Heart Attack Risk Prediction - In-Hospital Mortality ML Pipeline🫀

(A Modular, Production-Style ML Pipeline with Class-Imbalance Handling)

A reproducible ML system from end-to-end data ingestion, feature engineering, model comparison, SMOTE-based imbalance correction, and CLI-driven training & evaluation workflows.

Demonstrates how to go from raw CSV → cleaned features → baseline models → reproducible CLI pipeline, with optional SMOTE to address severe class imbalance.

---

🎯 Project Overview

Goal	Description
Prediction Task	Predict whether a patient died during hospitalization (DIED = 1).
Dataset	Clinical + administrative attributes (age, sex, length of stay, DRG, charges, diagnosis codes).
Class Imbalance	Positive class ≈ 9% → minority recall is the key metric.
Key Finding	SMOTE significantly improves DIED=1 recall across multiple models; Decision Tree delivered the strongest baseline.

---

📊 Architecture Overview

Architecture Diagram	Heart Attack ML Pipeline

---

🛠️ Tech Stack

• Core: Python, pandas, scikit-learn, imbalanced-learn
• ML Pipeline: modular scripts, SMOTE support, artifact saving
• Environment: virtualenv + VS Code / terminal
• Extras: Jupyter Notebook for EDA & visualizations

---

📂 Repository Structure

<project-dir>/
├─ data/
│  ├─ raw/               # Raw CSV dataset (place whole\_table.csv here)
│  └─ processed/         # Cleaned data, metrics, trained models (.joblib)
├─ notebooks/
│  └─ heart_attack_eda.ipynb   # EDA + visualizations
├─ src/
│  ├─ __init__.py
│  ├─ config.py           # Paths & constants
│  ├─ data_loader.py      # CSV loading + basic cleaning (e.g., coercing CHARGES numeric)
│  ├─ features.py         # Preprocessing (OHE for categoricals, scaling for numerics)
│  ├─ models.py           # Model factory, SMOTE, training logic,training implementation
│  └─ evaluate.py         # Metric utilities (accuracy, recall, reports)
├─ scripts/
│  ├─ train.py            # CLI script for model training
│  └─ evaluate_model.py   # CLI script for model evaluation
├─ requirements.txt
└─ README.md

---

⚙️ Models & Methods

The project trains six classic ML models:

• Naive Bayes
• KNN
• Decision Tree
• Logistic Regression
• SVM
• MLP (Neural Network)

SMOTE is optionally applied to mitigate heavy class imbalance.

Metrics evaluated:

• Accuracy
• Recall for class 0 and 1
• Emphasis: positive-class recall (DIED=1)

---

📈 Model Performance

ROC Curve & Confusion Matrix

📊 Feature Importance & SMOTE Impact

Below are the key model insights, including feature importance rankings and
a visualization of class distribution before and after applying SMOTE.

---

🏃 Quick Start — Run the Pipeline

1. Environment Setup

python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2. Place the Dataset

Put the dataset here:

<project-dir>/data/raw/whole_table.csv

Or pass a custom path via --raw-csv.

3. Train a Model

Example: Decision Tree + SMOTE

cd <project-dir>
python -m scripts.train --model decision_tree --smote

Example: Logistic Regression (no SMOTE)

bash
python -m scripts.train --model logistic_regression

Outputs generated:

• data/processed/trained_model.joblib
• data/processed/metrics.json

4. Evaluate a Saved Model

python -m scripts.evaluate_model --model-path data/processed/trained_model.joblib

5. Optional: Run the EDA Notebook

jupyter notebook
# open: notebooks/heart_attack_eda.ipynb

The notebook mirrors the full workflow (EDA → preprocessing → modeling → visualization).

---

🧪 Notes & Implementation Details

• CHARGES may contain "." — coerced to numeric with errors="coerce"
• Missing values handled in preprocessing
• SMOTE activated via --smote
• Minority recall emphasized because DIED=1 is medically critical
• Designed as a teaching example, not a clinical decision tool

---

⚠️ Disclaimer

This project is for educational purposes only and not intended for clinical or medical use.