Continuous electronic fetal monitoring (EFM) is used worldwide to visually assess whether a fetus is exhibiting signs of distress during labor, and may benefit from an emergency operative delivery (e.g. Cesarean section). Previously, computerized EFM assessment that mimics clinical experts showed no benefit in randomized clinical trials. However, as an example of routinely collected 'big' data, EFM interpretation should benefit from data-driven computational approaches, such as deep learning, which allow automated evaluation based on large clinical datasets. Here we report our investigation of long short term memory (LSTM) and convolutional neural networks (CNN) in analyzing EFM traces from over 35,000 labors for the prediction of fetal compromise. Of these, 85% are used for training with cross-validation and the remainder are set aside for testing. The results are compared with Clinical practice (reason for operative delivery recorded as fetal distress) and an earlier prototype system for computerized analysis of EFM (OxSys 1.5), developed on the same data. We demonstrate that CNN outperforms LSTM, Clinical practice, and OxSys 1.5 in predicting fetal compromise, with a sensitivity of 42% (30%, 34%, and 36% for the others, respectively), at comparable or lower false positive rates. We also show that increasing the size of the training set improves the sensitivity and stability of CNN's performance on the testing set. When tested on a small open-access external database, CNN moderately improves on the performance of published feature extraction based methods. We conclude that CNN could play an important role in the field of automated EFM analysis, but requires further work.
展开▼
