Automated interpretation of cardiotocography using deep learning in a nationwide multicenter study

Timely detection of abnormal cardiotocography (CTG) during labor plays a crucial role in enhancing fetal prognosis. Recent research has explored the use of deep learning for CTG interpretation, most studies rely on small, localized datasets or focus on outcomes less relevant to clinical practice. To address these limitations, we developed a clinically applicable model using a large-scale, nationwide CTG dataset with reliable annotations provided by a board-certified obstetrician. Our study utilized 22,522 deliveries from 14 hospitals, each including cardiotocography (CTG) recordings of up to 75 min in length. The CTG signals were segmented into 5-minute intervals, resulting in a total of 519,800 person-minutes of analyzed data. We trained and validated a deep learning model based on CTG segments for classifying normal and abnormal CTGs. In the independent test dataset, the model achieved an AUC (area under the receiver operating characteristic curve) of 0.880 and PRC (area under the precision-recall curve) of 0.625 in internal tests. External tests across three datasets achieved AUCs of 0.862, 0.895, and 0.862 and PRCs of 0.553, 0.615, and 0.601. Our study results show the potential of the deep learning for automated CTG interpretation. We will evaluate this model in future prospective studies to assess the model's clinical applicability.