also reinforces that you made the most of the hackathon’s offerings – a trait of a resourceful hacker.
Current goal: Your goal for the first few hours (and days) should be to get a simple version of the Symptom-to-ICD MCP server running, then iterate. Given your solid Python skills and background, you can handle the coding; the LLM will handle the medical heavy-lifting. Keep the project solo-buildable by avoiding getting stuck in rabbit holes – lean on pre-built models and services for anything outside your expertise (that’s the whole point of those credits!).
Next goal: Once the MVP is functional, focus on polish: improve the prompt for accuracy, add a few example cases in your README, maybe even a quick video demo of you typing in symptoms and getting JSON out. Showcasing the agent integration (even a basic one) will fulfill the MCP aspect and impress judges that your tool truly augments an AI agent’s capabilities. For instance, imagine a brief demo where an AI agent is asked “Patient has chest pain and shortness of breath, what could be the issue?” – the agent calls your MCP tool, then uses the returned JSON to answer “It might be unstable angina (ICD I20.0) and I recommend urgent evaluation.” Such a demo would clearly illustrate the power of your project.