File size: 2,853 Bytes
1bd8cdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
* [x] Initialize new Hugging Face Space with Gradio SDK 5.x

  * Add `mcp-server-track` tag in `README.md`
* [ ] Write Python function `symptom_to_diagnosis(symptom_text)`

  * Use OpenAI or Anthropic API to generate JSON
  * Format prompt to request JSON output
  * Parse model response into Python dict
  * Handle JSON formatting quirks (trim extra text, use `json.loads`)
  * Implement fallback rule-based mapping for demo cases
* [ ] Test `symptom_to_diagnosis` function

  * Input common symptom examples and combinations
  * Verify relevance and correctness of ICD codes and diagnoses
  * Tweak prompt to improve specificity and JSON validity
* [ ] Define confidence score methodology

  * Decide whether to use model’s self-reported scores or rank order as proxy
  * Document how confidence is calculated and interpreted
* [ ] Integrate function into Gradio Blocks interface

  * Use `gr.Interface` or `gr.ChatInterface` to accept symptom text input and display JSON output
  * Configure Gradio app metadata to expose MCP endpoint
* [ ] Build demonstration client or script (optional)

  * Create minimal client using `gradio.Client` or `requests` to call Space’s prediction API
  * Alternatively, build a second Gradio Space as a simple chatbot that calls the MCP tool
  * Prepare screen recording showing AI agent (e.g., Claude Desktop) calling the MCP endpoint with example query
* [ ] Update `README.md` documentation

  * Describe tool functionality and usage examples
  * Include `mcp-server-track` tag, link to video or client demo
  * List technologies used (e.g., “OpenAI GPT-4 API for symptom→ICD mapping”)
* [ ] Configure OpenAI/Anthropic API usage

  * Use cheaper models (e.g., GPT-3.5) during development
  * Reserve GPT-4 or Claude-2 for final demo queries to conserve credits
* [ ] Evaluate Hugging Face / Mistral credits for alternative inference

  * Identify open ICD-10 prediction models on HF Inference API (e.g., `AkshatSurolia/ICD-10-Code-Prediction`)
  * Consider running open-source models on Mistral if time allows
* [ ] Plan Modal Labs usage for cloud compute (optional)

  * Pre-compute ICD-10 embeddings in Modal job if semantic search is added
  * Host backend microservice or Gradio app on Modal if HF Space resources are insufficient
* [ ] Reserve Nebius or Hyperbolic Labs credits for GPU-intensive tasks (if needed)

  * Spin up GPU instance to host or fine-tune open-source model only if HF Space times out
* [ ] Consider LlamaIndex integration for retrieval-augmented generation (bonus)

  * Load ICD-10 dataset into LlamaIndex and test semantic search for candidate codes
  * Implement minimal index of common diagnoses for demo if time permits
* [ ] Record and document final demo

  * Capture symptom input, MCP tool invocation, and JSON output in a short video
  * Host video link in `README.md`