Multimodal Causal Machine Learning for Fetal Asphyxia Risk Prediction from Imperfect Monitoring Signals

Abstract Cardiotocography (CTG) is the cornerstone of intrapartum surveillance, yet its predictive value for perinatal asphyxia remains limited. Modern monitors record fetal heart rate (FHR) via Doppler ultrasound or intrapartum fetal electrocardiography (ECG), the latter usable only after cervical dilatation. We analysed 52,757 labours ≥36 weeks’ gestation across three hospitals—two in Australia and one in Europe—treating mode-specific FHR dropout (zero-value samples signifying missing valid biosignal) as an informative exposure rather than benign artifact. High dropout (>30% of the CTG recording available for analysis) was evaluated using doubly robust causal machine learning and survival analysis, incorporating inverse-probability weighting, targeted maximum-likelihood estimation (TMLE), and weighted Cox models. In the training cohort (n = 36,792), high ECG-mode dropout tripled the odds of perinatal asphyxia (OR 3.14, 95% CI: 2.20–4.48) and quadrupled the hazard (HR 4.76, 95% CI: 2.98–7.62) compared to ultrasound-mode low dropout. In crude analyses, this corresponded to an absolute risk increase of 5.0% (1 in 20 births). TMLE-adjusted models confirmed a 2.27% absolute risk increase—equivalent to one additional case per 44 labor episodes (95% CI: 28–83). External validation confirmed consistent associations in European (OR 4.02) and Australian (OR 2.22) cohorts. “Imperfect” ECG signals thus offer biologically plausible and trustworthy inputs for standalone or ensemble artificial intelligence models, enabling transparent, mode-aware decision support in real-time fetal monitoring.