Butterfly Network

A differentiated imaging device carries the business, and the company's own models are features on it. Butterfly's probe images the whole body with a semiconductor chip in place of piezoelectric crystals, and the device sells and images without any model. The learned components are real.

The Butterfly Gestational Age Tool (K252148, cleared 27 March 2026) feeds blind sweep cine loops into a deep learning model that outputs an estimated gestational age, and the Auto 3D Bladder Volume Tool (K200980, 2020) is a separately cleared measurement feature.

Around them sit two other layers: Compass AI, an enterprise workflow product that the company says is not intended to provide information for diagnostic or therapeutic decisions, and Butterfly Garden, a marketplace through which other developers deliver their own applications on the hardware. Clinical intelligence on a Butterfly device is therefore partly the company's own and partly its partners'.

The clinician acquires, reviews and decides, and one cleared tool moves the skill of view acquisition into the software. Most features assist: acquisition guidance, presets and automated measurement on images the operator reviews. The Butterfly Gestational Age Tool (K252148) changes the input rather than the decision.

The operator measures fundal height, which sets the number of sweeps, and follows on screen animations showing each sweep's path and probe orientation instead of a live B mode image; the model then returns an estimated gestational age.

The cleared indication limits it to a singleton intrauterine pregnancy presumed to be 16 to 37 weeks, used by qualified and trained healthcare professionals, and states that this adjunctive information is not intended to be used for prenatal management or delivery planning. That published boundary is a real oversight control.

The summary does not describe what the tool does with an inadequate sweep, whether it declines to return a result, or what confidence information reaches the operator, and an operator who cannot read standard views cannot judge a sweep's quality independently.

The company names the approach and the training data for its lead model, without a published version history. The 510(k) summary for the Butterfly Gestational Age Tool (K252148) says the collected blind sweep cine loops are input to a deep learning model that outputs gestational age, and that the model was developed on more than 100,000 cine loops, comprising millions of image frames, from thousands of patients at sites in the United States (the University of North Carolina) and Zambia, with clinical validation on a dataset independent of development.

The imaging approach, semiconductor ultrasound on a single probe, is described in detail on the company's pages. What is not published is the model's architecture beyond deep learning, a version history, or how model updates are communicated: capability reaches devices already in the field through cloud delivered software, so the model a clinician used last month may not be the one in use today.

The company's own clinical model is an in house, cleared build, and the layers around it are only partly disclosed. The Butterfly Gestational Age Tool is Butterfly's own deep learning model, cleared as a device (K252148) and developed on data from the University of North Carolina and sites in Zambia, so the model a clinician uses cannot be swapped without a regulatory consequence. Two other layers sit on the same platform.

Butterfly Garden delivers other developers' applications on the hardware; the company reports the ecosystem's size and that some partners hold their own clearances, but the partners are not listed, and what happens to a customer if a partner application is withdrawn is not described. And studies are stored in Butterfly's cloud, whose infrastructure provider and subprocessors are not named on the security page.

The strongest evidence is the lead model's clearance study, and it is the company's own. The 510(k) summary for the Butterfly Gestational Age Tool (K252148, March 2026) reports a prospective clinical evaluation of 111 pregnant participants at 16 to 37 weeks, at four US sites (two Butterfly offices, Thomas Jefferson University in Philadelphia and a primary care practice in Flagstaff), with 13 operators (7 physicians and 6 sonographers), on both the iQ+ and iQ3 probes, against gestational age from the last menstrual period.

Mean bias ranged from -3.1 to 2.6 days across gestational age windows and probes, with 95 percent limits of agreement widest in the 28 to 37 week window (-18.2 to 19.2 days on the iQ+, -20.0 to 13.9 on the iQ3). The summary also reports the development study at the University of North Carolina: a mean absolute error of 3.9 days for the model against 4.7 days for standard biometry.

Customer results are named: the University of Rochester reported a 116 percent increase in point of care ultrasound revenue after a systemwide deployment, an operational figure without method, and Indiana University School of Medicine runs a one probe per student program, which is scale of use rather than evidence of benefit. No independent or peer reviewed evaluation of the cleared tool as sold, and no patient outcome study, is on the company's pages.

