Extending the training window from the final hour to the full duration of labor and adding time from labor onset let a model identify 40.3% of hypoxic-ischemic encephalopathy cases at least 3 hours before delivery, an 8.4-point gain over the last-hour classifier

Progressively extending the training window from the final hour before delivery to the full duration of labor improves early identification of HIE and plateaus at 18 hours. Most prior AI models were trained only on the final hour before delivery, a window identifiable only retrospectively; this study extended the training window backward through labor, directly testing whether earlier evolving signs are learnable. A large cohort of 174,186 deliveries with 2,636 severe acidosis cases and 304 clinically validated HIE cases, evaluated continuously at a fixed 15% false positive rate, with the reported correspondence between window extension and performance and the 18-hour plateau.

Adding time from labor onset (TLO) as a feature yields the highest early-warning performance: 40.3% of HIE cases identified at least 3 hours and 27.3% at least 6 hours before delivery. Relative to the last-hour classifier, these represent absolute gains of 8.4% and 10.7%, indicating that time-aware modeling captures the temporal evolution of HIE patterns beyond simply lengthening the window. Results are reported from continuous evaluation at a fixed 15% false positive rate, with explicit absolute gains over the last-hour classifier.

55.7% of HIE cases can be identified at least 40 minutes before delivery, shifting the model from end-of-labor diagnosis toward early-warning decision support. This result makes the actionable lead time concrete and points to a clinical decision-support role rather than retrospective diagnosis. Derived from the same cohort under continuous evaluation at a fixed 15% false positive rate, with the proportion reported in the abstract.

The result applies to intrapartum CTG monitoring for deliveries at gestational age ≥35 weeks within Kaiser Permanente Northern California, aiming to give clinical teams early-warning signals hours before delivery at a fixed 15% false positive rate, thereby shifting intrapartum AI from end-of-labor diagnosis toward early-warning decision support. It fits obstetric monitoring settings with continuous CTG records and labor timing information, and offers a replicable approach for testing time-aware modeling in broader populations and additional health systems.

The reading scope here is incomplete, covering only the abstract, competing interest statement, ethics declarations, and data availability statement, without the main text methods, figures, or result tables, so feature definitions, model calibration, subgroup performance, and statistical uncertainty cannot be checked. Readers should still watch how time-aware models perform under different labor management routines and different CTG acquisition quality, and how early warnings would connect with existing intervention pathways in real clinical workflows. The data are not publicly available and cannot leave KPNC firewalls, so independent external replication requires a proposal and a data use agreement, which is a condition future validation will face.