175 - Machine Learning Analysis of Cardiotocographs (CTGs) for Predicting Neonatal Encephalopathy (NE) Among Term Infants
Saturday, April 26, 2025
2:30pm – 4:45pm HST
Publication Number: 175.5874
Xiaoxu Rong, University of Wisconsin School of Medicine and Public Health, Madison, WI, United States; Ryan M. McAdams, University of Wisconsin, Middleton, WI, United States; Daniel Pimentel Alarcon, University of Wisconsin-Madison, Madison, WI, United States; Claudette O. Adegboro, University of Wisconsin School of Medicine and Public Health, Madison, WI, United States
Assistant Professor University of Wisconsin School of Medicine and Public Health Madison, Wisconsin, United States
Background: Neonatal encephalopathy (NE), often caused by birth-related hypoxia or ischemia, can result in severe neurodevelopmental disabilities or death. Conventional cardiotocograph (CTG) monitoring has limited predictive accuracy, underscoring the need for enhanced diagnostic tools. Machine learning, particularly deep learning, offers a promising approach to detecting NE patterns in CTGs. However, robust model performance typically requires large datasets. Domain adaptation, which involves pretraining on broader datasets and fine-tuning with smaller, specific datasets, may improve model accuracy. Objective: To assess the feasibility of a deep learning model in identifying CTG patterns associated with NE among term infants. Design/Methods: We developed a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) model to capture the spatial and temporal features of CTGs. The CNN layers identify spatial variations in CTG patterns, while the LSTM layers detect sequential changes relevant to NE diagnosis. The model was pretrained on a public CTG dataset from the University Hospital in Brno (April 2010–August 2012) and fine-tuned using an IRB-approved private dataset from UnityPoint Health-Meriter Hospital, Madison, Wisconsin (January 2015–December 2021). NE cases were diagnosed using consistent inclusion criteria (Table 1), and model performance was evaluated using Area Under the Curve (AUC) metrics. Results: The public dataset included 552 CTG records, and the private dataset included 87 records, of which 10 were NE cases. There were no significant differences in gestational ages or birthweights between NE cases and controls (mean GA for NE cases = 39.57 ± 1.52 weeks, BW = 3314.75 ± 757.18 g; controls: GA = 39.23 ± 1.64 weeks, BW = 3315.07 ± 834.55 g). The baseline AUC of the CNN-LSTM model on the private dataset was 0.49 ± 0.05 without pretraining. After domain adaptation, the model’s AUC improved to 0.62 ± 0.07, indicating moderate enhancement in NE pattern detection.
Conclusion(s): Domain adaptation significantly increased the model's ability to leverage public CTG data to improve NE detection accuracy within a limited private dataset. While promising, the model’s predictive performance remains modest, indicating the need for larger datasets and broader validation. Future studies should explore model generalizability across diverse CTG datasets to develop a robust, real-time NE prediction tool that supports early intervention and improved neonatal outcomes.
