AI Is Racing Into the Delivery Room, but Most Models Have Never Been Tested Outside Their Own Data

Artificial intelligence has swept through nearly every corner of modern medicine, and now it is knocking on the doors of labor and delivery wards. A new scoping review published in Frontiers in Anesthesiology offers the most comprehensive picture yet of how AI and machine learning are being deployed across obstetric anesthesia, the subspecialty responsible for keeping mothers safe and comfortable through labor, cesarean delivery, and the most dangerous complications of childbirth. The verdict is striking: the field is exploding with promise, but almost none of that promise has been tested in the real world.

Researchers at Weill Cornell Medicine searched PubMed and Embase for studies published between January 2015 and December 2025, combining clinical terms for obstetric anesthesia and delivery with methodological terms for artificial intelligence and machine learning. From 3,135 records identified, two reviewers with consensus adjudication by board-certified anesthesiologists narrowed the field to 172 studies that met inclusion criteria: 154 original research studies, 17 systematic or narrative reviews, and a single case report. Twelve additional records, including nine registered clinical trials and three workforce-focused studies, were analyzed separately. The timing tells its own story. Roughly 69 percent of the included publications appeared after 2022, and more than half were published in 2024 or 2025 alone, a pace that mirrors the public release of large language models and the broader AI boom in medicine.

The geographic spread was equally revealing. The United States contributed 40 studies and China 35, followed by Israel, Italy, India, Singapore, and Spain, with eighteen multinational collaborations rounding out the picture. When the researchers classified the studies by machine learning approach, classification tasks dominated, followed by regression, neural networks, and large language models. Unsupervised learning, reinforcement learning, and dedicated radiomics pipelines appeared only rarely. The work clustered into distinct domains: maternal risk and complications prediction led with 61 studies, delivery mode prediction followed with 43, patient education and conversational AI claimed 15, fetal heart rate tracing analysis 12, imaging and radiomics 8, and provider education and simulation just 2.

The single largest domain, maternal risk and complications prediction, targets some of the most feared events in obstetrics. Postpartum hemorrhage accounts for 11 percent of U.S. maternal mortality and remains a leading preventable cause of maternal death, making accurate risk assessment a matter of life and death. Machine learning models predicting hemorrhage and transfusion have reported widely varying performance, with areas under the curve ranging from 0.67 to 0.99. Large language models have even been used to extract risk factors from prenatal notes and electronic health records, identifying features like multiple gestation and prior cesarean delivery. But the review uncovered a sobering technical problem: differing hemorrhage definitions across health systems can dramatically degrade model performance. One model’s C-statistic fell from 0.93 to 0.57 when applied to a contemporary cohort rather than an early-2000s cohort, a collapse that underscores how fragile these tools can be outside their training environment.

Placenta accreta spectrum, a life-threatening condition in which the placenta invades the uterine wall, illustrates both the power and the limits of current approaches. Although its prevalence is low at roughly 0.17 percent, it drives severe peripartum morbidity, with pooled incidences of 52.2 percent for hysterectomy and 46.9 percent for transfusion among affected women. Deep learning and MRI-based radiomics have shown genuine promise for earlier recognition, with one 2025 study using 15 radiomic features reaching an area under the curve of 0.95 for predicting the condition. Yet the same study found that random forest and k-nearest neighbor models using identical variables achieved AUCs of only 0.704 to 0.863, a reminder that algorithm choice matters enormously and that headline numbers can mask substantial intra-study variation.

Hypertensive disorders of pregnancy represent another major target. Pre-eclampsia affects 555 per 10,000 deliveries, and hypertensive disorders account for 6.8 percent of pregnancy-related deaths. A 2024 pilot study used routine biomarkers and clinical characteristics to predict severe pre-eclampsia requiring intensive care, achieving an AUC of 0.91 with an XGBoost model in an 81-patient cohort. Aspartate aminotransferase, uric acid, and body mass index emerged as the strongest indicators of severe disease. Meanwhile, peripartum cardiomyopathy, which complicates roughly 1 in 2,000 pregnancies worldwide and carries 76-fold higher odds of cardiogenic shock, has become the focus of one of the field’s most compelling real-world tests: a recent trial found that an AI-guided digital stethoscope doubled the diagnosis of pregnancy-related cardiomyopathy compared with usual care, with a number needed to screen of 47 to identify one new case of perinatal left ventricular systolic dysfunction.

