Related Experiment Video
Updated: Sep 15, 2025
04:36
Noninvasive Electrocardiography in the Perinatal Mouse
Published on: June 12, 2020
6.2K
Cross-Database Evaluation of Deep Learning Methods for Intrapartum Cardiotocography Classification.
Lochana Mendis1, Debjyoti Karmakar2, Marimuthu Palaniswami1
1Department of Electrical and Electronic EngineeringThe University of Melbourne Parkville VIC 3010 Australia.
Summary
Deep learning models show promise for detecting fetal compromise from cardiotocography (CTG) recordings. Using fetal heart rate (FHR) signals with specific preprocessing improved performance, with ResNet excelling in classification.
Area of Science:
- Perinatal medicine
- Artificial intelligence in healthcare
- Signal processing
Background:
- Cardiotocography (CTG) monitors fetal heart rate (FHR) and uterine contractions (UC) to assess fetal compromise during labor.
- Visual interpretation of CTG is challenging, leading to low sensitivity and necessitating automated methods.
- Limited datasets and lack of standardized evaluation hinder deep learning model development for CTG analysis.
Purpose of the Study:
- To conduct a cross-database evaluation of deep learning models for fetal compromise detection using CTG data.
- To investigate the impact of signal selection, preprocessing, and data sampling on model performance.
- To establish a standardized workflow for comparing and developing deep learning algorithms for CTG classification.
Main Methods:
- Utilized a private dataset (9,887 recordings) and the open-access CTU-UHB dataset (552 recordings).
- Evaluated six deep learning models, exploring variations in FHR/UC input, signal preprocessing (artefact removal, interpolation), downsampling frequency, and pH sample inclusion.
- Employed class activation maps to assess model interpretability by aligning with clinical knowledge.
Main Results:
- Preprocessing FHR signals with artefact removal and interpolation significantly improved classification performance for certain models.
- Excluding intermediate pH samples did not substantially enhance performance across any model.
- ResNet demonstrated superior fetal compromise classification performance across both datasets at a 1Hz downsampling rate.
- Class activation maps confirmed that ResNet focused on clinically relevant FHR patterns.
Conclusions:
- A standardized workflow for evaluating deep learning models in CTG classification is proposed.
- Optimized signal preprocessing and FHR-only input can enhance model performance.
- The ResNet model shows strong, interpretable performance, offering a potential benchmark for future research.
- Ensuring generalizability and interpretability is crucial for robust clinical application of these AI tools.
