Why-Be-So-Septic/mcp-server

MCP Server — Sepsis Outcome Prediction

Overview This repository implements a Model Context Protocol (MCP) server exposing a tool that predicts clinical outcomes for ICU/sepsis patients based on their longitudinal lab results and metadata. The tool hides implementation details (like HTTP requests) from the LLM and provides a clean interface: supply patient records, receive per-day outcome predictions.

Key Features

• MCP tool: predict_sepsis_outcome
• MCP tool: compute_lab_score
• Input schema validated with Zod; all fields optional and nullable with default null
• Gender normalization for backend ("male"/"female" → "M"/"F")
• FastAPI backend integration at http://localhost:8000/predict with required checkpoint_path
• Text-based tool output (JSON stringified) to comply with MCP response types

Tool: predict_sepsis_outcome Description Analyzes medical data for a patient with sepsis and predicts their clinical outcome across time.

Input The tool expects an object with key data containing an array of patient records. Each record represents a measurement day for a patient and accepts the following fields:

• id: string | number | null
• StartDate: string (YYYY-MM-DD) | null
• EndDate: string (YYYY-MM-DD) | null
• Aging: number | null
• Gender: "male" | "female" | null
• data: string (measurement date, YYYY-MM-DD) | null
• Optional numeric labs (all number | null): ALT, APTT, AST, Amilasa, CRP, Conj_bilirubin, Creatinin, D_dimer, Ferritin, Fibrinogen, Glucose, Hemoglobin, LDG, Lymphocytes, Monocytes, Neutrophils, Platelets, Potassium, Procalcitonin, Sodium, Total_protein, Urea, WBC

Important

• The tool does not accept Outcome or DaysBfOutcome in the request payload. These fields are intentionally excluded from the schema and will be rejected.
• Missing fields are allowed and default to null.

Output interpretation For each day t in a patient’s timeline the model predicts one of:

• 0 = no event within ≤7 days (patient remains hospitalized)
• 1 = discharge within ≤7 days
• 2 = death within ≤7 days

Backend Integration

• Endpoint: POST http://localhost:8000/predict
• Required body shape: { "checkpoint_path": "tsformer_best.pt", "patients": [ ...normalized records... ] }
• Gender values are mapped automatically by the tool: "male" → "M", "female" → "F". Null stays null.
• The backend should sanitize any NaN/Infinity values in responses (replace with null) to avoid JSON serialization errors.

Example request to the MCP tool { "data": [ { "id": "12345", "StartDate": "2025-01-10", "EndDate": "2025-01-20", "Aging": 67, "Gender": "male", "data": "2025-01-12", "ALT": 35.0, "APTT": 28.5, "AST": 22.0, "CRP": 8.7, "Creatinin": 1.1, "WBC": 7.8 } ] }

Example tool output (text) Patient analysis complete. Sepsis outcome predictions available. { "predictions": [ { "id": "12345", "date": "2025-01-12", "y_hat": 0 } ], "meta": { "model": "tsformer_best.pt" } }

Getting started Prerequisites

• Node.js 20+
• pnpm 9+ (recommended)
• Running FastAPI backend at http://localhost:8000 with /predict endpoint

Install dependencies pnpm install

Run in development pnpm dev

Build pnpm build

Type check pnpm typecheck

Notes

• The MCP server is implemented in src/index.ts. The tool description includes a detailed schema and example.
• The server returns MCP-compliant text content; JSON is stringified before returning.
• If you change the backend path or checkpoint file, update src/index.ts accordingly.

License Proprietary / Internal (update as appropriate).