Artificial Intelligence in Obstetrics: Overview, Concepts and Current Status

Artificial intelligence (AI) has made a rapid leap forward in many fields in the last decade, from image recognition to text generation, and this momentum has also been reflected in obstetrics and perinatology. A wide range of applications are available in the literature, from automated fetal biometry to preeclampsia risk models, from cardiotocography (CTG) interpretation to embryo selection and large language models writing clinical notes. However, this field is also open to exaggerated expectations. This article; It aims to clarify key concepts, map current practice, and honestly reveal both the opportunities and limitations that are often overshadowed. The other seven articles of the section discuss individual application areas in depth; This text offers the common framework and dictionary.

Core concepts: AI, machine learning, deep learning and LLM

Artificial intelligence (AI) is an umbrella term that covers computer systems that perform tasks that normally require human intelligence (pattern recognition, classification, prediction). Almost all of the systems we encounter in the clinic today fall into the machine learning class (methods that learn patterns in data from examples, without writing explicit rules) under this umbrella.

The most popular branch of machine learning is deep learning (deep learning - methods using multilayer artificial neural networks). Convolutional neural network (CNN), which is frequently used in image-based tasks, is a basic tool in tasks such as recognizing fetal ultrasonography planes or drawing organ boundaries (segmentation). According to the learning method, the methods are roughly divided into three: supervised learning that learns with labeled data (for example, training with labels "this image is the standard plane"), unsupervised learning that searches for structure in unlabeled data, and reinforcement learning that develops strategies with feedback.

Large language models (LLM - large language models; models that are trained with very large text corpuses and produce text close to human language, e.g. GPT family), which have been setting the agenda recently, are based on an architecture called transformer. Versions of these that can process images, sound, and text together are called large multi-modal models (LMM); The World Health Organization has discussed the governance of these models in health as a separate heading in its 2024 guide.

Map of AI applications in obstetrics

Gathering obstetric and perinatal AI applications into five main clusters makes it easier to understand the field. The table below summarizes this map and the related detailed article in our section.

A comprehensive screening review focusing on obstetric ultrasonography embodies the concentration of work in the field: Horgan et al (2023) identified 127 publications on this topic for ultrasonography alone and listed first trimester assessment, placental analysis, fetal biometry, fetal echocardiography, neurosonography, and fetal anatomy scanning as its primary uses. The common promise of these applications is; The aim is to reduce operator dependency, increase efficiency and improve scanning access, especially in low-resource environments.

How is performance measured in predictive models?

The success of predictive models is often reported in terms of AUC (area under the curve — a measure that summarizes how well the model discriminates between sick and healthy between 0.5 and 1.0), sensitivity, specificity, detection rate (DR) and false positive rate (FPR). For example, in a systematic review for preeclampsia, the AUC values ​​of the machine learning models included were reported between 0.860 and 0.973 (Ranjbar et al., 2024). While these figures may seem impressive, the same review highlights that all included studies were of a retrospective cohort design and were small in number (only four studies)—so high AUC alone is not evidence of clinical readiness.

Opportunities: What can AI really improve in obstetrics?

A balanced view also requires clearly defining the true potential of the field. The most consistent promises are:

• Reducing observer variability: Intraobserver and interobserver discrepancy is a known problem in subjective tasks such as ultrasonography biometry and CTG interpretation; AI-supported measurement standardization can contribute here.
• Efficiency and workflow: Automated plane capture, measurement and reporting can free up decision-making time by reducing scanning time and relieving the clinician of repetitive tasks.
• Access and equity potential: The possibility of expanding screening access by partially compensating for specialist shortages in low-resource settings is frequently emphasized (Horgan et al., 2023).
• Relieving documentation burden: Large language models and ambient systems hold the promise of reducing administrative burden by accelerating clinical note and summary production.

The common feature of these opportunities is that they are mostly supportive and workflow facilitating; The evidence that it directly improves clinical outcomes (such as neonatal morbidity, live birth) is much more limited.

Level of evidence and limitations

The most critical feature of the obstetric AI literature is the distance between performance promises and solid clinical evidence. The example that most clearly demonstrates this distance is the large-scale INFANT randomized controlled trial, which tested computerized decision support in intrapartum fetal heart rate monitoring. In this study analyzing 46,042 women at 24 centers in the United Kingdom and Ireland, decision support software added to CTG interpretation did not alter the rate of poor neonatal outcome (0.7 versus 0.7 percent; adjusted risk ratio 1.01; 95% CI 0.82–1.25) and no difference was found in developmental assessment at age 2 (INFANT Collaborative Group, Lancet 2017). The fact that a tool promising high diagnostic accuracy did not improve outcome in the randomized condition underscores the principle of “accuracy metric ≠ clinical benefit” in this field.

