Heart Disease Prediction Web App

Inspiration

Heart disease is one of the leading causes of death worldwide. I wanted to build a practical tool that uses machine learning to estimate heart attack risk from everyday health metrics, and make it accessible to anyone through a simple web interface.

What it does

The Heart Disease Prediction Web App takes 23 health and lifestyle inputs — age, cholesterol, blood pressure, BMI, smoking, exercise, diet, stress, and more — and classifies a patient as high or low risk for a heart attack in real time.

How I built it

• Dataset: 8,763 patient records from the Kaggle Heart Attack Prediction Dataset
• Preprocessing: Decomposed blood pressure into Systolic/Diastolic, encoded categorical features, scaled numerics, and handled missing values
• Models: Trained and compared Logistic Regression, Random Forest, and XGBoost with class-weight balancing to address a 64/36 class imbalance
• Metric focus: Prioritized Recall to minimize false negatives in a medical context
• Interface: Built an interactive Streamlit web app for real-time risk prediction
• Deployment: Containerized with Docker and deployed to Hugging Face Spaces

Challenges I ran into

• Class imbalance (64% vs 36%) required weighted loss functions and threshold tuning
• Preprocessing sequencing errors initially affected model training and had to be debugged
• Synthetic dataset limitations capped performance around 65% accuracy regardless of tuning

Accomplishments that I'm proud of

• Shipped a fully working end-to-end ML pipeline — from raw CSV to a live public demo
• Successfully containerized the app and deployed it to Hugging Face Spaces
• Built a clean, interactive UI that any non-technical user can use

What I learned

• How to handle class imbalance in real medical datasets
• End-to-end ML deployment using Docker and Hugging Face Spaces
• The importance of feature engineering (e.g., decomposing blood pressure into Systolic/Diastolic)
• How to balance Recall vs Precision in healthcare ML

What's next for Heart Disease Prediction Web App

• Train on a larger, real clinical dataset for stronger accuracy
• Add explainability with SHAP values so users can see why a prediction was made
• Integrate with wearables (Apple Health, WHOOP) for live vitals input

Built With

• docker
• hugging-face
• machine-learning
• matplotlib
• numpy
• pandas
• python
• scikit-learn
• seaborn
• streamlit
• xgboost