System-identification noise suppression for intra-partum cardiotocography to discriminate normal and hypoxic fetuses

We construct linear system-identification models of cardiotocography (CTG) data collected during labour and delivery. The models are the impulse response functions (IRFs) of the input-output system relating the uterine pressure (UP) stimulus to the fetal heart rate (FHR) response. We compare models obtained with and without applying noise suppression via the pseudo inverse technique. Finally, to determine the ability of the models to discriminate healthy from hypoxic fetuses, we use the average models as feature vectors of a support-vector-machine (SVM) classifier. Applying the pseudo-inverse resulted in cleaner models with lower variance accounted for (VAF), likely indicative of reduced overfitting. The area under curve of the receiver-operator characteristic (ROC) without applying pseudo-inverse was 0.695 ± 0.054. Similar results over a useful operating range of false-positive rates were observed with the pseudo-inverse applied.