産科および胎児モニタリングにおける AI

Cardiotocography (CTG) records the fetal heart rate and uterine contractions. Clinicians look at baseline rate, variability, accelerations and decelerations. The problem is that experts often disagree about the same trace, and CTG has a high false-positive rate, which contributes to interventions such as emergency caesarean sections. Computerized analysis is not new. The Dawes-Redman criteria, developed at Oxford, have long been used to analyze antenatal CTGs with rule-based software. For labor, the UK INFANT trial, published in 2017 with about 46,000 women, tested decision-support software that alerted staff to concerning CTG patterns. It found no improvement in poor neonatal outcomes compared with usual care. That result is a caution against assuming that automated interpretation improves care. Newer research, including machine learning work at Oxford, tries to learn patterns from large archives of traces linked to outcomes rather than encoding existing rules. Ultrasound assistance is more visibly successful. Manufacturers have added features that recognize standard views, place measurement calipers and check image quality. Researchers have also shown that AI can estimate gestational age and fetal position from blind sweeps performed by people without sonography training, which could help where trained sonographers are scarce. Preeclampsia risk prediction is often described as AI, but the best-established approach, from the Fetal Medicine Foundation, is a statistical competing-risks model combining maternal factors, mean arterial pressure, uterine artery pulsatility index and placental growth factor. The ASPRE trial showed that screening with this model and giving aspirin to high-risk women substantially reduced preterm preeclampsia. Machine learning models using health records are being studied as additions. A common misconception is that AI can already tell when a baby is in distress better than a midwife. Current evidence does not show that for labor CTG.