AI-assisted antenatal triage enters the randomized trial stage
Early pregnancy risk assessment still misses a substantial share of pregnancies that later develop serious complications. In routine obstetric care, the first-trimester label of “low risk” can be misleading, especially when maternal characteristics, obstetric history, and evolving clinical signals are considered together rather than in isolation. The CURAte trial is designed to test whether an AI-assisted stratification tool can improve the way clinicians sort pregnancies into high-risk and low-risk categories, and whether that change alters downstream maternal and fetal outcomes.
The protocol is notable because it moves beyond prediction accuracy alone. Many machine-learning systems perform well on retrospective datasets yet never reach the bedside in a way that changes actual care. Here, the intervention is not merely an algorithmic score sitting in parallel with clinical judgment; it is a decision-support process embedded into prenatal management, with predefined pathways for what clinicians should do when the tool flags a pregnancy as high risk. The trial therefore asks a more clinically meaningful question: can earlier and more consistent recognition of risk change what happens before delivery and in the early postpartum period?
Two decision points, one clinical pathway
Investigators will randomize 1,700 participants in a 1:1 ratio to AI-assisted risk stratification or standard antenatal care without AI support. Randomization will use block allocation and will be stratified by ethnicity and parity, a design choice that acknowledges how both variables can shape obstetric risk and health-system contact patterns. Participants will be blinded to allocation, while clinicians will receive the categorization output in the intervention arm and an undisclosed output in the control arm.
The AI assessment draws on maternal sociodemographic, clinical, and obstetrical data at two time points: the first trimester antenatal visit and again between 31 and 33 weeks’ gestation. Those two checkpoints are clinically sensible. The first visit captures preexisting vulnerability and prior pregnancy history, when preventive planning can still influence the trajectory of care. The late-third-trimester reassessment provides a second chance to detect women whose risk profile has shifted as pregnancy advances and complications emerge.
In practice, the tool will assign pregnancies in the intervention arm to either “high-risk” or “low-risk.” Clinicians then follow a predefined protocol tailored to each classification. For pregnancies categorized as high risk, clinicians work through referral criteria for the maternal-fetal medicine-led high-risk clinic and complete specified investigations and interventions before referral. The trial thus standardizes not just risk labeling, but the management cascade that follows the label.
A risk score changes nothing if it does not change care.
Why machine learning may matter in obstetrics
The biological and clinical rationale for this trial is straightforward. Adverse obstetric outcomes often arise from interacting processes rather than a single dominant signal. Maternal age, parity, prior pregnancy complications, medical comorbidity, and social determinants can all contribute to placental dysfunction, hypertensive disease, fetal growth problems, preterm birth, or postpartum complications. Human assessment can be strong, but it is also vulnerable to cognitive shortcuts, incomplete data synthesis, and variation between clinicians or sites.
A machine-learning model may help by weighting a broader set of variables simultaneously and by detecting patterns that are difficult to integrate mentally during a busy antenatal visit. If the tool is sufficiently calibrated, it could improve sensitivity for identifying pregnancies that truly need closer surveillance, referral, or targeted intervention. The protocol’s emphasis on first-trimester classification reflects a clear clinical aim: shift the point of recognition earlier, before a complication becomes entrenched.
The choice to reassess later in pregnancy also reflects obstetric reality. Risk is not fixed at booking. A pregnancy that appears uncomplicated early on may later reveal fetal growth restriction, rising blood pressure, symptomatic disease, or a change in obstetric circumstance. Re-stratification at 31 to 33 weeks offers a second opportunity to refine the care plan before the final stretch of gestation.
Primary outcome anchored in real-world maternal and neonatal events
The trial’s primary composite outcome is extracted from medical records six weeks after the end of pregnancy. While the protocol summary does not list each component in the excerpt provided, the composite is explicitly intended to capture adverse maternal and fetal/neonatal outcomes, making the endpoint clinically grounded rather than purely algorithmic. Follow-up through six weeks postpartum also matters because key maternal complications can emerge after delivery, when routine obstetric surveillance often begins to wane.
The two study arms will be compared on an intention-to-treat basis after the database is locked. No interim analysis of the primary outcome is planned, which reduces the risk of premature interpretation from incomplete data. This is a protocol for a pragmatic effectiveness trial, not a proof-of-concept accuracy study, and that distinction shapes every element of the design.
Ethical oversight and limits built into the protocol
Ethical approval has already been obtained from the National Healthcare Group Domain Specific Review Board, and written informed consent will be obtained from all participants. Data handling will follow institutional data-protection requirements and the Personal Data Protection Act. The AI system is explicitly positioned as clinical decision support rather than a replacement for clinician judgment, which is crucial in obstetrics where context, patient preference, and rapidly changing physiology still govern final decisions.
Subgroup analyses are planned to assess whether effects differ across key demographic and obstetric groups. This is especially important in a tool that is stratified by ethnicity at randomization and built from real-world clinical data, because performance can vary across populations and care settings. Any model used in prenatal care must prove that it helps across diverse groups, not only in the average participant.
The protocol’s limitations are already visible. As a randomized trial, it can test whether AI-assisted stratification changes outcomes in the setting studied, but it cannot by itself guarantee portability to other health systems or confirm that the model will retain the same performance as data practices evolve. It also depends on clinician adherence to the predefined response pathways, because a classification tool only has value when the care pathway that follows it is implemented consistently.
Still, the study is unusually important for one reason: it places an AI tool into the same evidentiary frame used for drugs, procedures, and care pathways. For obstetrics, where missed risk can carry consequences for both mother and baby, that is the level of proof the field has been waiting for.