AI Predicts Dangerous Complication for Moms After Delivery


Ob/gyns working with data scientists at Cedars-Sinai Medical Center in Los Angeles have developed an algorithm that can help predict which patients are at an increased risk for severe morbidity from bleeding after delivery.

The artificial intelligence (AI) model uses data that clinicians routinely collect to generate predictions at admission, during labor, and after delivery.

To train and validate the model, the researchers used retrospective data from 12,807 women, 386 of whom experienced severe morbidity from postpartum hemorrhage (PPH).

The researchers assessed how well the tool identified patients who would experience severe complications by calculating the area under the receiver operating characteristic curve (AUC ROC), where a score of 1 would indicate a perfect ability to distinguish severe cases.

The system's ability to predict complications improved as it considered more information obtained throughout the hospitalization.

At admission, the AUC ROC was 0.7. For the intrapartum period, it was 0.8. Postpartum, it improved to 0.88.

Cecilia B. Leggett, MD, a maternal-fetal medicine fellow at Stanford University, presented the findings last month in a poster at the 2024 Pregnancy Meeting of the Society for Maternal-Fetal Medicine.

Severe Complications

PPH complicates about a quarter of deliveries, Leggett said. But physicians lack a way to reliably predict which patients will experience severe complications from the condition, such as unscheduled hysterectomy, uterine artery embolization, ICU admission, massive transfusions, return to the operating room, or death.

"Typically, people who are pregnant are younger, healthier people," Leggett told Medscape Medical News. "Their bodies are resilient and can accommodate a lot of blood loss without having any severe consequence or morbidity. But we are not as good at predicting who is going to need extra support, who is going to end up needing to go back for an emergency hysterectomy because the bleeding just will not stop, or who is going to need to have multiple blood transfusions to stabilize them."

To assess whether AI could help identify such patients, Leggett and her colleagues at Cedars-Sinai conducted their study using an automated machine learning platform and time-series engineering, meaning the system analyzed data in a way that recognized it was examining a process occurring over time rather than assessing all data at once.

Key Features

They found that at admission, key predictive features for the AI model included age, type of insurance, and the Social Vulnerability Index, which is based on the demographics of a person's ZIP code, such as socioeconomic status, household characteristics, race and ethnicity, and housing types. Intrapartum, the duration of labor, average diastolic blood pressure, and crossmatch orders were important factors. After delivery, important features included the type of anesthesia and maximum heart rate.

The researchers said they intend to publish complete results of the study in a peer-reviewed journal.

It could be that AI will be able to pick up subtle but important changes in vital signs, Leggett said.

"Your blood pressure going down or your heart rate creeping up can predict that your body is starting to slowly not compensate as well for the blood loss you are experiencing," she said.

Further studies are needed to see if AI can help predict hemorrhages in real time and allow clinicians to intervene to improve outcomes, she added.

Original Article