Dive Brief:
- Machine learning algorithms may support clinical decisions during labor and improve predictions of neonatal outcomes, according to a retrospective study published in PLOS ONE.
- Mayo Clinic researchers used the Consortium on Safe Labor database to create labor prediction models designed to improve on established approaches for managing childbirth.
- On admission, models used to predict unfavorable labor outcomes achieved a sensitivity of 0.69 and specificity of 0.68. However, those figures increased as cervical dilation advanced and more data on labor progression was fed into the models.
Dive Insight:
The World Health Organization sought to standardize labor care by documenting measurements such as cervical dilation and fetal heart rate that healthcare professionals can refer to detect delays or deviations from the norm. However, a Cochrane review of three clinical trials found the use of the tool had no effect on outcomes such as the rate of unindicated cesarean delivery.
To try to improve on the WHO tool, researchers looked at 66,586 delivery episodes in a database of pregnancy and labor characteristics from 12 U.S. medical centers. The goal was to create a model that improves prediction of whether babies should be delivered vaginally or by Cesarean section, and the risk of postpartum hemorrhage and other adverse outcomes.
”A major conundrum every obstetrician faces in managing women in labor is weighing maternal and neonatal risks of delayed intervention against risks of unindicated Cesarean delivery,” the study’s authors wrote.
“Once validated with further research, we believe the algorithm will work in real time, meaning every input of new data during an expectant woman's labor automatically recalculate the risk of adverse outcome,” Abimbola Famuyide, a Mayo Clinic obstetrician and senior author of the study, said in a statement. “This may help reduce the rate of cesarean delivery, and maternal and neonatal complications.”
Periodic cervical examinations to gauge the progress of labor are an essential part of the delivery process, helping obstetricians predict the likelihood of a vaginal delivery in a specified period of time. But cervical dilation in labor varies from person to person, and the AI algorithms can help physicians decide more quickly which path to take.
The baseline model used variables known at the time of admission. To improve sensitivity and specificity, the researchers fed dynamic variables determined by pelvic examination into the prediction models. The data generated at a cervical dilation of 4 cm increased the sensitivity and specificity to 0.70 and 0.72, respectively. Once 10 cm dilation was reached, the figures rose to 0.79 and 0.84,
“Utilization of machine-learning–based algorithms may provide a dynamic, cumulative, and individualized model for prediction of outcomes of vaginal delivery and facilitation of intrapartum decision making,” the researchers concluded. “However, further prospective studies are warranted to assess outcomes of implementation of these models in labor units.”
In an emailed statement, Famuyide wrote that the algorithm is being validated using Mayo Clinic patients and then will be deployed prospectively in the practice to evaluate its impact on physicians’ and midwives’ decision making on women in labor.
The complexity of machine learning algorithms means the results of the study cannot be converted to a printed labor chart, like those used in healthcare today. However, the researchers are working on a digital application to facilitate clinical use of the model.
Famuyide wrote that they envisage the algorithm being embedded in electronic medical records where it will generate labor risk scores on admission and at every cervical dilation evaluation in labor over time.
