CTGDL: A Multi-source cardiotocography dataset for fetal stress prediction and CTG analysis

CTGDL: Cardiotocography Deep Learning Dataset CTGDL: Cardiotocography Deep Learning Dataset and Tools. CTGDL provides harmonized cardiotocography (CTG) datasets, metadata, and pretrained deep learning models for research on fetal monitoring, fetal stress prediction, and CTG signal analysis.The repository integrates three major CTG sources—SPaM, FHRMA, and CTU‑UHB—into a unified, standardized format suitable for machine learning and benchmarking. 📚 Dataset Descriptions SPaM Dataset (CTG Challenge 2017) The SPaM dataset consists of intrapartum CTG recordings collected for the 2nd Signal Processing and Monitoring (SPaM) in Labour Workshop held in Oxford, UK, in 2017. The associated CTG Challenge 2017 provided a shared benchmark for methods in fetal heart rate (FHR) and uterine contraction (UC) signal analysis, focusing on labour monitoring, fetal stress detection, and advanced signal processing and classification techniques. Source and documentation: Workshop page:https://users.ox.ac.uk/~ndog0178/spam2017.htm Data description:https://users.ox.ac.uk/~ndog0178/CTGchallenge2017.htmhttps://users.ox.ac.uk/~ndog0178/CTGchallenge2017.pdf Download (original workshop database):http://ctg.ciirc.cvut.cz/other_files/CTG_workshop_database.zip In this repository, the SPaM subset is used as preprocessed time‑series files (FHR and UC at 4 Hz, with missing data preserved via NaN markers where applicable), ready for use in machine learning pipelines for fetal stress prediction and CTG pattern analysis. FHRMA Dataset The FHRMA dataset contains fetal heart rate recordings with expert morphological annotations, including baseline, accelerations, and decelerations. It was originally introduced as a resource for training and evaluating algorithms for CTG morphological analysis and enables supervised and transfer‑learning approaches for detecting clinically relevant patterns in FHR. In this repository, FHRMA recordings are standardized to a common sampling rate and exported to tabular formats with aligned annotation channels to facilitate downstream modelling. CTU‑UHB Dataset The CTU‑UHB intrapartum CTG database comprises 552 recordings collected at Czech Technical University and University Hospital Brno, with detailed clinical outcome data including umbilical artery pH, Apgar scores, and maternal/neonatal parameters. Each recording includes FHR and UC signals acquired during labour, typically starting up to 90 minutes before delivery, and is accompanied by rich metadata suitable for supervised outcome prediction and retrospective CTG analysis. In this repository, CTU‑UHB data are provided as standardized 4 Hz FHR/UC time series and harmonized metadata tables, enabling consistent integration with SPaM and FHRMA for multi‑source learning, benchmarking, and fetal stress prediction experiments. 🧠 Fetus Stress Prediction Notebook 📘 Predict Fetal Stress Using a Trained PatchTST Model This repository includes the notebook pred_with_trained_PatchTST_and_plot.ipynb, which demonstrates how to apply a pretrained PatchTST model to fetal heart rate (FHR) and uterine contraction (UC) signals.The notebook performs the full end‑to‑end workflow for fetal stress prediction: What the notebook does: Loads the pretrained PatchTST classification model Downloads only the required folders from the CTGDL dataset using sparse checkout Loads preprocessed CTG signals (ctgdl_proc_samples) Extracts valid 1800‑point windows from each signal Converts each window into PatchTST patches Runs the model to compute fetal stress probability Saves prediction CSV files for each signal Visualizes: FHR and UC traces Predicted stress probability Optional CTU‑UHB metadata overlays The notebook is designed to run entirely in Google Colab, with no installation or local setup required. ▶️ Open the Notebook in Google Colab Run the full workflow directly in your browser: Open in Colab:https://colab.research.google.com/github/naomifridman/CTGDL/blob/main/pred_with_trained_PatchTST_and_plot.ipynb Notebook file:pred_with_trained_PatchTST_and_plot.ipynb (located in the root of the repository) 📦 Required Folders (Automatically Downloaded in Colab) The notebook automatically fetches only the necessary directories using sparse checkout: ctgdl_proc_samples/ — processed CTG signals trained_model/ — pretrained PatchTST classification model CTU_UHB/ — metadata for visualization No full repository download is required. 📜 Licensing CTU‑UHBThe license for the CTU‑UHB dataset is the Open Data Commons Attribution License v1.0 (ODC‑BY‑1.0). FHRMAFHRMA is distributed under the GPL‑3.0 license. SPaM (CTG Challenge 2017)The license for the SPaM dataset is defined by a Data Use Agreement (DUA).Therefore, it cannot be uploaded or redistributed in this repository.Users must download the data themselves and use the provided code to process it. SPaM source and documentation: Workshop page:https://users.ox.ac.uk/~ndog0178/spam2017.htm Data description:https://users.ox.ac.uk/~ndog0178/CTGchallenge2017.htmhttps://users.ox.ac.uk/~ndog0178/CTGchallenge2017.pdf Download (original workshop database):http://ctg.ciirc.cvut.cz/other_files/CTG_workshop_database.zip