Summary
A medRxiv preprint reports that a time-aware machine-learning model identified some cases of hypoxic-ischaemic encephalopathy (HIE) hours before delivery using labour monitoring data. The analysis covered 174,186 deliveries and evaluated performance at a fixed 15% false-positive rate.
A medRxiv preprint analysing 174,186 deliveries reports that a machine-learning model using the timing of labour identified 40.3% of hypoxic-ischaemic encephalopathy (HIE) cases at least three hours before delivery, and 27.3% at least six hours before delivery. Those results were absolute gains of 8.4 and 10.7 percentage points, respectively, over a model trained only on the final hour. Both were evaluated at a fixed 15% false-positive rate.
HIE is brain dysfunction associated with inadequate oxygen and blood flow around birth. The study examined cardiotocography (CTG), which monitors fetal heart rate and uterine contractions during labour. The authors describe CTG as standard intrapartum monitoring, while noting that its poor specificity can prompt unnecessary interventions.
How the time-aware model was evaluated
The researchers analysed deliveries at 35 or more weeks’ gestation in Kaiser Permanente Northern California. They extracted 40 CTG features from 20-minute periods and trained Random Forest classifiers to predict severe acidosis (2,636 cases) and clinically validated HIE (304 cases).
Rather than training only on the final hour before birth, they progressively expanded the training window, up to the full duration of labour. They also added time from labour onset as a feature. The study authors’ hypothesis was that a last-hour-only approach would favour patterns close to delivery, while a model trained across labour could learn how warning patterns develop over time.
Early identification of HIE cases improved as the training window grew, with performance plateauing at 18 hours. Adding time from labour onset produced the strongest early-warning results reported: 40.3% of HIE cases were identified at least three hours before birth and 27.3% at least six hours before. The researchers also report that 55.7% of HIE cases were identified at least 40 minutes before delivery.
What the results could mean for monitoring
The study points to a design consideration for AI-based fetal monitoring: the time context of a CTG segment may matter, not just the recorded pattern itself. A model that recognises how signals evolve during labour could potentially provide more time for clinical assessment than one focused on the period immediately before birth.
These percentages describe model identification of HIE cases at the study’s chosen operating point; they are not evidence that using the model’s warnings improves outcomes. The work is a preprint analysis of delivery records from one health system, rather than a reported prospective test of alerts in clinical care. Two authors were employed by PeriGen Inc.; the other authors declared no competing interests.