From data to delivery: a mini-review on the clinical applications and challenges of artificial intelligence in obstetric anesthesia and analgesia

1 Introduction

The rapid expansion of artificial intelligence (AI) across clinical medicine is reshaping how clinicians interpret data and make decisions. Within anesthesiology, machine learning (ML) tools have demonstrated value in perioperative risk stratification, prediction of complications, and pain management (1–8).

In parallel, obstetrics has witnessed a surge in predictive modeling for conditions such as preeclampsia, postpartum hemorrhage (PPH), and maternal cardiac disease (9–11).

However, in obstetric anesthesiology, integration of AI remains fragmented. Current AI models are often retrospective, lack external validation, and face challenges related to data quality, bias, interpretability, and workflow integration. Obstetric anesthesia presents complex clinical environment characterized by dynamic physiology, time-sensitive decision-making, and dual maternal-fetal responsibility. Emerging literature suggests AI may enable risk assessment, predictive modeling, early warning systems and individualized analgesic strategies (12–16). This narrative mini-review aims to evaluate current applications of AI in obstetric anesthesiology and highlight translational challenges.

2 Methods

A literature search was conducted across PubMed and Google Scholar to identify relevant articles on the applications of AI in the practice of obstetric anesthesiology. The search covered the period from database inception to April 2026, with no date restrictions applied but with emphasis on contemporary literature (post-2015). The search strategy combined keywords including “artificial intelligence,” “machine learning,” “obstetrics,” “obstetric anesthesia,” “obstetric analgesia,” “anesthesiology,” “perioperative care,” and “cesarean delivery.” OpenEvidence AI was used as a supplementary tool to identify additional relevant articles on the topic. Inclusion criteria comprised original research, systematic reviews, and case series in English language, while exclusion criteria included animal studies and editorials without primary data. Articles were screened independently by two authors, with disagreements resolved through consensus discussion. No formal risk-of-bias assessment tool was applied due to the narrative design of the review. Nevertheless, studies were informally appraised during synthesis based on methodological rigor, sample size, clarity of AI model description, clinical applicability, and consistency of findings. Duplicate citations retrieved from PubMed and Google Scholar were identified and removed before screening. After title/abstract screening and full-text assessment, approximately 50 relevant studies were included in the final narrative review. Key studies identified through this search, including investigations of postpartum hemorrhage prediction, neuraxial ultrasound guidance, spinal anesthesia-induced hypotension, and post-cesarean pain prediction, are summarized in Supplementary Table S1.

3 Current landscape of AI in anesthesiology and perioperative medicine

Anesthesiology and perioperative medicine is increasingly embedded with AI, with now more than a few models demonstrating strong performance in predicting anesthesia-related complications. These applications can predict hypotension, arrhythmias, hypoxemia, and other adverse events minutes before they occur, thus enabling proactive intervention by the anesthesiologists (1, 3, 17, 18).

Post-induction hypotension is among the most extensively studied applications of AI in perioperative medicine. Multiple machine learning approaches have demonstrated moderate-to-good predictive performance for identifying patients at risk before clinically significant hypotension occurs. Across studies, predictive performance appears relatively consistent regardless of model complexity, with some simpler and more interpretable approaches performing comparably to advanced neural network architectures (19–21).

Real-time intraoperative monitoring represents another potential field of application. AI applied to electroencephalogram data can assess depth of anesthesia, guide anesthetic drug dosing, and potentially mitigate postoperative delirium (18).

AI models for predicting various perioperative complications have achieved impressive performance metrics. For postoperative complications including pneumonia (area under the receiver operating characteristic curve [AUROC] 0.905), acute kidney injury (AUROC 0.848), deep vein thrombosis (AUROC 0.881), pulmonary embolism (AUROC 0.831), and delirium (AUROC 0.762), gradient boosting trees and deep neural networks demonstrated superior performance. Importantly, models using only preoperative data performed nearly as well as those combining preoperative and intraoperative data (22).

Meta-analysis by Mehta et al, suggested that AI-supported interventions are associated with improved perioperative outcomes, although the effect size varies across clinical applications and study designs. Strongest evidence is seen in intraoperative management, particularly in areas such as hemodynamic optimization and early detection of physiological deterioration. However, heterogeneity among studies limit the ability to draw definitive conclusions and thus leave a significant gap between algorithms and real-world integration (1). Intraoperatively, AI applications are increasingly focused on real-time monitoring and decision support. One of the most extensively studied examples is the aforementioned prediction of intraoperative hypotension (4, 19–21). More broadly, AI-driven tools are being explored for automated anesthesia delivery, depth-of-anesthesia monitoring and workflow optimization, supporting a shift toward semi-autonomous or closed-loop systems (7, 8, 18). In a systematic review by Lopes et al, the majority of AI methods tested showed superior performance results compared to traditional methods in all categories of clinical AI application: depth of anesthesia monitoring, image-guided techniques related to anesthesia, prediction of events/risks related to anesthesia and drug administration control (23).

