(237) Predicting fetal acidemia using convolutional neural networks trained on intrapartum electronic fetal monitoring data

Continuous electronic fetal monitoring (EFM) is widely used to assess fetal well-being during labor. However, its interpretation is subjective, inconsistent, and poorly correlated with neonatal outcome. We sought to develop and evaluate a convolutional neural network (CNN) to predict fetal acidemia using intrapartum fetal heart rate data.

Study Design:

We conducted a retrospective study of singleton, > 37 0/7 weeks, liveborn deliveries with EFM data at a tertiary care hospital from 2011–2023 for which pH or APGAR data were available. Only births with high-quality FHR signal in the 30 minutes prior to time of birth (defined as non-zero signal in 70% of samples) were included for training which restricted our dataset to vaginal births (Figure 1). We applied a convolutional neural network to predict a composite marker of neonatal distress, defined as pH < 7.2 or APGAR at 1 minute < 7. We evaluated performance on 50 sample held-out test-set using the area under the receiver operating characteristic curve (AUC). Both training and test sets were balanced with positive and negative samples. We assessed model performance on various input types and architectures of relevance to the field.

Results:
62,299 deliveries met inclusion criteria. Of these, 6,379 had pH test obtained and the remainder had APGAR available. Of these, 5,423 met criteria for our composite marker of neonatal distress (of which 3,391 had pH < 7.2). The model achieved an AUC of 0.83 + 0.01 in predicting acidemia on the holdout dataset. We found a single-channel model with only fetal heart rate performed better than a two-channel model that also included uterine activity and that the model performed best with filter-sizes of 2-minutes in length (Figure 2).

Conclusion:
A deep learning model can accurately predict fetal acidemia using short segments of FHR data in the final minutes of labor. We have further demonstrated operational proof of concept of EFM streaming, integrated this into our existing data pipeline. Future work will focus on prospective validation, integration into clinical decision making, and evaluation of impact on outcomes.