BACKGROUND
Cardiotocography (CTG) is a form of electronic fetal monitoring that provides continuous graphic representation of fetal heart rate (FHR) and uterus activity (UA)during pregnancy. CTG is a commonly used method to assess the well-being of the fetuses by clinicians.
OBJECTIVE
-
METHODS
A publicly available CTG dataset was generated by extracting numerical features from electronic signals of 2126 CTG records that were labeled as normal (n=1655, 77.8%), pathological (n=176, 8.3%) or suspect (n=295, 13.9%) by expert physicians. In this study, we applied Deep Neural Networks (DNNs), Support Vector Classifier (SVC), Random Forest, Extreme Gradient Boosting (XGBoost) and Light Gradient Boosting (LightGBM) classification algorithms to predict normal and pathological fetal states from CTG records.
RESULTS
Our results show that DNN outperformed conventional ML classifiers with an AUC value of 0.99.
CONCLUSIONS
Considering the unbalanced class distribution, we further applied decision threshold optimization and used the optimized DNN model to provide reliable probability estimates for the fetal state of suspect class samples.