CTG GPU Benchmark Pipeline: Transformer-Based Fetal Monitoring Models with Reproducible Python Scripts

1. Overview This repository provides a reproducible GPU benchmark pipeline for transformer-based fetal monitoring models using cardiotocography (CTG) data.It includes complete scripts for signal preprocessing, data segmentation, model training, and prediction scoring, enabling fair performance comparison across multiple deep learning architectures. All data are anonymized.The codebase is optimized for GPU environments such as Google Colab, Kaggle Notebooks, or local CUDA setups, and can be executed interactively (via Jupyter) or automatically (via Python CLI). 2. Repository Contents File Type Description 01_ExamData.csv Raw Data file Sample dataset containing fetal monitoring recordings with three synchronized channels: FHR (Fetal Heart Rate), UA (Uterine Activity), and AFM (Automatic Fetal Movement). 02_SplitData.ipynb Jupyter Notebook Preprocessing and segmentation notebook that splits the raw data into training, validation, and test sets. 19_GPU.ipynb Jupyter Notebook Interactive GPU training notebook for multiple time-series architectures (e.g., TSMixer, PatchTST, iTransformer, TCN, ResNet1D). Computes performance metrics such as MAE, RMSE, and R². ctg_gpu_benchmark_v3_full.py Python script Fully automated pipeline script for command-line execution. Handles model selection, dataset loading, GPU training loops, and output storage. _scored_output.xlsx Data file Contains model predictions and ground-truth values on the test set for subsequent statistical analysis and figure generation. 3. Usage Notes Dependencies: Python ≥ 3.10 PyTorch ≥ 2.1 scikit-learn, pandas, numpy, matplotlib, tqdm GPU with CUDA support recommended Suggested Execution Order:1️⃣ 01_ExamData.csv → raw data from NST signals2️⃣ 02_SplitData.ipynb → preprocessing & segmentation3️⃣ 19_GPU.ipynb → model training (interactive mode)4️⃣ ctg_gpu_benchmark_v3_full.py → automated benchmark5️⃣ _scored_output.xlsx → source dataset License & Citation:This dataset and code package are released for academic and research use.When referencing, please cite as: Karataş S. et al., “CTG GPU Benchmark Pipeline: Transformer-Based Fetal Monitoring Models with Reproducible Python Scripts,” Zenodo (2025). DOI: https://doi.org/10.5281/zenodo.17456942 4. Purpose The goal of this project is to establish a transparent and reproducible benchmark for deep learning models applied to fetal cardiotocography (CTG).By standardizing preprocessing, segmentation, and evaluation procedures, this pipeline enables objective performance comparison across transformer and classical time-series models, supporting future research in clinical fetal health prediction and perinatal data science. 5. Keywords Cardiotocography, Fetal Monitoring, Transformer Models, Deep Learning, GPU Benchmark, Reproducibility, Python, Time-Series Analysis