Artificial intelligence and decision support tools in obstetric anaesthesia: opportunities, challenges, future

Introduction

Alan Turing's original question “Can Machines Think?” on machinery and intelligence furthered research in artificial intelligence (AI) that makes up today's landscape of machine learning (1). Over the past decades, AI has made great strides in replicating intelligence within human-defined constraints. The four components of AI commonly used in medicine are machine learning (ML), neural networks (NN), deep learning (DL) and generative AI (2). ML utilises data to identify specific patterns with reinforcement learning to make decisions (3). While, NN use multiple layers of calculations to interpret and draw conclusions (4). DL is a subset of NN characterized by multiple neural layers that learn by abstracting representations of data to accurately represent complex relationships. Recently, we have witnessed the rise of generative AI that gave rise to chatbots powered by large language models (LLM) like ChatGPT, Claude and Gemini. Generative AI is enabled by deep learning with extended capability of producing new content including human like texts, images and videos (5).

While AI has made advances into medicine, this article focuses on the impact of AI on obstetric anaesthesia practice. In the antenatal setting, AI chatbots may enhance patient understanding of labour pain relief options and assist clinicians to perform preoperative risk stratification. While intraoperatively, AI may improve analysis of real time data and optimize neuraxial anaesthetic doses to ensure patient safety. AI also has the potential to streamline workflow and enhance healthcare resource allocation. We have summarised the usage of AI technologies and its application in Figure 1 and its evidence level in Figure 2. We also aim to identify potential challenges and define the future solutions for AI integration into clinical practice.

Methods

A narrative review of the literature was conducted to summarize evidence regarding artificial intelligence use within obstetric anaesthesia. Electronic databases including PubMed, Embase, and Cochrane were searched for relevant publications from inception to December 2025. Search terms included combinations keywords related to artificial intelligence (“artificial intelligence”, “machine learning”, “deep learning”, “neural networks”, “large language models”, and “generative AI”) and obstetric anaesthesia (“obstetric anesthesia” OR “obstetric anaesthesia”, “labour analgesia” OR “labor analgesia”). Reference lists of relevant articles were also manually screened to identify additional studies. We limited the search to published studies in English and excluded case reports, abstracts and letters.

The identified literature was reviewed and organized into themes including AI uses in the outpatient, intraoperative decision support tools, improving maternal care, challenges and future directions.

AI uses in the outpatient setting

Figure 1 shows the uses of LLM and DL in the preoperative setting of patient counselling, airway risk prediction and risk stratification within obstetric anaesthesia. LLMs like ChatGPT and Claude are trained on large data to analyse and generate human-like responses (6). Retrieval-augmented generation (RAG) enhances this process by searching for specific information amidst large amounts of text before generating a response making it highly suitable for pre-operative management (6), which often contains many guidelines and protocols. Ke et al. developed an AI chatbot on a secured government platform and evaluated the responses’ accuracy and risk of hallucinations within a clinical setting (7). In total, 240 clinical iterations of the AI chatbot’s responses were evaluated with an accuracy of 97.9% (235/240) compared against institutional guidelines and expert consensus. There were minimal 0.4% (1/240) hallucinations and 0.8% (2/240) deviations from guidelines (7). As seen in Figure 2, this was only conducted at a single institution using local guidelines limiting its generalisability. Comparing LLM’s accuracy against expert consensus and not prospective patient outcomes may introduce circular validation whereby the model replicates the expert’s opinion without improving patient care. Given the adverse consequences of hallucinations though minimal (0.4%), clinician oversight remains essential to override erroneous judgement. Therefore, LLM shows potential as a clinical support tool in preoperative obstetric care but further collaborative multi centre studies benchmarked against patient outcomes with iterative feedback will be required to improve the model’s robustness.

