CTG Classification & Regression Benchmark (Classical ML & Transformer Models)

1. Overview This repository contains the full experimental pipeline used to evaluate three-class clinical classification and 0–100 continuous clinical score regression on real-world cardiotocography (CTG) time-series data. The project compares: Five classical machine learning models, and Five transformer-based deep learning models, plus A transformer-based regression framework predicting the continuous clinical score (0–100). All analyses are performed on segmented CTG tracings consisting of FHR (Fetal Heart Rate), UA (Uterine Activity), and AFM (Automatic Fetal Movement) channels. 2. Repository Structure ├── ClassicModels.ipynb ├── Transformer_iTransformer.ipynb ├── Transformer_PatchCTG.ipynb ├── Transformer_PatchTST.ipynb ├── Transformer_TimesNetLite.ipynb ├── Transformer_ETCNN.ipynb ├── Transformer_Regression.ipynb ├── _scored_output.xlsx └── README.md 3. Description of Files A) Classical Machine Learning — Three-Class Classification ClassicModels.ipynb Trains and evaluates five widely used classical ML algorithms: Logistic Regression LightGBM Random Forest SVM (RBF kernel) XGBoost Includes: Preprocessing and feature handling SMOTE-balanced training data, original distribution test data Accuracy, Precision, Recall, Macro-F1 Class-wise ROC curves and macro-ROC metrics Confusion matrices B) Transformer-Based Models — Three-Class Classification Each transformer model is implemented in a separate notebook for modularity and reproducibility: Transformer_iTransformer.ipynb Implements the iTransformer architecture for multivariate long-range time-series modeling. Transformer_PatchCTG.ipynb Patch-based transformer optimized specifically for CTG segmentation and feature aggregation. Transformer_PatchTST.ipynb Implementation of the PatchTST architecture for multivariate physiological time-series. Transformer_TimesNetLite.ipynb Lightweight version of TimesNet using multi-scale temporal decomposition. Transformer_ETCNN.ipynb Hybrid encoder combining CNN feature extraction and transformer-based temporal modeling. All transformer notebooks include: Train/validation loop Optimizers (AdamW / Lion) LR scheduling Early stopping ROC curves (per class + macro) Complete classification metrics C) Transformer-Based Regression — 0–100 Continuous Score Transformer_Regression.ipynb Predicts the continuous 0–100 clinical evaluation score (regression task), using: PatchTST PatchCTG iTransformer TimesNetLite ETCNN Outputs: MAE, RMSE, R² Pearson & Spearman correlations Bland–Altman analysis Calibration curves Scatter plots for predicted vs. true scores This notebook corresponds to the Regression Axis described in the associated thesis/manuscript. D) Additional Output _scored_output.xlsx Contains: True class labels True continuous clinical scores Predicted probabilities (3-class models) Predicted regression outputs Threshold-based validation metrics Data used to generate all figures (ROC, PR, Bland–Altman, calibration, etc.) 4. Tasks & Experimental Axes This repository covers three analytical axes: Axis 1 — Three-Class Classification (Classical ML) Evaluates baseline machine learning algorithms on CTG-derived segments. Axis 2 — Three-Class Classification (Transformers) Assesses the discriminative performance of modern transformer architectures. Axis 3 — Regression (0–100 Continuous Clinical Score) Predicts clinician-derived continuous assessment values and compares regression accuracy across transformer-based models. 5. Requirements This project requires: Python >= 3.10 PyTorch >= 2.0 scikit-learn xgboost lightgbm imbalanced-learn numpy, pandas einops matplotlib, seaborn GPU acceleration (Colab / A100 / CUDA-enabled environment) is strongly recommended for transformer models. 6. How to Use Provide dataset paths in the notebooks. Run preprocessing and segmentation (if needed). Execute classical models via ClassicModels.ipynb. Execute transformer-based classification models individually. Run Transformer_Regression.ipynb for continuous score prediction. Use _scored_output.xlsx for analysis, visualization, and statistical comparisons. 7. Citation Once published on Zenodo, please cite: APA Karataş, F. (2025). CTG Classification & Regression Benchmark: Classical Machine Learning and Transformer Models [Code and Analysis]. Zenodo. https://doi.org/10.5281/zenodo.17612016 BibTeX @dataset{karatas2025ctgbenchmark, author = {Karataş, Ferhat}, title = {CTG Classification and Regression Benchmark: Classical Machine Learning and Transformer Models}, year = {2025}, publisher = {Zenodo}, doi = {https://doi.org/10.5281/zenodo.17612016}, } 8. License This work is shared under the Creative Commons Attribution 4.0 (CC-BY 4.0) license.You are free to reuse, modify, and distribute with proper attribution. 9. Contact Dr. Ferhat KarataşDepartment of Computer EngineeringFirat University, TürkiyeORCID: 0000-0002-9462-6092