Related Experiment Video
Updated: Aug 4, 2025
08:07
Author Spotlight: Advancing Labor Management Through Electromyometrial Imaging for Understanding Uterine Contractions
Published on: May 26, 2023
1.2K
End-to-end learning with interpretation on electrohysterography data to predict preterm birth
A M Fischer1, A L Rietveld2, P W Teunissen3
1Department of Computer Science, Vrije Universiteit, De Boelelaan 1105, Amsterdam, 1081 HV, The Netherlands; Department of Obstetrics and Gynecology, Amsterdam UMC Location AMC, Meibergdreef 9, Amsterdam, 1105 AZ, The Netherlands.
Computers in Biology and Medicine
|April 5, 2023
Summary
Deep learning models for predicting preterm birth using electrohysterography (EHG) show promise. Adding clinical data did not improve EHG model performance, but an interpretability framework was developed.
Area of Science:
- Biomedical Engineering
- Machine Learning in Healthcare
- Obstetrics and Gynecology
Background:
- Predicting preterm birth remains a clinical challenge.
- Electrohysterography (EHG) signals uterine electrical activity, offering potential for preterm birth prediction.
- Interpreting EHG signals requires specialized signal processing knowledge, limiting clinical utility.
Purpose of the Study:
- To apply Deep Learning models, specifically Long Short-Term Memory (LSTM) and Temporal Convolutional Networks (TCN), to electrohysterography (EHG) data for preterm birth prediction.
- To evaluate the impact of incorporating clinical data alongside EHG data on model performance.
- To develop and validate an interpretability framework for time series classification models in low-data scenarios.
Main Methods:
- Utilized the Term-Preterm Electrohysterogram database for model training and evaluation.
- Implemented end-to-end Deep Learning models (LSTM and TCN) for EHG signal analysis.
- Assessed model performance with and without the inclusion of available clinical data.
- Developed a novel interpretability framework for time series classification.
Main Results:
- End-to-end Deep Learning models achieved an Area Under the Curve (AUC) of 0.58, comparable to traditional machine learning models using handcrafted features.
- The addition of clinical data to EHG data did not lead to a significant performance improvement.
- The proposed interpretability framework demonstrated suitability for limited datasets.
Conclusions:
- Deep Learning models offer a viable approach for analyzing EHG data for preterm birth prediction, achieving performance comparable to existing methods.
- Clinical data, in its current form, does not enhance the predictive power of EHG-based models.
- The developed interpretability framework provides clinicians with insights into model predictions and highlights the need for high-risk patient datasets to reduce false positives.