Currently, about 6% of all caesarean sections still require general anaesthesia with tracheal intubation bearing the risk of a difficult intubation at 2%. (8, 9) It is well known that risk factors like obesity, edematous oropharyngeal soft tissue are associated with a physiologically and anatomically difficult airway (10, 11). Xia et al. utilised deep learning facial analysis on 5849 healthy predominantly Han Chinese patients of whom 514 had Cormack-Lehane grades III or IV (12). Their facial model using the Light Gradient Boosting Machine algorithm showed the highest area under the curve (95%CI) of 0.779 as compared to bedside examination and other machine algorithms (12). Interestingly, their model performed better than traditional clinical predictors; the model is able to analyse facial features that exceeds the capacity of manual multivariable assessments (13, 14). Furthermore, standard bedside airway examination have comparatively low sensitivity and specificity (15). There are several limitations in this retrospective study which comprised of mainly Han Chinese limiting generalisability. An area under the curve of 0.779 falls below the threshold for clinical adoption and since pregnant patients were not included; the differences in airway anatomy may further lower the model performance. By using cameras, predictive algorithms like the facial model by Xia, AI can be used to develop applications to determine airway management preoperatively (14). Currently, there is limited data on deep learning models on the obstetric airway. Further prospective studies are needed to validate these predictive variables before clinical utility can be established in this high-risk population.

AI chatbots in healthcare have contributed to triaging of symptoms (16), diagnosis (17) and facilitating patient communications (18). Responses provided by AI chatbots like Gemini and ChatGPT are accurate and easily readable when tasked with answering common questions on labour epidural (19, 20). However, notable differences exist between various AI platforms. Gemini’s sentiment analysis (tone and formality) is more serious compared with ChatGPT (20). These differences likely reflect different training datasets, models and objectives that underlie each chatbot. The tone differences though nuanced can affect users differently- one that is too casual may sound flippant while one that sounds too clinical may reduce comprehension. As AI chatbots may provide different responses, comparing the quality and consistency between different AI chatbots is important. Lee et al reported the readability, quality, and accuracy of the responses between ChatGPT 3.5 and Gemini on 20 frequently asked questions about labour epidurals (19) Overall, Gemini had better readability, actionability, and detailed answers yet conveyed similar accuracy as compared to ChatGPT 3.5 (19). The use of ChatGPT 3.5 is now superseded with newer models within a short span of 2 years which limits the relevance of the results today.

A further concern arises when AI chatbots are used in languages apart from English. Gonzalez et al evaluated the accuracy of ChatGPT 3.5 (older version) responses to common questions (different set of questions from the previous studies) on the topic of labour epidurals in both Spanish and English (21). Three areas were evaluated: impact on labour and delivery, maternal and fetal harm. The Spanish responses had significantly lower accuracy scores compared to English, especially in questions regarding the effect of epidural analgesia on labour and delivery (21). The inaccuracies could stem from unequal representation in training data and the differences in language nuances that may not fully depict the complexities of medical information (21). This perpetuates healthcare inequality as non-English speaking patients who already face barriers to healthcare may also be receiving inferior AI content. Future research should move to incorporate patient centred outcomes like satisfaction and comprehension of chatbot responses to assess clinical utility.

Within the larger context of women’s health, a systematic review studying AI chatbots and health outcomes in women health’s showed significant effects in reducing anxiety (p = 0.017) among women who are conceiving or diagnosed with cancer (22). Similarly, a randomised controlled trial among 100 patients who interacted with ChatGPT preoperatively as compared to receiving information from anaesthesia nurses experienced lesser anxiety measured by the overall STAI score (p = 0.015) (18). This RCT was limited by the relatively small sample size with no possibility of blinding potentially introducing performance bias. While there are potential benefits in reducing anxiety in surgical or oncology patients, no study to date have specifically examined their application in reducing anxiety among parturients. As AI chatbots can gather information extensively in real time and respond in a manner akin to humans, patients’ immediate concerns in the obstetric settings can be addressed. This may be helpful as a preoperative tool to assist in risk counselling and to allay patients’ concerns.