Beyond risk prediction, AI is being applied to the mechanics of birth itself. Approximately 30 percent of the reviewed articles addressed prediction of delivery mode and timing, spanning emergency cesarean, successful trial of labor after cesarean, and vaginal versus cesarean delivery. Because about one in three U.S. births occurs by cesarean and most of those are unplanned, carrying higher complication rates than planned surgery, accurate forecasting could transform labor ward planning. Machine learning models for intrapartum delivery mode prediction showed variable performance, with AUCs from 0.69 to 0.932, and one study found a k-nearest neighbor model predicting vaginal delivery four hours after admission with an AUC of 0.812, outperforming traditional logistic regression. For trial of labor after cesarean, three machine learning models significantly outperformed the established Maternal Fetal Medicine Unit calculator, though the margins were narrow and external validation remained limited. The Artificial Intelligence Dystocia Algorithm, designed to assess transverse fetal head position and cesarean risk, reached 95.5 percent accuracy on just 135 cases, a tantalizing but preliminary result.

Fetal heart rate tracing analysis, the backbone of intrapartum monitoring and the primary indication behind 27 percent of cesarean deliveries in the United States, has attracted its own wave of machine learning. Deep learning and neural network models for cardiotocography analysis have reported strong performance, with AUCs between 0.87 and 0.987, and one model categorized delivery mode from fetal heart rate signals with an AUC of 0.967. Imaging and radiomics studies clustered around two themes: MRI and ultrasound models for placenta accreta, and procedural ultrasound for anesthesia itself, including software for lumbar level identification and a neural network for real-time cricothyrotomy landmark identification using wireless handheld ultrasound, relevant to failed-airway preparation on labor and delivery. Notably, fusion models combining radiomic features with clinical and laboratory data consistently outperformed either data type alone, suggesting the most useful clinical applications will not rely on imaging in isolation.

The most rapidly emerging domain is patient education and conversational AI, with 15 of 16 studies published in 2024 or 2025. Multiple evaluations of ChatGPT and other chatbots found that responses to common questions about labor epidural analgesia were generally accurate and appropriately framed but often exceeded the recommended sixth-to-eighth grade reading level, and occasionally included factual errors or omitted clinically important caveats. A 2025 evaluation of ChatGPT-5 in obstetric and gynecologic emergencies highlighted reliability gaps that persist even in the newest models. Innovative tools like the Mothers’ Assistant for Labor analgesia, an interactive avatar-based education platform, point toward more sophisticated applications, but no included study prospectively evaluated whether large language model-supported education actually changed patient knowledge, anxiety, or satisfaction compared with standard counseling. On the provider side, a 2025 study comparing four large language models on complex anesthesia decision-making across 30 high-risk scenarios found that all four generated structured plans, with DeepSeek-R1 performing best in the obstetric subgroup, though the study was limited by small sample size and no outcome-based validation.

Perhaps the most important finding of the entire review is what is missing. Of the 154 original research studies, only 22, or 14 percent, reported external validation in an independent dataset. Prospective evaluations were rare, and almost none described integration into clinical workflow, meaning that even the most accurate models have rarely been tested in the hands of actual anesthesiologists caring for actual patients. The nine registered clinical trials identified cover narrow topics such as labor analgesia sequelae, lung ultrasound interpretation, and automated spinal landmark identification, a pipeline that looks thin relative to the sheer volume of model development activity. The authors also flag serious ethical concerns: algorithmic biases arising from unbalanced training datasets, historical health inequities, race-based clinical adjustments, and opaque black-box methodology could exacerbate the very racial and ethnic disparities in maternal outcomes that persist in the United States, where maternal mortality reached 32.6 deaths per 100,000 live births in 2022. The path forward, the researchers argue, requires a deliberate shift from building models to validating them externally, testing them prospectively, and embedding them responsibly into clinical workflows, with obstetric anesthesiologists serving as designers, evaluators, and stewards of the tools that will increasingly shape their practice. Until then, the extraordinary numbers on paper remain exactly that: numbers on paper.

Subject of Research: Applications of artificial intelligence and machine learning in obstetric anesthesia, including maternal risk prediction, delivery mode forecasting, fetal monitoring, and patient education

Article Title: Artificial intelligence and obstetric anesthesia: a scoping review

Article References: Givens, M. F., Rodriguez Caceres, E., Jotwani, R., Rubin, J. E., Goldstein, P. A., & White, R. S. (2026). Artificial intelligence and obstetric anesthesia: a scoping review. Frontiers in Anesthesiology, 5, Article 1896595. https://doi.org/10.3389/fanes.2026.1896595

Image Credits: AI Generated

DOI: 10.3389/fanes.2026.1896595

Keywords: artificial intelligence, machine learning, obstetric anesthesia, maternal mortality, postpartum hemorrhage, placenta accreta, pre-eclampsia, fetal heart rate monitoring, large language models, external validation, clinical decision support, scoping review