Fetal Heart Rate Classification with Convolutional Neural Networks and the Effect of Gap Imputation on Their Performance

Cardiotocography (CTG) is widely used to monitor fetal heart rate (FHR) during labor and assess the wellbeing of the baby. Visual interpretation of the CTG signals is challenging and computer-based methods have been developed to detect abnormal CTG patterns. More recently, data-driven approaches using deep learning methods have shown promising performance in CTG classification. However, gaps that occur due to signal noise and loss severely affect both visual and automated CTG interpretations, resulting in missed opportunities to prevent harm as well as leading to unnecessary interventions. This study utilises routinely collected CTGs from 51,449 births at term to investigate the performance of time series gap imputation techniques (GIT) when applied to FHR: Linear interpolation; Gaussian processes; and Autoregressive modelling. The implemented GITs are compared by studying their impact on the convolutional neural network (CNN) classification accuracy, as well as on their ability to correctly recover artificially introduced gaps. The Autoregressive model has been shown to be more reliable in the classification and recovery of artificial gaps when compared to the Linear and Gaussian interpolation. However, the improvement in the classification accuracy is relatively modest and does not reach statistical significance. The median (interquartile range) of sensitivity at 0.95 specificity is 0.17 (0.14,0.18) and 0.16 (0.13, 0.17) for the Autoregressive model and the zero imputations (baseline method) respectively (Mann-Whitney U = 69, P = 0.16). Future work include investigation and evaluation of other gap imputation methods to improve the classification performance of CNN on larger dataset.