Postoperative care represents another key domain for AI integration, particularly in the prediction and management of pain and opioid use. AI models have been developed to identify patients at risk of severe postoperative pain or prolonged opioid consumption, facilitating more personalized analgesic strategies (5).

Despite all the mentioned advances, a huge gap remains between model development and clinical implementation. Moreover, recent perspectives emphasize the “double-edged sword” of AI in patient safety: while these technologies have the potential to enhance decision-making and reduce human error, they may also introduce new risks such as overreliance, automation bias and system failures (6).

4 AI predicitive modeling in obstetrics

AI is transforming obstetrics through applications in ultrasound imaging, risk prediction for pregnancy complications, fetal monitoring, and clinical decision support systems (24, 25).

AI models show promising results for early screening of spontaneous preterm birth, gestational diabetes mellitus, fetal growth restriction as well as for prediction of risks such as PPH and preeclampsia (26–28).

In recent years, several studies have demonstrated promising performance of AI models for prediction of postpartum hemorrhage (PPH) using routinely available clinical and imaging data (29–31). Wu et al. combined magnetic resonance imaging-based placental radiomics with clinical factors to identify severe PPH in women with suspected placenta accreta spectrum disorders (29). Large cohort studies by Venkatesh et al. and Westcott et al. further showed that ML approaches, particularly gradient boosting models, achieve good predictive discrimination for PPH risk at the time of labor admission (30, 31). These methods may better capture complex interactions among maternal risk factors compared with traditional statistical techniques. Collectively, these studies suggest that gradient boosting approaches demonstrate strong discrimination for PPH prediction, although differences between ML and traditional statistical models are often incremental rather than transformative (9, 30, 31).

Hypertensive disorders of pregnancy affect 5%–8% of pregnancies and encompass a spectrum of conditions. Up until now, many AI models have been developed to predict blood pressure using various surrogate markers, as well as to predict response to antihypertensive treatment, but mostly in non-pregnant population (11). In regards to prediction of preeclampsia, studies show that ML models achieve high predictive performance, with area under curve (AUC) values ranging from 0.86 to 0.97. Despite promising accuracy, limitations include small number of studies, retrospective designs, and need for external validation. Overall, ML approaches show strong potential for early prediction of preeclampsia, but further research is needed before clinical implementation (10).

In the contex of prevention, diagnostics, risk stratification and predicition of complications, another growing subspecialty is the cardio-obstetrics (11). Maternal cardiac disease is estimated to complicate up to 4% of pregnancies globally and is the leading cause of indirect maternal mortality worldwide (32). Besides well known scoring systems such as CARPREG I and II, ZAHARA and modified World Health Organisation (mWHO) classification, few AI models have been developed that are worth mentioning (11). Shara et al. Developed AI model capable to accurately predict conditons such as preeclampsia, myocardial infarction and both cardiomyopathy and cardiac failure many days before a formal diagnosis being entered into electronic health record (11, 33). Jung et al. demonstrated an AI model able to predict left ventricular systolic dysfunction and identify peripartum cardiomyopathy from electrocardiografic (ECG) data. Their model showed high level of accuracy with both 12-lead and single lead ECG data for screening of cardiomyopathy (34).

5 Translation of AI into practice of obstetric anesthesiology

Besides previously discussed potential for prediction of maternal complications such as PPH and preeclampsia, potential application of AI in obstetric anesthesiology primarily lies in the domain of anesthetic and analgesic techniques (12–16, 35). Neuraxial analgesia and anesthesia remain the cornerstone of obstetric anesthetic practice. With increasing prevalence of obesity in obstetric population, optimizing the success of neuraxial techniques has become a growing priority. Strategies to enhance procedural accuracy and outcomes are therefore of significant clinical interest. One emerging area is the application of AI to support ultrasound-assisted neuraxial procedures (13, 36, 37). Chan et al. evaluated ML approach to identify the optimal needle insertion site for spinal anesthesia in obese parturients. In this prospective cohort study of 48 obese parturients undergoing cesarean delivery (CD), a ML-based ultrasound was used to automatically identify spinal landmarks and guide needle insertion.The system demonstrated high first-attempt success rates and strong agreement between automated and clinician-estimated spinal landmark measurements, suggesting potential utility in technically difficult obstetric neuraxial procedures (38).

