17 November 2026 · 13 min

Company Spotlight: Butterfly Network Gestational Age Tool - AI Prenatal Biometry

Radiology & ImagingFDA Clearances

We unpack the FDA 510(k) summary for the Butterfly Gestational Age Tool, an AI software that estimates gestational age from blind ultrasound sweeps. We explore the clinical trial data, the limits of agreement across different pregnancy windows, and the specific labeling that restricts its use in delivery planning.

Key points

Source: FDA 510(k) summary K252148: Butterfly Gestational Age Tool (Butterfly Network) - U.S. Food and Drug Administration, 2026

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Company: Butterfly Network (butterflynetinc.com) · Product: Butterfly Gestational Age Tool · FDA 510(k) K252148

Your company here. This podcast is looking for its first sponsors: reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsorship options and rates →

Transcript

Sam: Imagine estimating a patient's gestational age just by sweeping a handheld ultrasound probe across their abdomen, without ever needing to interpret a live image on the screen.

Maya: That is exactly what we are looking at today. We are breaking down the FDA 510(k) summary for the Butterfly Gestational Age Tool, a document published by the U.S. Food and Drug Administration on March 27, 2026.

Sam: And before we dive in, a quick reminder that our voices are AI-generated, and this episode is a summary of a public document. This is a Company Spotlight episode, meaning we are focusing on one specific product clearance.

Maya: Importantly, we are analyzing the FDA's public 510(k) summary, not the company's marketing material. We will look at the real data, the specific indications for use, and the limitations the agency outlined.

Sam: Alright, let us start with the basics. What exactly is the Butterfly Gestational Age Tool, and who is it for?

Maya: It is a software application designed to be used with Butterfly's ultrasound systems, specifically the iQ+ or iQ3 probes. It provides an output of gestational age for a singleton intrauterine pregnancy presumed to be between 16-37 weeks gestation.

Sam: So it is not for the first trimester, and it is strictly for singleton pregnancies, meaning no twins or multiples. How does the workflow actually operate at the bedside?

Maya: This is where it gets very interesting from a clinical workflow perspective. The user launches the tool within the iQ App's OB scan presets. First, they measure the fundal height in centimeters.

Sam: Just with a standard tape measure?

Maya: The document does not specify the physical measuring device, just that they measure the fundal height in centimeters. That measurement then determines the number of ultrasound videos, which they call sweeps, that are needed.

Sam: And these sweeps are where the AI comes in.

Maya: Exactly. The user captures short cines by moving the probe across the abdomen. But here is the massive shift: they do this without relying on the live B-mode. Instead, the system presents users with animations for each sweep to communicate the intended path and probe orientation.

Sam: Wow. So a traditional sonographer is constantly looking at the live B-mode scan to find specific anatomical landmarks, like the fetal head or the femur length, to calculate age. This tool removes that requirement completely.

Maya: Right. The operator just follows the animation, making direct contact with the patient's skin using a coupling medium like ultrasound gel. The collected sweeps are then input into a deep-learning model within the tool, which outputs an estimated gestational age.

Sam: That is a textbook example of democratizing a specialized skill. You are moving the capability of biometry from specialized sonographers using cart-based machines to primary care clinicians or emergency personnel using a handheld tablet.

Maya: It is, but there is a very strict boundary on how this information can be used. The Indications for Use section explicitly states: This adjunctive information is not intended to be used for prenatal management and/or delivery planning.

Sam: Wait, really? You get a gestational age, but you cannot use it for delivery planning? Does the FDA summary say why that restriction is there?

Maya: It does not. If the document does not explain the reason for a warning, we stop there. We can only report that the limitation is explicitly written into the cleared indications.

Sam: Fair enough. So it is for use by qualified and trained healthcare professionals in environments where healthcare is provided, but it acts more as an initial assessment or triage tool rather than the final word on managing the pregnancy.

Maya: That is the practical takeaway for clinicians and health systems buying this technology. Now, let us talk about how this AI was actually developed and trained, because the numbers are impressive.

Sam: Let us hear the data. Where did the deep learning model come from?

Maya: It was developed by the University of North Carolina in the Fetal Age Machine Learning Initiative, or FAMLI protocol. They conducted a prospective study collecting over 100,000 cine loops comprising millions of image frames from thousands of patients.

Sam: That is a massive training dataset. Where were those patients located?

Maya: The sites were within the United States, specifically the University of North Carolina, and in Zambia. Point-of-care ultrasound users obtained blind sweeps alongside standard fetal biometry performed by sonographers.

Sam: So they compared the AI's blind sweeps directly against the gold standard of expert biometry. How did the algorithm perform in their main test set?

Maya: The mean absolute error was 3.9±0.12 days for the model, versus 4.7±0.15 days for biometry. That is a difference of -0.8 days, with a P-value of <0.001.

Sam: So in that specific development cohort, the AI actually had a slightly lower mean absolute error than traditional biometry.

Maya: Yes, and the summary notes the results were similar in North Carolina, with a difference of -0.6 days, and Zambia, with a difference of -1.0 days.

Sam: Okay, that is the training and development side. But what did the company submit for this specific 510(k) clearance? They had to run their own clinical performance test, right?

Maya: They did. The clinical performance evaluation tested the tool on pregnant subjects with confirmed gestational age ranging from 16 to 37 weeks. They explicitly state this was a dataset totally independent from the development data.

Sam: How many patients were in this independent test?

Maya: There were 111 unique subjects enrolled between March 2025 and February 2026 across 4 USA locations.

Sam: 111 subjects is a relatively small sample size compared to the thousands in the training data. Where were these 4 locations?