A similar pattern is seen in reproductive medicine: systematic review comparing AI to embryologists in embryo selection reports that models perform consistently well on image/morphology tasks; however, the authors clearly state that most of the studies were not evaluated prospectively in a clinical setting and many lack external validation (Salih et al., 2023). As a matter of fact, randomized studies comparing time-lapse + AI-based selection with standard morphology have generally failed to show superiority in live birth/pregnancy rate. These null findings need to be highlighted for an honest picture of the field.

The main limitations that recur in all applications are as follows:

• Data quality and representation: Models “learn” within the boundaries of the data on which they are trained. Data derived from a single center or narrow population results in performance degradation across different devices, protocols, and populations.
• Lack of external and prospective validation: The preponderance of the literature is limited to retrospective and internal validation; Validation in independent centers and prospective design is still the exception.
• Algorithmic bias: Unbalanced data sets can lead to inaccurate performance across racial/ethnic/socioeconomic subgroups, deepening existing maternal health disparities.
• Explainability: The “black box” nature of deep learning models creates challenges in terms of clinical justification, error analysis, and liability.
• Automation bias: Clinician over-reliance on AI output (automation bias) can impair independent clinical judgment.
• Distribution drift: Changing population, device, and application patterns over time can silently erode model performance; This requires constant monitoring.

Risks specific to large language models

LLMs carry unique risks when entering obstetric practice. In a study evaluating common obstetric questions for pregnant women, the average rating experts gave to ChatGPT responses was 4 out of 5, with 75 percent of responses rated positive (≥4); However, the dimensions of completeness (mean 3.8; positive rate 46.7 percent) and safety (mean 3.9; positive rate 53.3 percent) fell behind accuracy (mean 4.2) (Peled et al., 2024). The authors particularly emphasize the need for caution in questions involving maternal or fetal safety. Hallucination (producing unrealistic but convincing output), limited timeliness of evidence, privacy/KVKK concerns and bias are the main factors that necessitate the responsible use of these models.

Reporting and evaluation standards

These limitations have necessitated a common quality and transparency framework for the field to mature. Two standards are central to the scope of this article:

• TRIPOD+AI (Collins et al., BMJ 2024): An updated checklist for transparent, complete and accurate reporting of clinical prediction models developed with regression or machine learning methods. Transparent reporting is a prerequisite for any quality assessment.
• CONSORT-AI (Liu, Cruz Rivera et al., Nat Med 2020): 14 new items added to CONSORT 2010 for clinical trial reports containing an AI component; It requires reporting of elements such as the definition of the intervention, handling of inputs/outputs, human-AI interaction and analysis of error cases.

In addition to these two frameworks, tools such as SPIRIT-AI for protocols, DECIDE-AI for early clinical assessment, and PROBAST-AI for bias assessment play a complementary role; The regulatory dimension (FDA's AI/ML-based software medical device framework, CE/IVDR) and institutional stances (ISUOG's AI position statement, FIGO, WHO AI ethical guidelines) determine the basis for clinical use. Details of these standards are discussed in our department's "Ethics, Bias, Validation and Regulation" article.

Its reflection on clinical practice and hyperlink to other articles

The practical implication for the clinician is clear: AI tools play a supporting role where they are most powerful today—reducing observer variability, measurement standardization, and workflow efficiency. Questions to ask when evaluating a tool include: For what population and with what devices was it developed? Has it been prospectively validated at an independent, external center? What is the approval/CE status and level of evidence? How does its output affect clinical decision and how is liability defined in case of error? TRIPOD+AI and CONSORT-AI provide a practical checklist to answer these questions.

Building on this general framework, the chapter's other papers delve into each application area: AI in Fetal Ultrasonography , AI in Cardiotocography (CTG/NST) Interpretation , AI in Preeclampsia Prediction , AI in Preterm Birth Prediction , AI in Embryo Selection in IVF/ART , Large Language Models in Obstetric Practice and Ethics, Bias, Validation, and Regulation . They all share the same message: take the promise seriously, but take the evidence even more seriously. (Ethics, Bias, Validation and Regulation)

← Return to Artificial Intelligence in Obstetrics