The role of AI in predicting neuraxial anesthetic dosing has been studied in pregnant women by Wei et al. (13, 39). In a study of 684 term parturients undergoing elective CD, ML algorithm was used to develop a decision-support model for optimizing dosing of intrathecal 0.5% hyperbaric bupivacaine. An appropriate sensory block level (T4–T6) was achieved in 75.44% of patients, with a mean bupivacaine volume of 1.965 mL (95% CI 1.945–1.984). The final model, incorporating vertebral column length and abdominal girth, demonstrated a good predictive performance (39).

In regards to identification of epidural space, the most developed AI application is a hybrid ML system that automatically localizes the needle target in ultrasound images of the spine. This system uses deep neural network architecture with feature augmentation to identify anatomical landmarks of the epidural space. The model achieved high localization accuracy compared with expert sonographer annotations. However, this study did not specifically focus on obstetric patients (40).

While AI-specific trials are scarce, especially in obstetrics, there is robust evidence for conventional preprocedural ultrasound in obstetric neuraxial procedures. A 2021 meta-analysis of 22 trials with 2,462 obstetric patients found that preprocedural ultrasound increased first-pass success rate by 46% (RR 1.46, 95% CI 1.16–1.82), compared to landmark palpation (41). A 2023 network meta-analysis showed that real-time ultrasound guidance had the highest probability of first-pass success (82.8%), followed by ultrasound-assisted (67.1%) and traditional positioning (0.1%) (42).

Given the strong body of evidence supporting ultrasound guidance for neuraxial blocks, alongside the increasing adoption of imaging enhanced with AI, it is reasonable to expect further advancements in this field. A recent 2026 review highlights that AI-assisted ultrasound may enhance the accuracy of neuraxial and regional blocks in obstetric anesthesia, with emerging technologies, such as integrated three-dimensional and AI-guided systems showing potential to improve first-attempt success rates, particularly in patients with obesity. Importantly, these AI-enabled tools are intended to augment, rather than replace clinician expertise (16). Findings across ultrasound-assisted neuraxial studies are generally consistent in demonstrating improved procedural accuracy and first-pass success, particularly in technically challenging obstetric populations (38, 41, 42).

A 2025 study on epidural analgesia for vaginal delivery developed AI model to predict epidural-related maternal fever, with Categorical Boosting (CatBoost) demonstrating the highest performance (AUC 0.94; 95% CI 0.86–0.98) (43).

Frassanito et al. provide a timely and insightful narrative review on the evolving role of AI in anesthesia for CD. The authors effectively synthesize current evidence on how AI-driven tools can enhance perioperative decision-making, patient monitoring, and risk stratification in obstetric anesthesia (35). CD is the most performed inpatient surgery worldwide, with a rising rate, and projected to account for almost one third of all deliveries by 2030 (35, 44, 45).

Besides already discussed potential to enhance administration of neuraxial blocks, multiple other domains within obstetric anesthesiology may benefit from the application of AI.

Hypotension during spinal anesthesia for CD is reported to be up to 82% in some reports (46, 47). In a 2020 study, a single hidden-layer neural network slightly outperformed a discrete-feature approach in predicting hypotension during spinal anesthesia for CD (AUC 0.89 vs. 0.87), using non-invasive hemodynamic data from pulse contour analysis. These findings suggest that arterial stiffness variability may serve as a predictor of hypotension and help guide phenylephrine dosing during spinal anesthesia for CD (48).

Another potential field of AI application is the post-cesarean pain management. Intrathecal morphine is widely recommended and considered the “gold standard” for post-cesarean analgesia, providing prolonged pain relief lasting 20–48 h, with multiple professional societies recommending it's use (49–51). A study by Tan et al. evaluated six ML models to predict significant post-cesarean pain (score ≥3) within 13–24 h, utilizing data from 6,561 patients. Ridge regression demonstrated the highest performance (AUC 0.719) with key predictors including pain scores from the first 12 h (52). In a study by Zhang et al., ML model for postoperative analgesia and adverse reactions (MLPIAAR) managed to successfully predict high-risk factors for inadequate analgesia, as well as postoperative nausea and vomiting (53). These findings are clinically relevant because post-cesarean pain remains highly variable between patients despite standardized analgesic protocols. AI-based prediction models may facilitate individualized multimodal analgesia, optimize opioid-sparing strategies, and identify patients requiring enhanced postoperative monitoring. However, prospective validation and assessment of clinical utility are still lacking.

6 Discussion

AI in obstetric anesthesiology is transitioning from conceptual promise to early-stage clinical applicability. However, most currently available models should still be considered investigational rather than implementation-ready, as prospective multicenter validation, workflow integration, and demonstration of real-world clinical benefit remain limited. Evidence from other fields in anesthesiology demonstrates that AI models can reliably predict perioperative complications with moderate-to-high accuracy (1, 19–22). Importantly, simpler models such as logistic regression have in some cases performed comparably to more complex architectures (19).