Intraoperative: clinical decision support, neuraxial anaesthesia insertion and dosing assistance

There is moderate evidence that the usage of ultrasound in neuraxial anaesthesia is associated with higher first pass success in obese parturients (Figure 2). As obesity becomes more prevalent, spinal anaesthesia becomes technically more challenging due to excess adipose tissue preventing palpation of bony landmarks and the increased depth to subarachnoid space. Ultrasound assisted landmarking to locate the intrathecal space pre-procedurally are becoming more important to minimise failure rates. A failed neuraxial block may necessitate conversion to general anaesthesia which carries the risk of aspiration and significantly increased risk of difficult airway (8). Chan et al. developed a machine learning algorithm to determine spinal needle insertion point by training the machine to identify the posterior dura complex in the transverse view and the L3/4 interspinous space in the longitudinal view using automated ultrasound-guided images (23). Overall, there was good correlation between the recorded depth of skin to posterior complex and clinical measured depth (r = 0.915) with a 79.1% first attempt success rate for spinal anaesthesia among obese parturients (23). While promising, this study was conducted at a single centre without a control group in which conventional ultrasound or landmark technique was done which makes it difficult to quantify the added benefit. Further studies can evaluate this application within education and training to provide real time feedback to trainees on anatomical identification in obese parturients.

Traditionally to determine the epidural space, the anaesthetist relied on manual palpation and loss of resistance techniques (air or water) with a variable failure rate ranging from 6% to 20% (24). Of late, computer vision and deep learning allow the identification of the small epidural space precisely. Pesteie et al. described a hybrid machine learning system that performs pixel level classification of ultrasound images from human subjects to identify the epidural space (25). On comparison with expert annotation on standard 2-dimensional ultrasound systems, the model had a localization error of 1.7 mm and 0.8 mm in the lateral and vertical planes which are within acceptable limits (25). In another study, Lin et al. developed an intelligent recognition system (i-RS), based on linear discriminant analysis of data from optical reflectance spectroscopy to aid identification of the ligamentum flavum and epidural space during needle insertion (26). The sensitivity and specificity of i-RS for identifying the correct anatomical structures was similar to an expert anaesthetist with a sensitivity and specificity of 80% (26). Pesteie's study was validated against expert annotation of ultrasound images rather than prospective clinical patients which limits its application. A localisation error within acceptable limits on an image may not translate into accurate needle placement under clinical conditions (labour pain, movement and tissue deformation) during the neuraxial procedure. Moreover, neither study has undergone prospective randomised evaluation in the obstetric population. However, translating these systems into clinical workflows and integration with existing equipment remains challenging due to time sensitivity in the obstetric population. Prospective studies in a clinical setting will be required to determine if the improved accuracy of identifying the epidural space may reduce complications related to accidental dural puncture.

As seen in Figure 1, within the intraoperative setting AI has been utilised to determine the optimal neuraxial dosing for labour analgesia and Caesarean section. While the labour experience is complex and multifaceted, epidural analgesia is the gold standard for labour pain relief (27). There are numerous ways to administer the epidural infusion (like continuous infusion, automated bolus or patient controlled epidural analgesia) (28), but they remain limited in individualising pain relief as these regimes depend on predetermined doses. The first computer integrated patient controlled epidural analgesia (CI- PCEA) automatically titrates the background infusion based on the individual analgesia requirements that measures the demand bolus that the parturient needs (29). Overall, the CI-PCEA had higher maternal satisfaction scores 93 vs. 86 out of 100, p = 0.04 and does not increase the use of local anaesthetic when compared with PCEA (29). This shows that individualised dosing by computer integration can improve patient outcomes without compromising safety. While this system does not incorporate AI, the automated feedback that is driven by patient's analgesic demands establishes a foundation in which future AI driven algorithms can be built upon.

Leong et al. also developed a vital sign controlled, patient assisted intravenous analgesia for administration of remifentanil during labour (30). This interactive feedback system evaluated real time maternal oxygen saturation and heart rate to monitor for adverse effects like respiratory depression. The authors commented that the algorithm was not as responsive due to the factors such as the slight delay in the pharmacokinetics of remifentanil coinciding with the dynamic nature of pain and the lack of integration of cardiotocography (CTG) (30). Rather than relying on rigid algorithms, AI may be more suited to predict contraction pain from analysing CTG patterns and trigger dosing. However, the absence of CTG integration within the model is a gap in fetal monitoring during labour analgesia. Therefore, this system remains at the stage of experimental proof of concept.

