Lead-Time Dependence of Fetal Acidaemia Prediction from Intrapartum Cardiotocography: A Balanced-Panel Benchmark of Feature-Based and Deep Classifiers Against Clinical Baselines

Models that predict fetal acidaemia from intrapartum cardiotocography (CTG) are almost always fitted to the segment of trace immediately preceding delivery and summarised by a single figure of discrimination, a design that cannot show whether a warning could be produced early enough to act upon. This study makes the interval to delivery an explicit variable. From 552 recordings in an open-access intrapartum database, a balanced panel of 318 women with 100 acidaemic cases was formed in which four non-overlapping 15-minute windows, ending 45, 30, 15 and 0 minutes before birth, were all of usable quality. One cross-validation partition was reused for every window, feature set and model. Six feature-based learners on 61 physiological descriptors and four sequence networks on the waveform were compared with twelve maternal and obstetric variables known before labour and with a window-adapted FIGO 2015 reading of the same windows. The cross-validated area under the receiver operating characteristic curve (AUROC) was 0.527, 0.492, 0.667 and 0.636 along the axis. Against an empirical permutation null, the two earliest windows were indistinguishable from chance and the two latest were not. The area under the precision–recall curve in the early windows (0.36 and 0.34) barely exceeded the prevalence of 0.31, and the FIGO reading traced the same curve (AUROC 0.53 to 0.63), which the learned models never significantly exceeded. Net benefit was positive only in the final two windows. Information useful for predicting acidaemia is concentrated in the last quarter of an hour before birth.