In obstetric population, predictive modeling has shown strong performance in identifying patients at risk of major complications such as PPH, preeclampsia and heart disease (9, 10, 30, 31). These developments provide a valuable foundation for anesthesiologists, as many of these complications directly influence peripartal management. However, despite high reported accuracies, most models remain limited by retrospective designs, lack of external validation and variability in datasets (9, 10). Disconnection between model development and clinical implementation mirrors challenges observed in the broader perioperative AI landscape (1, 6).

Within obstetric anesthesiology, AI applications are currently most advanced in the optimization of neuraxial techniques and perioperative risk prediction. Studies exploring AI-assisted ultrasound guidance for spinal anesthesia demonstrate encouraging improvements (38). Similarly, AI-driven dosing models for intrathecal local anesthetics highlight the potential for personalized approach, moving beyond traditional weight- or height-based approaches (39). These applications are particularly relevant given the increasing complexity of the obstetric population, including rising rates of obesity and comorbidities.

Another promising domain is the prediction and management of spinal anesthesia-induced hypotension, a common and clinically significant complication during CD (46–48). Additionally, AI-driven approaches to post-cesarean pain prediction and management suggest a future in which analgesic strategies can be tailored in real time (52, 53). Collectively, these applications illustrate that AI in obstetric anesthesiology extends beyond procedural optimization and hemodynamic prediction toward broader perioperative personalization, including recovery trajectories, analgesic planning, and patient safety.

Despite these advances, several critical barriers must be addressed before widespread clinical adoption can occur. First, data quality and representativeness remain major limitations. Many models are trained on single-center datasets that may not reflect broader populations. Second, algorithmic bias poses a significant risk, particularly in obstetrics, where disparities in maternal outcomes are well documented. Third, the “black box” nature of many AI models limits clinician trust and may hinder integration into high-stakes clinical environments. Finally, workflow integration remains a practical challenge, as AI tools must align with existing clinical systems (6).

Another important consideration is the risk of automation bias and overreliance on AI-generated recommendations. While AI has the potential to enhance patient safety by reducing human error, inappropriate dependence on these systems may introduce new types of errors, particularly in rapidly evolving obstetric scenarios. Therefore, AI should be viewed as a decision-support tool that augments, rather than replaces, clinician expertise.

Lack of integration of fetal outcomes into AI models used in obstetric anesthesiology is another major gap. Future systems should incorporate fetal monitoring data to better reflect the dual-patient nature in obstetrics. Importantly, strong predictive accuracy does not necessarily translate into improved maternal or fetal outcomes, and future studies should evaluate whether AI-assisted interventions meaningfully improve clinical decision-making and patient safety.

Finally, important ethical considerations must be addressed. These include informed consent during labor, where patients may have limited capacity for complex decision-making, and algorithmic bias, particularly given known disparities in maternal outcomes across racial and socioeconomic groups. Additionally, medicolegal responsibility for AI-assisted decisions remains unclear and requires regulatory clarification. Other concerns include data privacy, cybersecurity, and ownership of maternal-fetal health data within large AI training datasets. Regulatory oversight and standardized reporting frameworks such as TRIPOD-AI and CONSORT-AI will likely play an important role in ensuring safe implementation. Furthermore, disparities in access to AI-enabled technologies between high-resource and low-resource healthcare settings may widen existing inequities in maternal care.

This review is subject to several limitations. First, the exclusion of preprints and non-English studies may have introduced language and publication biases, potentially overlooking relevant data that could affect the overall synthesis. Furthermore, the search strategy was limited to published literature, which may skew results toward studies with statistically significant or positive outcomes. These limitations should be considered when interpreting the findings of this review.

7 Conclusion

AI holds significant promise for transforming obstetric anesthesiology by enabling predictive, individualised and proactive care. Current evidence demonstrates strong potential across multiple domains, including risk stratification, optimization of neuraxial techniques, prediction of spinal anesthesia-induced hypotension, and individualized postoperative pain management.

Future research should prioritize prospective multicenter validation of existing models, adherence to new AI reporting standards such as Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD-AI) and Consolidated Standards of Reporting Trials (CONSORT-AI), development of explainable AI (XAI) models suitable for clinical use, as well as integration with real-time electronic health records and monitoring systems.

Bridging the gap between algorithm development and bedside implementation calls for priority for policy, robust prospective validation, improved data quality, enhanced model interpretability, and integration into clinical practice.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fanes.2026.1863454/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Tao Xu, International Peace Maternity and Child Health Hospital, China

Reviewed by: Dereje Bayissa Demissie, Stellenbosch University, South Africa

Elhadi Miskeen, University of Bisha, Saudi Arabia