Studies suggested the feasibility for implementing AI to predict breakthrough pain. Breakthrough pain during labour affects 15%–55% of parturients and is defined as an unplanned epidural supplementation from patient reported pain despite a functioning epidural (31). Breakthrough pain is clinically important as it negatively impacts the childbirth experience and requires anaesthetist intervention. Predicting which parturients are at risk can allow earlier intervention. However, machine learning did not improve prediction of breakthrough pain during labour as compared to multivariable regression methods. A single centre retrospective study compared two machine learning predictive models (XGBoost, random forest) and logistic regression with all models performing equally AUC 0.763–0.772 in assessing the risk of breakthrough pain (32). While the models do reflect a lower than necessary discriminative ability to predict breakthrough pain, the convergence of the AUC suggest that the limiting factor is likely data quality. Multicentre prospective trials with the inclusion of dynamic intrapartum data like cervical progression or CTG parameters can improve data quality and the model's generalisability. Despite the advent of AI, ensuring optimal epidural analgesia is a closer possibility but to predict the dynamic nature of breakthrough pain requires larger datasets and variables for AI to be trained on.

Currently, there is no accurate way to dose neuraxial anaesthesia for Caesarean section to achieve that precise T4 height (33). Anaesthetists may determine the dose based on the patient's parameters like age, height, position, weight, abdominal girth, duration of surgery while attempting to minimise the catastrophic risk of a high spinal block (34). Wei et al. used a machine learning model to determine the optimal intrathecal 0.5% hyperbaric bupivacaine dose during Caesarean section. The model which incorporated vertebral column length and abdominal girth explained 80% of the variability in optimal bupivacaine dose and predicted the required volume with relatively small error of 0.008 (35). AI shows potential in determining the appropriate neuraxial dose especially with the incorporation of other variables such as complexity of surgery, fetal weight and the use of opioids.

One of the notable side effects after neuraxial anaesthesia is hypotension with an incidence rate as high as 70% that is associated with maternal and fetal morbidity (36, 37). Brief episodes of hypotension can result in lower fetal Apgar scores and acidosis (36, 37). Gratz et al. evaluated a neural network that analysed pulse waves and arterial stiffness from continuous non-invasive blood pressure measurements to predict hypotension and vasopressor requirements after spinal anaesthesia for Caesarean section (38). Their proposed model demonstrated modest discriminatory ability compared with a single-feature extraction approach (area under the receiver operating characteristic curve 0.89 vs. 0.87, P < 0.001) (38). However, the investigators noted that there was limited transparency within the model regarding the physiological cause of hypotension and proposed to deepen the neural network to include multiple layers and to incorporate more physiological variables (like heart rate variability) to enhance the model's performance (38). Therefore, as AI progresses, it has the potential to more reliably predict hypotension which enables accurate vasopressor administration to maintain blood pressure. However, the lack of widespread availability of non invasive blood pressure waveform analysis in obstetric theatres may limit its worldwide adoption.

Use of AI to improve maternal care

Severe postpartum haemorrhage (PPH) is defined as blood loss of 1,000 mL or greater accompanied by symptoms and signs of hypovolemia within 24 h after birth (39). It remains one of the leading cause of maternal mortality worldwide (39). As seen in Figures 1, 2, AI can improve PPH prediction by integrating numerous risk factors involved; a recent meta-analysis identified up to 47 potential risk factors for PPH (40). Given the number of risk factors and its complex interactions, accurate prediction of PPH is difficult without the assistance of complex modelling; an area that AI is most suited for. A systematic review demonstrated machine learning had a higher likelihood of PPH prediction compared with conventional methods when applied to the same dataset (OR 1.95, 95% CI: 1.88–2.01, p < 0.001) (41). Current machine learning studies used predictive variables such as BMI, fetal size, parity, maternal weight, age, type of delivery and ethnicity (41). However, what remains to be seen is the incorporation of more dynamic data such as fetal sonography, graphic representations (like fetal CTG, maternal heart rate variability) which may potentially improve the AI model's performance (42). While individual assessment of PPH in the operating theatre is prone to human error or biases, the introduction of standardized checklist has provided a structure for risk stratification. In the future, complex patterns analysed by AI may augment the traditional checklist approach to improve the prevention of severe PPH.

