The safe introduction of artificial intelligence tools into obstetric practice depends not only on performance metrics; depends on monitoring for bias, external and prospective validation, explainability, chain of responsibility and regulatory/reporting standards. This article addresses this critical but relatively little-talked-about dimension of the field in an evidence-based manner.
⚕️ Target audience: This content is aimed at healthcare professionals and is for educational/current-information purposes. AI tools offer clinical decision support; It does not replace clinician responsibility and approved clinical workflows.
As applications of artificial intelligence (AI) in obstetrics and perinatology proliferate, attention often focuses on performance figures such as sensitivity, specificity, and AUC (area under the curve—a measure of a model's discriminatory power). However, the real factors that determine whether a tool is clinically reliable and fair are the questions behind these figures: In which population was the model developed? Does it perform equally across different ethnic and socioeconomic groups? Has it been prospectively validated at an independent center? Who is responsible for the mistake? This article holistically addresses the ethics, bias, validation, and regulatory framework of obstetric AI.
Algorithmic bias — when a model systematically produces inaccurate or disadvantageous results against certain groups — is not an abstract concern in maternal health. In a field where significant ethnicity-based disparities in maternal mortality and morbidity persist, models trained on historical data can not only reflect but reinforce these disparities at scale. Because machine learning models learn by accepting patterns in past decisions and access differences as "correct."
The most concrete obstetric example in this field is the vaginal birth after cesarean section (VBAC) success prediction calculator. The original calculator developed by the MFMU (Maternal-Fetal Medicine Units) network included race/ethnicity as a variable, lowering the predicted probability of success for Black and Hispanic pregnancies; There was concern that this could unequally discourage the recommendation of attempted labor (TOLAC). The variable was removed from the model in 2021. In an interrupted time-series analysis using US data, TOLAC and VBAC rates after removal were unchanged overall and across the four major ethnic groups; whereas the ethnic gap in successful VBAC rates (e.g., Hispanic 75.0%; Black 67.0%; white 77.2% versus other non-white groups) has been shown to persist (Hikade et al., Obstet Gynecol 2025). The lesson is twofold: removing the race variable is a necessary step, but alone it does not resolve the underlying clinical disparity.
An example of how bias can be reduced, not just detected, is in the prediction of perinatal mood and anxiety disorders (PMAD). In models trained from electronic health records in a center, data is reweighted in the pre-processing stage; It significantly reduced the demographic parity difference and the inter-ethnic false-negative difference, but did not significantly reduce the discriminatory power (AUROC ~0.61–0.64) (Wong et al., JAMA Netw Open 2024). This single-center study suggests that the “fairness–performance trade-off” is not as sharp as is often assumed; but external validation is required.
Just because a model performs well on the data for which it was developed (internal validation) does not mean that it is ready for clinical use. Two separate layers are required: external validation (testing the model under different devices/protocols/demographics in a population outside the center where it was developed) and prospective validation (measuring the effect on clinical outcome in a prospective, real-life workflow, ideally in a randomized controlled trial). A significant portion of obstetric prediction models are published only retrospectively, single-center, and with internal validation; This creates the risk of decreasing performance when moved to another center due to distribution shift.
The classic obstetric example of how prospective evidence can limit “high performance” claims is in the field of intrapartum cardiotocography: in the large multicenter INFANT randomized trial with the addition of computerized CTG decision support, the system was shown not to reduce poor perinatal outcomes. This is a reminder of the gap between in silico or retrospective success and true clinical benefit. Therefore, prospective, outcome-oriented evaluation—not just diagnostic accuracy—is indispensable.
Most deep learning models are “black box”; How they arrive at a recommendation cannot be seen directly. Explainable AI—methods that attempt to present to the clinician in an interpretable format the factors that contribute to the model's decision; e.g., heat maps, feature importance scores—aim to close this gap. But explainability alone is not assurance: post-hoc explanations can be misleading or inconsistent, and a “plausible-looking” explanation can lend unwarranted credence to a faulty proposal.
A related critical danger is automation bias — the clinician's over-reliance on the automated system's recommendation to the exclusion of his own judgment. In an experimental study on clinical decision-making, it was shown that clinician accuracy increases when the AI recommendation is correct, but when incorrect recommendations are presented, clinicians are misled and their diagnostic accuracy decreases (Kücking et al., Int J Med Inform 2025). Practical takeaway: AI should be positioned as a “second opinion,” preserving workflows where the final decision rests with the clinician, and communicating the error profile of the system transparently to the user.
Who is responsible if a decision made by an AI tool makes an error? Clinician, institution, manufacturer? The standard position in current medical-legal frameworks is that AI provides decision support, with ultimate responsibility remaining with the clinician. This makes it an ethical obligation for the clinician to know the indication, limitations and approval status of the tool. A prerequisite for accountability is traceability: the model version used, training/validation datasets, and the population for which performance was demonstrated must be clearly documented. This expectation of transparency also forms the core of the following reporting standards.
