The birth of a child, which can be one of the most joyous moments in a parent’s life, is also one that can be accompanied by profound anxiety.
Despite medical advances, maternal and fetal risks remain a critical challenge. Pregnancy-related complications can result in maternal deaths and serious health complications, and severe maternal morbidity continues to affect women and families across the U.S. These challenges underscore the importance of continued efforts to improve the safety and quality of perinatal care.
Product and technical teams at GE HealthCare1 are working to better understand these challenges as they explore the potential of AI in perinatal care at one of the most profoundly important moments in a human life[1]: when the cry of a newborn baby stills the bustle of activity in a hospital room and the cacophony of sounds that characterize our modern world.
A fragmented system
Improving the quality of perinatal care is no easy undertaking: few areas of medicine require clinicians to process as much real-time data, as quickly, as labor and delivery. One of the signals assessed during birth is the fetal heart rate, which clinicians consider alongside other clinical information. Abnormal fetal heart rate patterns may require further clinical evaluation.
However, monitoring the fetal heart rate can be a complicated process. Nurses and physicians navigate separate interfaces, documentation workflows, and device outputs that are not always aligned. Information often has to be located, interpreted, and reconciled across screens, creating operational friction at the point of care.
Product, technical, and clinical teams across the healthcare industry are examining how AI may be evaluated responsibly in complex clinical environments. That work requires careful attention to the clinical problem, the information clinicians use, the design of the workflow, and the methods used to assess performance and usability.
Over the last few years, I’ve seen institutions exploring AI consider rollout across the five stages outlined below.
1. Define the clinical problem together
Before anything, I recommend that you sit down with frontline clinicians to understand where caring for patients may be more difficult than it needs to be. For example, fetal monitoring requires clinicians to assess several features together: the baby’s baseline heart rate, fluctuations around that baseline, and any increases or decreases in heart rate in relation to uterine contractions.
Two clinicians reviewing the same tracing may differ in how they classify a slowdown or identify the baseline. Even experienced clinicians can interpret the same fetal monitoring data differently.
Further complicating matters is the inherently fragmented nature of the workflow. Nurses often have to review information from multiple sources, including maternal vital signs and medication infusions, while also documenting the assessment and care provided. When those details sit in separate interfaces, clinicians have to bring them together while continuing to care for the patient.
Working backward from these day-to-day challenges can help define the problem you’re trying to solve: the need to bring disparate details together, while also exploring ways to support interpretation and simplify documentation.
2. Design with the clinicians who will use it
A nurse reviewing a fetal heart rate slowdown needs to make sense of that information in context. When does the heart rate begin to fall? When does it recover? How does this episode compare with the preceding tracing? The way information appears on a screen affects how readily a clinician can examine those relationships.
This is the kind of clinical perspective that should ideally inform every aspect of development. These perspectives can be gathered during meetings between product and clinical teams where attendees review workflows, usability, and the presentation of information. These meetings can surface important insights. To give just one example, nurses might ask for larger fetal monitoring strips so they can assess the tracings more readily.
Recurring discussions allow clinicians to influence the design as it develops.
3. Build trust by explaining what the AI is intended to contribute
For a nurse caring for a woman in labor, an AI assessment introduces a practical question: how much weight should this information carry in the care of this patient? That question can have significant clinical implications when the algorithm’s interpretation differs from what the nurse sees in the tracing.
For example, imagine a nurse might identify a heart rate slowdown that an AI assessment has not marked. The nurse must retain responsibility for reviewing the tracing in context and determining what attention the patient needs. The discrepancy could also give an oversight team a specific case to examine.
Trust, then, depends on making these boundaries clear. When we explain what the AI is intended to evaluate, are candid about where its performance remains uncertain, and consider ways for colleagues to raise concerns when an assessment does not match their clinical judgment, nurses may better understand the intended role of the technology, with their clinical judgment, not the algorithm, guiding care at the bedside.
4. Give clinical concerns a clear route to review and action
Concerns raised by a clinician might require follow-up from several team members. A nurse might question an interpretation, have difficulty finding relevant information, or report that a workflow takes longer than expected. Each concern needs someone with the authority and expertise to follow up.
I’ve seen organizations exploring AI bring together clinical leadership, nursing, IT, informatics, quality, patient safety, and operational teams early to define success measures, monitoring processes, and escalation pathways. The purpose is to connect what clinicians experience during evaluation with decisions about training, changes to workflows, and readiness for further evaluation.
Consider a situation in which a nurse reports that an AI assessment differs from their reading of a fetal tracing. The clinical concern is immediate, and clinical judgment takes precedence. A subsequent review can then examine the disagreement and determine what follow-up is appropriate. Was the issue with the interpretation itself? Was the displayed information difficult to understand? Does the team need further training on what the algorithm assesses? These are questions for investigation, and they require input from different disciplines.
Establishing those responsibilities before a potential evaluation can give users a defined route for raising concerns. Continuing the review afterward could allow the organization to decide whether a problem calls for additional training, a change to the application or workflow, or further investigation before any expansion.
5. Measure the work clinicians do before expanding
Making an application available across a hospital tells us little about whether it may help a nurse complete a clinical task. We need to understand whether colleagues use it during patient care, whether it may make information easier to review, and where it may add work.
That’s why we need to agree on measures of adoption, workflow impact, clinician experience, and potential operational value before an evaluation begins. This is the right time to be meticulous about tracking the time required to complete key tasks, clinician feedback, adoption and utilization, and consistency across sites.
For example, we could observe a nurse reviewing a fetal tracing alongside maternal clinical information. We could time that review before and after introducing an application in an appropriately controlled evaluation, then ask the nurse where they ran into difficulty. If they still had to pull information from another system, that would help explain the result and show what may need further evaluation. This way, we could evaluate whether a single view might affect the time required instead of assuming that it does.
Together, these five steps can help guide the responsible evaluation and potential rollout of AI across a health system. The stakes could not be higher. Maternal and fetal mortality are more than clinical problems; they are tragedies for families and communities. That’s why I am proud of GE HealthCare’s ongoing efforts to explore how AI might help nurses and physicians review fragmented information, with the goal of studying its potential role in the interpretation of fetal monitoring and clinical workflows.
Gregory Frank will be participating in a panel at HLTH 2026 on evaluating enterprise AI in healthcare.
- Technology in development. Not cleared or approved by the U.S. FDA. Not available for sale. Its safety and effectiveness have not been established. ↩︎
