deep learning method for locating fetal heart rate decelerations during labour using crowd-sourced data

Monitoring the heart rate of a fetus is the only method for continuously monitoring fetal well-being during labour. Decelerations in the fetal heart rate, usually corresponding with contractions, provide information regarding the ability of the fetus to cope with a reduced oxygen supply caused by contractions. For this reason, visual and computer-based assessments of fetal well-being depend upon the correct identification of decelerations. Agreement between clinical professionals when identifying these features is poor. Automated methods offer a more consistent alternative, but must perform at a level comparable to a trained human observer. Collecting data to train these systems can be difficult due to time restraints on experts in fetal medicine. In this study, we designed a website to crowd-source data from clinical professionals at their convenience. This data was combined with that from other publicly available sources to produce a dataset that was used to train a convolutional auto-encoder to locate decelerations. Crowd-sourcing was found to be an effective method for collecting data that was suitable to successfully train a deep learning model to locate decelerations. The model trained using this data provided a comparable level of accuracy to users of the website and other models used for the same purpose.