Early warning scores identify hospitalized patients who are at high risk of clinical deterioration. Although there are scores for the general population, specific scores like modified early obstetric warning system (MEOWS), maternal early warning criteria (MEWC) and maternal early warning triggers (MEWT) have also been developed for obstetric care (43). This is due to differences in normal vital sign ranges such as an increase of 10–15 heart beats per minute, respiratory alkalosis, initial lower blood pressure and maternal hyperthermia in the first trimester as compared to non-pregnant women (44–46). Arnolds et al. compared the performance of general and obstetric early warning scores and a score developed using machine learning [electronic Cardiac Arrest Risk Triage score (eCART)] among patients hospitalized in obstetric wards over a 10 year period (43). Of the 19,611 patients, 43 (0.2%) required ICU transfer or died while 88 (0.4%) had an infection. eCART demonstrated superior discrimination for both deterioration (AUC 0.86) and infection (AUC 0.77) compared to all other scoring systems (p < 0.05) (43). However, the low incident of adverse events (0.2%) may overestimate the model discriminatory ability. AI modelling of large physiological datasets may enhance early clinical deterioration but this is a single centred retrospective study that will require prospective multicentre trials before there can be formal clinical implementation.

Synthesis across the perioperative pathway

Figure 2 summarises the AI applications across all three domains showing the strength of the evidence and its limitations. Most of these studies require prospective multicentre validation and the performance relies mainly on AUC and not a change in clinical actions. Till date, there is no published study of fully deployed AI systems within obstetric anaesthesia and the field is mainly in the developmental phrases. In clinical practice, a parturient undergoing LSCS may be simultaneously supported independently by various AI models. Future work should therefore move beyond individual model validation to interoperability with electronic health records. Healthcare systems can institute shadow deployments of AI models running parallel with standard care to evaluate how AI recommendations influence clinical decision so as to calibrate alert thresholds.

Challenges and future considerations

Despite the advancements in AI, many challenges including lack of transparency, accuracy, privacy risks and difficulties with clinical integration remain. Firstly, certain high complexity models of AI including deep neural networks adopt a “black box” approach in terms of its inputs and outputs (47). This is problematic because users have no information on how certain predictions were made. Traditionally, clinicians explain to patients the basis of medical decisions before patients can agree to being cared for. If the clinician is unable to understand the process that shapes a decision, this hinders explainability and understanding. Hence, this may undermine the principle of informed consent (the right to know and patient autonomy). Therefore, efforts are underway in explainable AI—an area that seeks to improve the transparency of its algorithms. For example, Lundberg et al. used an explainable AI system that quantified the individual contribution of each risk factor to intraoperative hypoxaemia (48). SHAP (SHapley Additive exPlanations) is a tool that aims to bridge the gap between algorithmic prediction to clinical interpretation by quantifying each factor's contribution to a prediction. Rather than simply providing a conclusion, every prediction is accompanied by a visual breakdown of factors allowing clinicians to decide on the modifiable actions (48). Therefore, integrating explainable AI tools into anaesthesia systems may improve transparency restoring the clinician's ability to justify decisions. As AI becomes increasingly prevalent, explainability should be mandatory rather than encouraged, black box models should not be deployed into practice regardless of its performance.

The accuracy of an AI output is dependent on the quality and characteristics of its training dataset. Increasingly, evidence suggests that under-representation of certain minorities may skew datasets giving rise to biases. For example, ChatGPT 3.5 produced less accurate Spanish responses than English when answering questions on labour epidural which can perpetuate misinformation and the low epidural rates among minorities (21). These findings underscore the need for firstly accurate translation of patient educational materials for patients and AI to learn from. Additionally, developers and regulatory bodies should require performance data broken down based on language, ethnicity as a requirement to ensure minimal biases in the development and validation of AI algorithms.

Obstetric care is unique in that the clinician is responsible for both the mother and fetus with linked but distinct physiologies. This poses challenges in accurately predicting patient outcomes based on retrospective data. In a systematic review appraising the use of composite outcomes in obstetrics randomised controlled trials (RCTs), there were only 22 trials studying both maternal and perinatal outcomes out of a total of 156; with the other trials solely focusing on either maternal or fetal outcomes (49). Hence, most algorithms will be trained on narrow siloed datasets poorly reflecting real life scenarios, whereby maternal and fetal interest may be in conflict. For example, giving glyceryl trinitrate to reduce uterine tone facilitates easy fetal delivery, but may further lower maternal blood pressure. It is precisely at such moments that the anaesthetist's clinical judgement and real-time assessment of haemodynamics become crucial that remains beyond the reach of AI. Therefore, AI algorithms trained on compartmentalised datasets may fail to capture the dynamic interplay of maternal and fetal well-being and may struggle to encode contextual priorities which are commonly considered during clinical settings. This is a structural problem that requires AI to be built with integrated maternal-fetal outcomes.

