MLP-Heart-Disease-Classification

🫀 MLP Heart Disease Classification — A Comparative Study

A Comparative Study of Multi-Layer Perceptron Architectures for Heart Disease Classification

This repository contains the complete code, results, and IEEE-style research report for a systematic evaluation of 18 MLP configurations on the UCI Cleveland Heart Disease dataset.

---

📊 Key Results

Best Model: ReLU activation, 2 hidden layers (64→32), 30 epochs — 83.61% accuracy, F1=0.6667 Worst Model: Sigmoid activation, 3 hidden layers — completely failed to detect disease cases

---

🧪 Experiment Design

18 total experiments combining:

• 3 Activation Functions: ReLU, Sigmoid, Tanh
• 3 Hidden Layer Depths: 1 layer (64), 2 layers (64→32), 3 layers (64→32→16)
• 2 Epoch Settings: 20 and 30

Fixed settings: Batch size=32, Adam optimizer (lr=0.001), Validation split=0.2, StandardScaler normalization

---

📁 Repository Structure

├── mlp_heart_disease.py          # Complete Python code (fully commented)
├── heart.csv                      # Heart Disease dataset
├── experiment_output.log          # Full experiment console output
├── outputs/
│   ├── IEEE_Research_Report.md    # Full IEEE-style research paper
│   ├── comparison_results.csv     # Results table (CSV)
│   ├── detailed_results.txt       # Detailed text results
│   ├── graphs/
│   │   ├── accuracy_comparison_bar.png
│   │   ├── loss_comparison_bar.png
│   │   ├── accuracy_heatmap.png
│   │   ├── epoch_comparison.png
│   │   ├── grouped_activation_comparison.png
│   │   └── top5_metrics_comparison.png
│   ├── training_curves/           # 18 accuracy & loss curve plots
│   │   ├── RELU_-_1_layer_-_20ep.png
│   │   ├── RELU_-_2_layers_-_30ep.png
│   │   └── ... (18 total)
│   └── confusion_matrices/        # 18 confusion matrix plots
│       ├── CM_RELU_-_1_layer_-_20ep.png
│       ├── CM_RELU_-_2_layers_-_30ep.png
│       └── ... (18 total)

---

🔑 Key Findings

1. ReLU is the best activation function — consistently highest accuracy and F1-scores
2. 2 hidden layers is optimal — single layers underfit, 3 layers overfit on this small dataset
3. Sigmoid fails in deep networks — vanishing gradient problem causes complete failure with 3 layers
4. 30 epochs ≥ 20 epochs — modest improvement in most cases
5. Class imbalance hurts recall — all models struggle to detect positive (disease) cases

---

⚙️ How to Run

pip install tensorflow scikit-learn pandas matplotlib seaborn numpy
python mlp_heart_disease.py

---

📄 Research Report

The full IEEE-style research report is available in outputs/IEEE_Research_Report.md, containing:

• Abstract, Introduction, Literature Review
• Methodology with dataset description and preprocessing steps
• Model Development with architecture details
• Complete Results with tables and analysis
• Discussion comparing all 18 configurations
• Conclusion with key insights and future work
• References

---

📈 Sample Visualizations

Accuracy Comparison

Accuracy Heatmap (Activation × Layers)

Epoch Effect Comparison

Best Model Training Curves

Best Model Confusion Matrix

---

Built with TensorFlow 2.21, scikit-learn 1.8, Python 3.x