Heart-Attack-Risk-Prediction

🫀 Heart Attack Risk Prediction – BRFSS 2023

A full-stack ML pipeline to identify heart attack risk using CDC’s 2023 BRFSS dataset. Built with a production-first mindset, this project integrates clean data engineering, interpretable ML, and real-time deployment using AWS services.

Project Outcome: A transparent, explainable, and high-performance heart-risk scoring engine — powered by XGBoost, deployed on SageMaker, and monitored via CloudWatch — ready for insurer use cases like underwriting and population health management.

---

🎯 Problem & Objective

Heart attacks are preventable yet costly. Traditional risk scoring misses early behavioral and comorbidity signals. This solution leverages public health data to deliver accurate, explainable risk predictions.

• Business Need: Insurance underwriting & preventative care
• Goal: Predict risk of heart attack from lifestyle + clinical factors
• End Product: Interpretable ML model with <1s API latency

---

📊 Dataset

• Source: CDC BRFSS 2023
• ~430,000 responses, 350+ fields
• 33 curated features selected post codebook mapping
• Target variable: cvdinfra4 (ever diagnosed with a heart attack)

---

🧱 Project Phases

📥 1. Data Extraction & Preprocessing

• Extracted .xpt survey files, filtered relevant columns
• Cleaned missing/ambiguous codes (7=DK, 9=Refused)
• Standardized binary indicators, renamed columns, created modeling-ready dataset

🧹 2. Data Cleaning & Mapping

• Dropped columns with excessive nulls
• Applied log + z-score transformation on skewed features (e.g., bmi_log_z)
• Mapped categorical codes to readable labels using BRFSS codebook

🧠 3. Feature Insights

• Identified high-risk groups using positive rate profiling (e.g., widowed, low-income)
• Used Cohen’s d, t-tests, chi-squared tests to rank feature importance

📈 4. Exploratory Data Analysis

• Bivariate risk profiling for both categorical and numerical features
• Visual checks, outlier detection, and initial pruning
• Selected statistically significant features for modeling

🧪 5. Model Development (XGBoost)

• Pipeline: log_z numerics + binned features + binary flags
• Baseline AUROC: 0.9885, Recall: 1.00
• Optuna tuning: 150+ trials, minor lift (best scale_pos_weight ≈ 1.2)
• Compared with LGBM & Logistic Regression — XGBoost selected for performance + ecosystem

🧠 6. SHAP Interpretation & Feature Engineering

• Used SHAP for global and individual explanations
• Created engineered features: chronic_burden, preventive_neglect, bmi_risk_code
• Validated model robustness after removing dominant feature (prev_chd_or_mi)

🚀 7. SageMaker Deployment

• Model saved as xgb_top25_shap.joblib and deployed using built-in XGBoost 1.5-1 container
• REST API exposed via ml.m5.large endpoint
• Smoke tested with sample input/output
• Latency consistently <150ms (p95)

📉 8. Monitoring with CloudWatch

• Configured InvocationLatency > 1s alarm via Boto3
• Programmatically triggered and reset on deployment
• No threshold breaches observed in load testing

📊 9. Experiment Tracking with MLflow

• Auto-logged: metrics, parameters, Conda env, artifacts
• Compared baseline vs tuned runs
• Enabled one-click rollback & reproducibility
• Stored trained models + test splits in artifact store

🌐 10. Streamlit Web App

• CSV upload → real-time inference via deployed API
• SHAP visualizations (global + per-row with GPT insights)
• XGBoost vs Logistic Regression comparison toggle

🔁 11. Airflow Orchestration

• Parameterized DAG: data prep → training → deploy → monitor
• Papermill-executed notebooks, IAM-secured tasks
• Airflow DAG updates model + CloudWatch alarm dynamically

---

💡 Key Results

Metric	Value
AUROC	0.9884
Recall	1.00
Precision	~0.66
Inference Time	< 150 ms
SHAP Validated	✅

---

🧰 Tech Stack

Data & Modeling: Python, Pandas, Scikit-learn, XGBoost, Optuna
Explainability: SHAP, OpenAI GPT API
Deployment: AWS SageMaker, Boto3
Monitoring: AWS CloudWatch, SNS
Tracking: MLflow
Orchestration: Airflow + Papermill
Web UI: Streamlit