There are privacy concerns with the collection, transmission and storage of health and demographic data. However, these concerns have to be balanced against the need for a larger database that is essential for better algorithmic analysis to improve patient outcomes. Of late, federated learning which is a privacy preserving approach to AI training addresses these challenges by enabling decentralised model training. This allows hospital to collaborate yet keep patient data locally by sharing only the model weightage of each factor (50). Data loss during transmission is reduced while each institute can focus on enhancing their own data encryption protocols. In summary, the integration of federated learning offers a potential framework for robust data analysis and collaboration.

Integration of any new technologies such as AI poses legal and workflow automation challenges. The current AI regulatory landscape is a mix of diverse approaches that aim to ensure its safe development but may hinder international collaboration. For example, the USA Food and Drug Administration (FDA) takes a risk stratified approach using the predetermined change control plan which allows continuous improvements to AI-enabled devices even after approval thereby providing a reasonable assurance of safety and effectiveness (51). However, the European Union framework operates across regulatory bodies (like the European Medicines Agency) which potentially creates unresolved regulatory issues between the Medical Device Regulation and AI act (51). A practical implication of the regulatory difference would mean that an AI tool validated in one country cannot be readily deployed in another without incurring additional regulatory and legal reviews. Traditionally, medical negligence stems from a breach of duty and that medical care provided was below the standard expected. An unresolved question is whether liability may extend beyond a decision by a clinician who acted on an incorrect AI recommendation to the developers of the AI system. The absence of clarity shapes how clinicians engage with AI as a clinician who fears liability for acting based on AI's recommendation may reject its use. Therefore, a clear legal definition of shared liability, establishing governance on AI recommendations and an explicit requirement that patients are informed that AI contribute to their care are essential to provide accountability and safety.

The rapid pace of obstetric emergencies means hospital healthcare systems must provide timely information so that early intervention can be taken. Any AI tool must integrate seamlessly into existing workflows without adding cognitive burden during already stressful situations. A study showed that clinicians were less likely to accept alerts by 30% for each additional electronic health record reminder received (52). This is particularly important to address as AI alert systems may increase the rates of false alarms which could possibly worsen alert fatigue among clinicians. Hence, AI warning systems should firstly be designed with appropriate alert thresholds into clinical workflow as the foundation to assist clinicians. By limiting unnecessary notifications, such AI systems may improve clinician acceptance and enhance decision making.

To date, no studies have evaluated the cost effectiveness of AI integration specifically within obstetric anaesthesia. Developing and implementing AI within healthcare comes with substantial upfront and recurring costs, which vary depending on the complexity of the AI model and deployment strategy. Basic AI integration into an existing application can cost around $40,000 while a custom deep learning model may cost up to $100,000 (53). With AI successful implementation, estimates suggest potential savings of $200–360 billion across the United States (54). However, financial constraints remain a significant barrier to low income nations where over 92% of preventable global maternal deaths occur (55). Additionally, Africa hosts less than 1% of global data centre capacity and only about 5% of African researchers have access to high performance computers (56). Until structural and cost barriers are addressed, the technology remains out of reach from those who have the greatest need for it.

In conclusion, the use of AI in obstetric anaesthesia has the potential to change how we evaluate, and provide obstetric care. AI has been vital across the entire perioperative pathway from relieving anxiety, refining neuraxial techniques to optimising labour analgesia and improving maternal care. To ensure effective clinical AI implementation, external validation of existing models across ethnicities is essential to ensure generalisability. Adaptive regulatory frameworks have to be a priority to ensure that AI models balance innovation with safety by addressing privacy and transparency. Lastly, anaesthetists should be equipped with AI literacy as a formal competency so that they can critically evaluate AI at work especially when clinical intuition overrides AI recommendations. The path ahead requires the collaborative work of ensuring that AI enters our field not on it terms but on what mothers deserve and what clinicians deem safe.

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.

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