Incomplete or selective reporting of AI research hinders critical evaluation and replication of evidence. Reporting guidelines that were developed to close this gap and are considered standard today are:
These standards form a hierarchy: predictive model development/evaluation (TRIPOD+AI) → early clinical evaluation (DECIDE-AI) → randomized intervention trial protocol and report (SPIRIT-AI / CONSORT-AI). When evaluating a tool, asking which guidance is followed at which stage is a practical way to quickly read the maturity of the evidence.
AI-based software is regulated as “medical device software” (SaMD) in most jurisdictions. In the USA, the FDA has adopted the total product lifecycle (TPLC) approach for software that can learn and adapt after distribution with the "AI/ML-Based SaMD Action Plan" dated 2021; With the concept of "Algorithm Change Protocol" it allows the manufacturer to pre-define future model changes. In the draft guideline (AI-enabled device software functions: lifecycle management and marketing application recommendations) published on January 6, 2025, FDA further focused on lifecycle management and post-deployment performance monitoring.
In Europe, AI-based medical devices require CE marking under the Medical Device Regulation (MDR) or the In Vitro Diagnostics Regulation (IVDR), depending on their nature; also, at a horizontal level, the EU Artificial Intelligence Act (AI Act) subjects many AI systems used in healthcare to additional obligations in the “high risk” category. The clinically important distinction is that regulatory approval (FDA clearance/CE mark) indicates that the tool meets a minimum safety/performance threshold for a specific intended use and population; It does not guarantee the existence of evidence of clinical benefit or external validation in your patient group. These two concepts (approval ≠ proven clinical superiority) should be kept separate.
The World Health Organization's 2021 guide "Ethics and governance of artificial intelligence for health" sets out six key ethical principles for AI in health: protection of autonomy; consideration of human welfare, security and public interest; transparency, explainability and understandability; responsibility and accountability; inclusion and equality; responsive and sustainable AI. WHO has extended this framework to the era of generative AI with additional guidance published on January 18, 2024, containing more than 40 recommendations on large multi-modal models (LMM).
On the ultrasonography side, ISUOG's (International Society for Ultrasonography in Obstetrics and Gynecology) AI positioning statement recognizes the opportunity technology offers for women's health; It adopts careful evaluation of strengths/weaknesses before implementation, transparent identification of methods and data sets used in development and validation, and equality as its guiding principle. Professional organizations such as FIGO (International Federation of Gynecology and Obstetrics) and ACOG (American College of Obstetricians and Gynecologists) also take a cautious line that AI tools should be adopted only with appropriate validation, bias control and clinician supervision. The common emphasis is clear: excitement should not replace evidence.
Practical summary — things to ask when evaluating an AI tool: (1) In what population was it developed and can it be generalized to my patient group? (2) Does it have external and ideally prospective validation at independent centres? (3) Have differences in performance and false-negatives been reported across ethnic/socioeconomic subgroups? (4) Has benefit on clinical outcome been demonstrated or is it merely diagnostic accuracy? (5) What is the regulatory approval status (FDA/CE) and approved intended use? (6) What standard (TRIPOD+AI / DECIDE-AI / CONSORT-AI) does the reporting comply with?
The reporting standards (TRIPOD+AI, CONSORT-AI, SPIRIT-AI, DECIDE-AI) and institutional guidelines (WHO, ISUOG) cited in this article are high-authority, consensus-based documents and have been verified. In contrast, most clinical evidence on bias reduction and fairness metrics is single-center, retrospective, and has limited external validation; The reported PMAD bias-reduction results (Wong et al. 2024) are from a single center and their generalizability is uncertain. Findings in the VBAC sample are specific to the US population; Ethnic categories and inequality dynamics vary from country to country. Data on automation bias come from experimental settings and vary in magnitude in actual clinical workflow. As regulatory frameworks are rapidly evolving (especially the FDA's 2025 draft guidance and the EU AI Act implementation), current official texts should always be checked.
Ethics, validation and regulation; It is a prerequisite for AI adoption, not an “optional add-on.” The practical attitude for the clinician is to evaluate a tool not only on its impressive performance figures; to evaluate it based on its external/prospective validation in its population, subgroup fairness evidence, explainability, validation status, and quality of reporting. AI recommendations should be treated as a second opinion, one should be aware of automation bias, and the final clinical decision—with responsibility—must remain with the physician. Fair, transparent, and externally-validated tools have the potential to improve maternal outcomes; however, this potential only translates into clinical benefit when the above assurances are met.
Last updated: . Sources have been confirmed with their publication dates and current versions.