🫀 Cardiovascular Disease Risk Prediction

A machine learning project for early cardiovascular disease risk screening, optimized for high recall to minimize missed diagnoses in healthcare settings.

---

Project Overview

This project implements an end-to-end machine learning pipeline for predicting cardiovascular disease risk. The system is designed as a decision-support tool for healthcare screening, prioritizing the identification of at-risk patients over overall accuracy.

Key Features:

• Recall-optimized Gradient Boosting model with probability calibration
• Custom decision threshold (0.36) tuned for high sensitivity
• Interactive Streamlit web application for real-time predictions
• Automated BMI calculation from height and weight inputs
• Professional preprocessing pipeline with StandardScaler

---

Problem Statement

Cardiovascular disease (CVD) is one of the leading causes of death globally. Early detection and risk assessment are critical for preventive care. In medical screening scenarios, missing a positive case (false negative) is more costly than a false alarm (false positive).

This project addresses the need for:

• High-sensitivity risk screening to catch potential CVD cases
• A user-friendly interface for healthcare decision support
• Transparent probability scores rather than binary predictions

---

Dataset Description

Source: Kaggle Cardiovascular Disease Dataset
Records: ~70,000 patient records
Target Variable: cardio (1 = cardiovascular disease present, 0 = absent)

Features

Feature	Description	Type
age	Age in years	Continuous
gender	0 = Female, 1 = Male	Categorical
height	Height in cm	Continuous
bmi	Body Mass Index (calculated)	Continuous
ap_hi	Systolic blood pressure	Continuous
ap_lo	Diastolic blood pressure	Continuous
cholesterol	1 = Normal, 2 = Above Normal, 3 = Well Above Normal	Ordinal
gluc	Glucose level (1-3 scale)	Ordinal
smoke	Smoking status (0/1)	Binary
alco	Alcohol intake (0/1)	Binary
active	Physical activity (0/1)	Binary

Preprocessing Applied:

• Removed id column and converted age from days to years
• Replaced weight with calculated BMI feature
• Cleaned outliers in blood pressure readings
• Final processed dataset: ~65,000 records with 11 features

---

Machine Learning Pipeline

1. Data Preprocessing

Raw Data → Outlier Removal → Feature Engineering (BMI) → Train/Test Split (80/20) → StandardScaler

• Stratified split to maintain class balance
• StandardScaler fitted only on training data to prevent data leakage
• Scaler is bundled with the trained model inside models/final_model.pkl

2. Baseline Model Comparison

Multiple algorithms evaluated with focus on recall:

Model	Recall
Logistic Regression	~0.70
K-Nearest Neighbors	~0.65
Support Vector Machine	~0.68
Decision Tree	~0.62
Random Forest	~0.68
Gradient Boosting	~0.72

3. Hyperparameter Tuning

GridSearchCV with 5-fold cross-validation optimized for recall:

Logistic Regression:

param_grid = {'C': [0.01, 0.1, 1, 10]}
# class_weight='balanced'

Gradient Boosting (Selected):

param_grid = {
    'n_estimators': [100, 200],
    'learning_rate': [0.05, 0.1],
    'max_depth': [3, 4]
}

---

Model Training and Optimization

Final Model: Calibrated Gradient Boosting Classifier

The Gradient Boosting model was selected based on:

• Highest baseline recall among ensemble methods
• Good feature importance interpretability
• Stable performance across cross-validation folds

Probability Calibration

Applied Isotonic Calibration using CalibratedClassifierCV to ensure predicted probabilities are reliable and well-calibrated:

from sklearn.calibration import CalibratedClassifierCV

cal_gb = CalibratedClassifierCV(best_gb, method='isotonic', cv=5)
cal_gb.fit(X_train, y_train)

---

Evaluation Results

Final Model Performance (Test Set)

Metric	Value
Recall	~0.83
Precision	~0.66
ROC-AUC	~0.78

Confusion Matrix Interpretation

With the optimized threshold (0.36):

• True Positives: High identification of actual CVD cases
• False Negatives: Minimized (critical for healthcare)
• False Positives: Acceptable trade-off for higher sensitivity

---

Decision Threshold Strategy

Why Not Use Default 0.5?

