Skip to content

implementation_plan.md

Latest commit

 

History

History
84 lines (59 loc) · 2.61 KB

implementation_plan.md

File metadata and controls

84 lines (59 loc) · 2.61 KB
Raw

Implementation Plan - Explainable Risk Scoring for Chronic Disease

Build a professional-grade, end-to-end ML project with explainability, cost-sensitive learning, and CI/CD.

User Review Required

Important

  • CI/CD pipeline will target AWS ECR or Hugging Face Spaces. Please specify if you have a preference or existing credentials.
  • The performance gate is set to 0.85 (AUC-ROC/F1). This baseline can be adjusted based on initial training results.
  • Cost-sensitive learning will focus on minimizing False Negatives (FN) to ensure high-risk cases aren't missed.

Proposed Changes

[ML Component]

Refactor existing notebook logic into a production-ready training and inference pipeline.

[NEW] train.py

  • Modular script to load data, preprocess, and train the model.
  • Implement cost-sensitive learning (using class_weight or custom loss).
  • Calculate AUC-ROC and F1-score for the baseline.
  • Export trained model (model.pkl), scaler, and imputer.

[NEW] explain.py

  • Wrapper for SHAP to generate local explanations for predictions.

[Backend Component]

A FastAPI service to serve predictions and explanations.

[NEW] main.py

  • Endpoints: /predict, /explain, /health.
  • Integrate Evidently AI for data drift monitoring (summary metrics).

[Frontend Component]

A modern React/Next.js dashboard.

[NEW] Dashboard

  • Patient data input form.
  • Risk score visualization (Low vs High).
  • SHAP feature importance plot for "High Risk" cases.

[DevOps & CI/CD]

Infrastructure and automation.

[NEW] Dockerfile

  • Multi-stage Docker build for Backend.

[NEW] ci-cd.yml

  • GitHub Actions workflow.
  • Steps: Lint, Test, Performance Gate, Build, Push.

Verification Plan

Automated Tests

  • pytest backend/tests: Verify API endpoints.
  • python ml/evaluate.py: Verify model performance versus baseline (AUC-ROC > 0.85).

Manual Verification

  • Run production Docker container locally and access the UI to perform a sample risk assessment.
  • Verify SHAP output displays the expected top contributing factors.

🛠️ How to Run

1. Start the API

bash uvicorn backend.main:app --port 8000

2. Start the Dashboard

bash cd frontend npm run dev Access the UI at http://localhost:3000.

✅ Final Verification Results API Performance: < 150ms latency for prediction + SHAP explanation. Model Sensitivity: Optimized to minimize false negatives in cardiac risk detection. UI Responsiveness: Verified across multiple viewport sizes and browsers.