Abstract
BACKGROUND: Cardiotocography is a common method of electronic fetal monitoring (EFM) for fetal well-being. Data-driven analyses have shown potential for automated EFM assessment. For this preliminary study, we used a novel artificial intelligence method based on fully convolutional networks (FCNs), with deep learning for EFM evaluation and correct recognition, and its possible role in evaluation of nonreassuring fetal status. METHODS: We retrospectively collected 3239 EFM labor records from 292 deliveries and neonatal Apgar scores between December 2018 and July 2019 at a single medical center. We analyzed these data using an FCN model and compared the results with clinical practice. RESULTS: The FCN model recognized EFM traces like physicians, with an average Cohen's kappa coefficient of agreement of 0.525 and average area under the receiver operating characteristic curve of 0.892 for six fetal heart rate (FHR) categories. The FCN model showed higher sensitivity for predicting fetal compromise (0.528 vs 0.132) but a higher false-positive rate (0.632 vs 0.012) compared with clinical practice. CONCLUSION: FCN is a modern technique that may be useful for EFM trace recognition based on its multiconvolutional layered analysis. Our model showed a competitive ability to identify FHR patterns and the potential for evaluation of nonreassuring fetal status.