Maya: This is an important detail. Site 1 was the Butterfly offices in Burlington-MA. Site 2 was Thomas Jefferson University in Philadelphia-PA. Site 3 was Remedy Direct Primary Care in Flagstaff-AZ. And Site 4 was the Butterfly offices in NYC-NY.

Sam: Let me stop you there. Two of the four clinical trial sites were the company's own offices?

Maya: Yes. And looking at the breakdown in Table 1, out of the 111 total subjects, 58 were at Site 1, and 36 were at Site 4.

Sam: That is something health-system leaders should definitely note. The independent validation relies heavily on data collected in the manufacturer's own facilities, rather than widespread external hospital environments. Who actually performed the scans on these 111 subjects?

Maya: All subjects were scanned with the Butterfly tool by 13 trained health care practitioners, consisting of 7 Physicians and 6 Sonographers. Traditional biometry was also performed by those 6 sonographers at the same visit.

Sam: What about the patient demographics in this group of 111? Was it a diverse cohort?

Maya: The document provides a detailed breakdown. For race, there were 72 White subjects, 22 Black or African American, 5 Asian, 1 Native Hawaiian or Other Pacific Islander, 1 American Indian or Alaska Native, and 10 Other. For ethnicity, 23 were Hispanic and 88 were Non-Hispanic.

Sam: And body mass index? BMI is always a critical factor in ultrasound penetration and image quality.

Maya: They tracked that as well. There were 13 subjects with a BMI under 25, 64 subjects with a BMI between 25 and 29.9, and 34 subjects with a BMI over 30.

Sam: Alright, so how did they measure whether the AI tool was actually effective on these 111 subjects? What was the benchmark for success?

Maya: The clinical evaluation compared the AI tool's error against the pre-defined established clinical acceptable error of ultrasound measured gestational age. They broke this down into three time windows.

Sam: Let us hear the acceptable error windows first.

Maya: The acceptable error is +/-10 days for Week 16 to 21 6/7. It increases to +/- 14 days for Week 22 to 27 6/7. And it goes up to +/- 30 days for Week 28 to 37 6/7.

Sam: So as the pregnancy gets further along, the acceptable margin of error gets much wider, up to a full month in the third trimester. How did the Butterfly tool perform against those benchmarks?

Maya: They provide Bland-Altman bias and limits of agreement estimates. Let us look at the results for the iQ3 probe compared to the Last Menstrual Period, or LMP.

Sam: Go for it.

Maya: In the earliest window, Week 16 to 21 6/7, the bias was -0.50 days. The limits of agreement were -10.68 to 9.68. So that sits very closely around that +/-10 days acceptable error mark.

Sam: And the next window? Week 22 to 27 6/7?

Maya: For that window, the bias was 0.61 days, with limits of agreement from -8.84 to 10.05. That is well within the +/- 14 days standard.

Sam: And the final window, where the acceptable error is +/- 30 days?

Maya: For Week 28 to 37 6/7, the bias was -3.09 days, with limits of agreement from -20.03 to 13.85. Again, securely within the +/- 30 days bound.

Sam: Did the FDA document mention if the patient's BMI or the type of clinician holding the probe changed these results?

Maya: The summary notes that subgroup analyses looking at the mean GA error within various covariates, including gestational age window, sites, BMI category, and healthcare provider type, do not appear to show any appreciable performance difference compared to Biometry.

Sam: That is reassuring. It implies the AI performs consistently whether a physician or a sonographer is doing the blind sweeps, which is the entire value proposition of the product.

Maya: Exactly. Now, we need to touch on the regulatory pathway. This is a Class II device, cleared under Regulation Number 21 CFR 892.1550, which is for an Ultrasonic Pulsed Doppler Imaging System.

Sam: And since it is a 510(k) clearance, that means the FDA determined it is substantially equivalent to existing predicate devices. What were those predicates?

Maya: The primary predicate is the company's own Butterfly iQ3 Ultrasound System, cleared under K232808. The reference predicate is the Voluson Expert Series by GE Healthcare, cleared under K231965.

Sam: So they are demonstrating equivalence not just to their own hardware, but to the SonoBiometry feature on the Voluson, which is an AI-powered tool that automates standard fetal measurements during exams.

Maya: That is exactly right. The summary concludes that based on the intended use, technological characteristics, and performance testing, the Butterfly Gestational Age Tool does not introduce any new questions of safety or effectiveness.

Sam: And the FDA letter reminds the manufacturer that this substantial equivalence determination does not mean the FDA has evaluated contract liability warranties, and that all device labeling must be truthful and not misleading.

Maya: The document also lists several standards used for evaluating safety, including IEC 60601-1 for basic safety, and notably, IEC 60601-1-12 for medical equipment intended for use in the emergency medical services environment.

Sam: Testing for the emergency medical services environment makes a lot of sense for a portable, handheld device. It hints at where this could be incredibly useful in practice, even with the restriction against delivery planning.

Maya: Let us land this with a memorable takeaway. For clinicians and health-system buyers, this clearance represents a tangible shift in point-of-care ultrasound. You are using AI to estimate gestational age accurately through blind sweeps, eliminating the need for the operator to interpret live B-mode images.

Sam: But with the critical caveats that the independent clinical validation was small—just 111 subjects, mostly at the company's own sites—and the clear labeling that this is adjunctive information only, not for managing prenatal care or planning a delivery.

Maya: That is the perfect recap. As always, you can find a link to the full FDA 510(k) summary K252148 in our show notes to review the performance tables yourself.

Sam: And remember, this podcast is for informational purposes only and is not medical advice. Thanks for listening to this Company Spotlight on AI in Medicine - Smart Summaries.