Abstract
Abstract
Cardiotocography (CTG) is the standard of care for intrapartum fetal monitoring, yet its poor specificity drives unnecessary intervention without reducing hypoxic-ischemic encephalopathy (HIE). Most artificial intelligence (AI) models are trained exclusively on the final hour before delivery, a window only identifiable retrospectively. We hypothesized that this framework prevents models from learning the early evolving signs of HIE. We analyzed 174, 186 deliveries, with gestational age GA ≥ 35 weeks, from Kaiser Permanente Northern California. Forty CTG features were extracted from 20-minute epochs to train Random Forest classifiers targeting severe acidosis (
N
= 2, 636) and clinically validated HIE (
N
= 304). The training window was progressively extended from the final hour to the full duration of labor, and time from labor onset (TLO) was added as a feature. Models were evaluated continuously throughout labor at a fixed 15% false positive rate. Our results showed that the early identification of fetuses at risk of HIE improved as the training window was extended, plateauing at 18 hours. Adding TLO gave the highest early-warning performance: 40.3% of HIE cases identified at least 3 hours and 27.3% at least 6 hours before delivery, absolute gains of 8.4% and 10.7% over the last-hour classifier. Finally, 55.7% of HIE cases were identified at least 40 minutes before delivery. In conclusion, classifiers trained only on the last hour may identify HIE patterns that occur near delivery; time-aware models trained across labor map the temporal evolution of these patterns and provide actionable early warnings. Thus, AI systems for intrapartum monitoring must account for the temporal evolution of labor to shift from end-of-labor diagnosis to early-warning decision support.
Author summary
During labor, fetal monitoring relies on cardiotocography (CTG) to assess fetal well-being and their risk of developing hypoxic-ischemic encephalopathy (HIE). However, current clinical guidelines suffer from poor specificity, driving unnecessary Caesarean deliveries without decreasing HIE rates. Moreover, existing artificial intelligence (AI) systems are trained exclusively on the final hour of labor, which is only available retrospectively and assume signal stationarity. The fetal heart rate (FHR) evolves continuously through labor, rendering the assumption of stationarity reductive. This biased assumption also limits the early-warning capabilities of such AI systems. Here, we analyzed 174,186 deliveries to train random-forest (RF) models that target severe acidosis and HIE. We found that expanding the training window beyond the final hour and indexing the time of labor onset into the classification features brought significant advantage. This joint approach resulted in over 10.7% more HIE cases being detected at six hours before delivery when compared to the conventional last-hour approach. Such achievement carries important clinical implications. For preventive interventions to be appropriate and efficient, clinicians must be able to stratify risk as early as possible. Thus, the goal of the intrapartum decision support systems should not be limited to accurate detections at the time of delivery but also to optimize for the lead time of the detections. Our approach demonstrated that to significantly improve the early identification of fetuses at risk of HIE, novel AI systems should be exposed to longer labor training windows and be able to track the time from labor onset to provide temporal context to their decisions.