PTB-XL ECG Classification Project

Multi-class ECG classification using machine learning on the PTB-XL dataset, focusing on handling missing data, feature selection, and class imbalance.

Project Structure

PTB-XL+ paper/
├── src/
│   ├── preprocessing/       # Data preprocessing scripts (run in order)
│   │   ├── 01_merge_and_label.py
│   │   ├── 02_clean_columns.py
│   │   ├── 03_split_data.py
│   │   └── 04_impute_and_flag.py
│   │
│   ├── analysis/           # Analysis and feature selection
│   │   ├── 05_feature_selection_rf.py  # Feature selection (Top-50, Top-100, Top-200)
│   │   ├── missing_data_analysis.py
│   │   ├── outlier_analysis_per_label.py
│   │   └── create_class_weight_visualization.py
│   │
│   └── modeling/           # Model training scripts
│       └── 06_train_models.py  # Main model training (RF Class Weight & Ensemble Undersampling)
│
├── data/
│   └── processed/          # Preprocessed data (generated by preprocessing scripts)
│
├── reports/
│   ├── feature_selection/       # Feature selection plots and lists
│   ├── missing_data_analysis/   # Missing data visualizations
│   └── model_comparison/        # Model performance plots (ROC, Precision/Recall)
│
├── results/                # CSV results of model experiments
│
├── requirements.txt        # Python dependencies
└── README.md              # This file

Setup

1. Install dependencies:

   pip install -r requirements.txt

2. Download datasets:

   • PTB-XL: https://physionet.org/content/ptb-xl/1.0.3/
   • PTB-XL+: https://physionet.org/content/ptb-xl-plus/1.0.1/

   Note: Processed CSV files are not included in the repository. You must run the preprocessing pipeline to generate them.

Pipeline & Usage

The project follows a sequential pipeline. Scripts are located in src/.

1. Preprocessing

Run preprocessing scripts in order:

# Step 1: Merge datasets and process labels
python src/preprocessing/01_merge_and_label.py

# Step 2: Clean columns
python src/preprocessing/02_clean_columns.py

# Step 3: Split data into train/val/test
python src/preprocessing/03_split_data.py

# Step 4: Impute missing values
python src/preprocessing/04_impute_and_flag.py

This pipeline performs:

• Merging PTB-XL metadata with PTB-XL+ features
• Processing labels and creating multi-label columns (NORM, MI, STTC, CD, HYP)
• Cleaning columns (dropping sparse/metadata columns)
• Splitting data into train/validation/test sets using stratified folds
• Imputing missing values and creating missing flags

2. Feature Selection

• Variance Threshold: Removes low-variance features (< 0.05).
• Random Forest Importance: Selects Top-50, Top-100, and Top-200 features based on importance scores.

Select top features using Random Forest importance:

python src/analysis/05_feature_selection_rf.py

This generates:

• Top-50, Top-100, and Top-200 feature lists
• Feature importance plots
• Feature selection reports in reports/feature_selection/

3. Model Training

Train and evaluate models:

python src/modeling/06_train_models.py

This script:

• Trains Random Forest models with Class Weighting (Top-50, 100, 200)
• Trains Random Forest with Ensemble Undersampling (Top-200)
• Generates performance metrics and visualizations

4. Additional Analysis (Optional)

# Missing data analysis
python src/analysis/missing_data_analysis.py

# Outlier analysis per label
python src/analysis/outlier_analysis_per_label.py

# Class weight visualization
python src/analysis/create_class_weight_visualization.py

Methodology

Missing Data Handling

• Analysis: Identified that missingness in P-wave features is correlated with MI and STTC classes (MNAR).
• Strategy: Instead of dropping rows, missing values were imputed with 0 (for signal absence) or median, and binary flags (is_missing_*) were added to preserve the information of "missingness".

Feature Selection

• VarianceThreshold: Removed 186 features with variance < 0.05.
• Random Forest Importance: Selected the top 200 features from the remaining set. PCA was avoided to maintain interpretability.

Class Imbalance Handling

• Class Weighting: Assigning higher weights to minority classes (HYP, CD).
• Ensemble Undersampling: Training multiple models on balanced subsets to reduce bias towards the majority class (NORM).
• Balanced Random Forest: Automatically balances bootstrap samples.

Models Evaluated

1. Random Forest with Class Weighting (Top-50, Top-100, Top-200 features)
2. Random Forest with Ensemble Undersampling (Top-200 features)

Results

The experiments showed that handling class imbalance is crucial. The Random Forest with Top-200 features and Class Weighting provided the best balance between Precision and Recall.

Model	Accuracy	Precision	Recall	F1 Score	ROC-AUC
RF (Top-50, CW)	53.20%	73.58%	66.13%	69.53%	0.904
RF (Top-100, CW)	54.82%	75.40%	67.64%	71.17%	0.909
RF (Top-200, CW)	55.24%	75.91%	67.87%	71.50%	0.915
RF (Top-200, Undersampling)	57.09%	69.23%	68.12%	69.23%	0.915

Note: Metrics are macro-averaged. Detailed per-class metrics are available in the report.

Report

All experiments can be reproduced by running the scripts in order:

1. src/preprocessing/01_merge_and_label.py → 04_impute_and_flag.py
2. src/analysis/05_feature_selection_rf.py
3. src/modeling/06_train_models.py

Results are saved in:

• results/ - CSV files with performance metrics
• reports/ - Visualizations and analysis reports

Presentation is Prepered in:

• Sunum/ - pdf and tex files

Code Environment

• Python: 3.11+
• Key Libraries:
  • pandas, numpy - Data processing
  • scikit-learn - Modeling and evaluation
  • matplotlib, seaborn - Visualization

References

1. Wagner, P., et al. (2020). PTB-XL, a large publicly available electrocardiography dataset. Scientific Data.
2. Strodthoff, N., et al. (2021). PTB-XL+, a comprehensive electrocardiographic feature dataset. Scientific Data.

License

This project is for educational purposes at Istanbul University.