The default probability threshold of 0.5 optimizes for accuracy, but in medical screening:

• Missing a sick patient (FN) → Delayed treatment, worse outcomes
• Extra screening for healthy patient (FP) → Additional tests, minor inconvenience

Threshold Selection Process

Evaluated recall across thresholds from 0.10 to 0.90:

thresholds = np.arange(0.1, 0.9, 0.05)
for t in thresholds:
    preds = (probabilities >= t).astype(int)
    recalls.append(recall_score(y_test, preds))

Selected Threshold: 0.36

• Achieves ~83% recall (catches most CVD cases)
• Maintains ~66% precision (acceptable false positive rate)
• Balances sensitivity with practical usability

---

Streamlit Application

Features

• User-Friendly Interface: Dropdown menus for categorical inputs
• Automatic BMI Calculation: Enter height and weight, BMI computed automatically
• Probability Score: Shows exact risk probability (0.00 - 1.00)
• Risk Classification: High/Low risk based on 0.36 threshold
• Transparency: Displays threshold and model optimization strategy
• Risk Bands (Low / Moderate / High) for interpretability, separate from screening decision

Input Fields

Input	Type	Range
Age	Number	Years
Gender	Dropdown	Male/Female
Height	Number	50-250 cm
Weight	Number	20-300 kg
Systolic BP	Number	mmHg
Diastolic BP	Number	mmHg
Cholesterol	Dropdown	Normal/Above Normal/Well Above Normal
Glucose	Dropdown	Normal/Above Normal/Well Above Normal
Smoker	Dropdown	Yes/No
Alcohol	Dropdown	Yes/No
Physical Activity	Dropdown	Yes/No

---

How to Run Locally

Prerequisites

• Python 3.10+
• pip package manager

Installation

# Clone the repository
git clone https://github.com/yourusername/cardio-disease-ml.git
cd cardio-disease-ml

# Install dependencies
pip install -r requirements.txt

Train the Model (if not already trained)

Run the modeling notebook or training script to generate model artifacts:

# Option 1: Run Jupyter notebooks
jupyter notebook notebooks/04_modeling_and_evaluation.ipynb

Run the Streamlit App

streamlit run app.py

The app will open at http://localhost:8501

---

Deployment

Streamlit Cloud

1. Push your repository to GitHub
2. Go to share.streamlit.io
3. Connect your GitHub repository
4. Set main file path: app.py
5. Deploy

Required files for deployment:

• app.py
• requirements.txt
• models/final_model.pkl
• models/scaler.pkl

Project Structure

cardio-disease-ml/
├── app.py                    # Streamlit web application
├── Dockerfile                # Docker configuration
├── requirements.txt          # Python dependencies
├── README.md                 # Project documentation
├── data/
│   ├── raw/
│   │   └── cardio_train.csv  # Original dataset
│   └── processed/
│       └── cardio_processed.csv  # Cleaned dataset
├── models/
│   └── final_model.pkl       # Trained model bundle  
├── notebooks/
│   ├── 01_business_and_planning.ipynb
│   ├── 02_data_validation_and_eda.ipynb
│   ├── 03_preprocessing_and_features.ipynb
│   └── 04_modeling_and_evaluation.ipynb
├── src/
│    └── model_training.py     # Model training utilities
└── deployment/
    ├── predict.py            # Prediction utilities
    ├── monitoring.py         # Conceptual monitoring plan (not automated)
    └── retraining.py         # Planned retraining strategy (not automated)

---

Limitations

⚠️ This is a decision-support prototype, NOT a medical diagnosis system.

• Not FDA Approved: This tool has not undergone regulatory approval for clinical use
• Dataset Limitations: Model trained on a specific demographic; may not generalize to all populations
• Feature Limitations: Does not include all clinically relevant factors (family history, ECG, biomarkers)
• No Longitudinal Data: Predictions are point-in-time, not tracking disease progression
• Threshold Trade-off: High recall comes at the cost of more false positives
• Requires Valid Input: Garbage in, garbage out — unrealistic values will produce unreliable predictions

Intended Use:

• Educational and demonstration purposes
• Initial screening to flag patients for further evaluation
• Decision support for healthcare professionals (not replacement)

---

Future Improvements

• Add SHAP explanations for individual predictions
• Implement confidence intervals for probability scores
• Include additional clinical features (if data available)
• Add batch prediction capability for CSV uploads
• Implement model monitoring and drift detection
• A/B testing framework for threshold optimization
• Multi-language support for broader accessibility

---

Tech Stack

• Python 3.10+
• scikit-learn - Machine learning
• pandas / numpy - Data processing
• Streamlit - Web application
• matplotlib / seaborn - Visualization

---

License

This project is for educational and demonstration purposes.

---

Acknowledgments

• Dataset: Kaggle Cardiovascular Disease Dataset
• Inspiration: Healthcare AI best practices for high-stakes prediction

---

Built with ❤️ for better healthcare outcomes