Comparative Evaluation of Machine Learning Models for Predicting Treatment Non-response in Idiopathic Premature Ventricular Contractions

A multicenter study comparing six machine learning approaches — Logistic Regression, MLP, XGBoost, TabTransformer, TabNet, and Kolmogorov-Arnold Network (KAN) — for predicting pharmacological treatment non-response in idiopathic PVC patients.

Key Results (External Validation, N=366)

Model	AUROC	Brier Score	ICI	F1 Score
XGBoost	0.987	0.034	0.018	0.930
KAN	0.971	0.068	0.053	0.872
TabTransformer	0.959	0.055	0.023	0.885
TabNet	0.958	0.064	0.046	0.868
MLP	0.946	0.077	0.025	0.833
Logistic Regression	0.927	0.090	0.032	0.782

Clinical Calculator

An open-access web-based prediction tool is available at:

https://huggingface.co/spaces/halil21/ipvc-treatment-predictor

Clinicians can input patient-specific parameters and obtain individualized risk predictions from multiple models, with no software installation required.

Project Structure

VES-Deep-Learning/
├── notebooks/
│   └── VES_deep_learning_revised.ipynb   # Training & evaluation notebook
├── r_scripts/
│   ├── analysis_revised.R                # Performance metrics & figures
│   └── perform33.R                       # 34-metric evaluation functions
├── model_weights/                        # Trained model checkpoints
│   ├── logistic_regression_model.pkl
│   ├── xgboost_model.pkl
│   ├── tabtransformer_model.pth
│   ├── kan_model.pth
│   ├── mlp_model.pkl
│   ├── tabnet_model.zip
│   ├── scaler.pkl
│   └── label_encoders.pkl
├── figures/                              # All publication figures
│   ├── figure2_roc_pre_recalib.*         # ROC curves
│   ├── figure3_calibration_pre_recalib.* # Calibration plots
│   ├── figure4_dca_pre_recalib.*         # Decision curve analysis
│   ├── figure5_calibration_post_recalib.*
│   ├── figure6_roc_post_recalib.*
│   ├── figure7_dca_post_recalib.*
│   ├── shap_*.png                        # SHAP summary plots
│   └── supp_*.png                        # Supplementary figures
├── requirements.txt
└── .gitignore

Setup

Python Environment

git clone https://github.com/ihtanboga/VES-Deep-Learning.git
cd VES-Deep-Learning
pip install -r requirements.txt

Running the Analysis

1. Training & Evaluation (Python): Open notebooks/VES_deep_learning_revised.ipynb in Jupyter or Google Colab.

2. Performance Metrics & Figures (R):

   Rscript r_scripts/analysis_revised.R

   Required R packages: pROC, CalibrationCurves, ggplot2, dplyr, dcurves, gridExtra, viridis, boot, tidyr

Methodology

• Training cohort: 1,278 patients from 8 centers across Istanbul
• External validation cohort: 366 patients from 3 centers in the Anatolia region
• Internal validation: Stratified 80/20 train-validation split within Istanbul cohort for deep learning models; 5-fold cross-validation for classical ML models
• Preprocessing: Label encoding and standard scaling fit exclusively on training data
• Recalibration: Logistic recalibration (Platt scaling) on Istanbul validation subset
• Sensitivity analysis: Models retrained without PVC burden to assess feature dependency

Features (20 predictors)

Numeric (14): PVC burden (%), PVC QRS duration, LVEF, Age, PVC prematurity index, QRS ratio, Mean heart rate, Symptom duration, QTc sinus, PVC coupling interval dispersion, CI variability, PVC peak QRS duration, PVC coupling interval, PVC compensatory interval

Categorical (6): Multifocal PVC, Non-sustained VT, Gender, Hypertension, Diabetes mellitus, Full compensation

Citation

If you use this code or the clinical calculator, please cite:

Tanboga IH, et al. Comparative Evaluation of Machine Learning Models Including TabTransformer and Kolmogorov-Arnold Networks for Predicting Treatment Non-response in Idiopathic Premature Ventricular Contractions: A Multicenter Study. [Journal], 2025.

Contact

For questions, contact haliltanboga@yahoo.com.

License

This project is provided for academic and research purposes.