Heart Disease Prediction API — MLOps OPPE2

End-to-end MLOps pipeline for heart disease classification, deployed on GKE with CI/CD, explainability, fairness testing, and drift detection.

Project Info

• Model: Logistic Regression (baseline)
• Dataset: 14 clinical attributes (age, gender, cp, trestbps, chol, etc.)
• Accuracy: 83.05%

Stack

• Training: scikit-learn, MLflow (http://136.111.167.244:8100/)
• API: FastAPI + Uvicorn (Docker)
• Deployment: GKE (oppe2final, us-central1), HPA max 3 pods
• CI/CD: GitHub Actions → GCR → GKE

Deliverables

D2 — Model Explainability (SHAP)

• Analysed false negatives using SHAP LinearExplainer
• cp (chest pain type) had the strongest negative SHAP values — the model is least dependent on low cp values for predicting heart disease

D3 — Fairness Testing (Fairlearn)

• Sensitive attribute: age
• Demographic Parity Difference: 0.2105 ⚠ (bias detected)
• Equalized Odds Difference: 0.3889

D4 — Dockerized API on GKE

• FastAPI app containerized and pushed to GCR
• Deployed on GKE with HPA (min 1, max 3 pods, CPU threshold 60%)
• CI/CD via GitHub Actions on every push to main

D5 — Per-Sample Inference with Logging

• 100 random samples run through deployed API
• Each request logged with input, prediction, probability, latency
• Avg latency: ~0.27ms | P95: ~0.34ms

D6 — Stress Testing (wrk)

• Tool: wrk -t10 -c2000 -d60s
• 2000 concurrent connections, 60 seconds
• HPA scaled pods during load test

D7 — Input Drift Detection (KS Test)

• All 13 features showed significant drift (p < 0.05)
• Synthetic data vs training data compared using Kolmogorov-Smirnov test

Repository Structure

├── train.py                        # Training, explainability, fairness, drift
├── inference.py                    # 100-sample inference against deployed API
├── app/main.py                     # FastAPI application
├── Dockerfile                      # Container definition
├── requirements.txt                # Python dependencies
├── k8s-deployment.yaml             # GKE Deployment + Service + HPA
├── wrk_script.lua                  # wrk load test script
└── .github/workflows/ci-cd.yml     # GitHub Actions CI/CD pipeline

API Endpoints

Endpoint	Method	Description
/health	GET	Health check
/predict	POST	Single prediction
/predict_batch	POST	Batch predictions

Quick Start

curl -X POST http://<EXTERNAL-IP>/predict \
  -H "Content-Type: application/json" \
  -d '{"age":55,"gender":1,"cp":2,"trestbps":130.0,"chol":250.0,"fbs":0,"restecg":1,"thalach":150.0,"exang":0,"oldpeak":1.5,"slope":1,"ca":0,"thal":2}'