Heart Disease Prediction

Kaggle Competition | AUC: 0.95524 (Private LB) | Rank: 604 / 4371

---

Competition Performance

 Metric                        Score
 ----------------------------- ----------------------------
 Private Leaderboard AUC       0.95524
 Public Leaderboard AUC        0.95377
 Final Rank                    604 / 4371 (Top \~14%)

This repository contains my structured experimentation pipeline developed for the Kaggle competition:

Playground Series - Season 6, Episode 2: Predicting Heart Disease

The focus was on disciplined cross-validation, model evolution, and competitive generalization performance.

---

Project Objective

Predict the probability of heart disease using structured clinical features by:

• Exploring multiple model families
• Applying deep learning techniques to tabular data
• Performing hyperparameter optimization
• Building a robust Out-of-Fold (OOF) validation pipeline
• Maintaining leaderboard stability

The goal was not just leaderboard score --- but strong validation methodology and reproducibility.

---

Models Explored

Gradient Boosting

• XGBoost
• LightGBM
• CatBoost

Neural Networks

• Dense Neural Networks (TensorFlow / Keras)
• Regularization tuning (Dropout, BatchNorm)
• Depth/width exploration
• Learning rate scheduling

---

Technical Strategy

1️ Preprocessing

• Missing value handling
• Scaling & normalization
• Encoding strategies
• Distribution adjustments

2️ Feature Engineering

• Interaction features
• Binning
• Frequency Encoding
• GroupMean Encoding
• Statistical combinations
• Iterative refinement
• Feature selection experiments

3️ Optimization

• Hyperparameter tuning with Optuna
• Early stopping strategies
• Cross-validation stability monitoring
• Overfitting mitigation

---

Validation Framework

To avoid leaderboard overfitting:

• K-Fold Cross Validation
• Out-of-Fold (OOF) prediction tracking
• Consistent AUC comparison across folds
• Submission pipeline from OOF-trained models

This ensured strong alignment between public and private leaderboard scores.

---

Project Structure

Heart-Disease-Prediction/
│
├── submision_outputs/
├── medal-winning-notebooks/
├── heart-disesase-prediction/
├── best_documented_iteration/
├── best_performing_iteration_and_submission/
└── README.md

---

Tech Stack

• Python
• TensorFlow / Keras
• XGBoost
• LightGBM
• CatBoost
• Optuna
• Scikit-learn
• Pandas / NumPy
• Jupyter Notebook

---

What This Project Demonstrates

• ✔ Competitive machine learning workflow
• ✔ Multi-model experimentation
• ✔ Deep learning applied to structured tabular data
• ✔ Strong cross-validation discipline
• ✔ Hyperparameter optimization pipelines
• ✔ Clean experiment iteration structure

---

Competition Citation

Yao Yan, Walter Reade, Elizabeth Park. Predicting Heart Disease. https://kaggle.com/competitions/playground-series-s6e2, 2026. Kaggle.

---

Future Improvements

• Model ensembling (stacking / blending)
• Feature importance stability analysis
• Experiment tracking
• Deployment-ready inference pipeline

---

Author

Adwait Tagalpalewar