Studies acquired on the probes are stored in Butterfly's cloud, and what may be done with that archive is not published. The security page describes a program run against NIST, SOC 2, HIPAA and HITECH frameworks and a global privacy program that meets the GDPR, which are general assurances.

Nothing on the security page or the home page says whether customer images are used to train, tune or evaluate the company's models, whether partner developers can reach them, how long studies are retained, or whether images are de identified before any such use.

The question is sharper because the cleared gestational age model was developed on research data (the University of North Carolina and sites in Zambia), and the company now holds a growing archive of clinical ultrasound from hospitals, medical schools, veterinary practices and global health programs. Images are captured on phones and tablets before reaching the cloud, and local retention on those devices is not described. Safety engineering, meaning guardrails and how a safety event is handled, is not published. The privacy notice linked in the site footer is where a training use position would be stated.

HIPAA is named as a framework the company works to, with no business associate position published. The security page says Butterfly enforces and monitors its cybersecurity operations using frameworks including HIPAA and HITECH, alongside NIST and SOC 2. That is a compliance claim; no business associate status, agreement template or scope is stated with it. The privacy notice linked in the site footer is where the company's role for the studies it stores would be set out.

Studies are stored in Butterfly's cloud on behalf of the organizations that acquire them, which is the relationship a business associate agreement governs. Two features of this product widen the question. Acquisition happens on a clinician's phone or tablet, often a personal device, so protected health information is created on endpoints outside a hospital's management. And individual clinicians and small practices can buy probes through the online store, where an agreement may never be negotiated and the terms of use become the whole arrangement.

A SOC 2 Type 2 report is named on the company's own security page, and the artifact is not open to download. Butterfly states that it holds a SOC 2 Type 2 attestation report issued by an independent third party auditor, with a scope encompassing the iQ+ ecosystem across the entire organization, and that its program follows NIST, SOC 2, HIPAA and HITECH frameworks; security questions go to a named security mailbox.

The report's period, and whether its scope reaches the iQ3 and the gestational age tool, are not stated. The page does not describe penetration testing, a vulnerability disclosure program or device security. The device layer is its own surface: these are connected medical devices pairing with personal phones and tablets over consumer networks in hospitals, homes and clinics abroad, and firmware update integrity, pairing and a lost probe are questions separate from cloud security. The partner marketplace adds a third surface.

Each device and the lead AI tool carry their own 510(k), and the list can be checked in the FDA database. Butterfly Network's clearances are K163510 (the original Poseidon ultrasound system, 6 September 2017), K200980 (Auto 3D Bladder Volume Tool, 11 June 2020, 21 CFR 892.1560), K202406 (Butterfly iQ, 16 September 2020), K220068 (Butterfly iQ and iQ+, 31 March 2023), K232808 (Butterfly iQ3, 4 January 2024, 892.1550) and K252148 (Butterfly Gestational Age Tool, 27 March 2026, 892.1550, product code IYN).

K252148 covers an adjunctive gestational age output for a singleton intrauterine pregnancy presumed to be 16 to 37 weeks, used by qualified and trained healthcare professionals, and states that it is not intended for prenatal management or delivery planning; its predicate is the iQ3 (K232808), with a GE Voluson system (K231965) as reference device. The company describes it as the first cleared blind sweep tool, which is its own claim.

Compass AI is stated not to be intended to provide information for diagnostic or therapeutic decisions. Partner applications on Butterfly Garden carry their developers' own clearances, not Butterfly's. New capability reaches fielded devices through software updates, so a given probe's regulatory state depends on its software version and market.

The clearance study for the lead model names its subgroups and states a conclusion, without the per group figures. The 510(k) summary for the Butterfly Gestational Age Tool (K252148) reports accuracy by gestational age window with counts (28, 37 or 38, and 45 participants per window on each probe) and states that analyses by gestational age window, site, BMI category and provider type do not appear to show any appreciable performance difference.

It gives the participants' demographics, including race and ethnicity, but no results by race, ethnicity or BMI. The population question is real for this product. Ultrasound image quality varies with body habitus, and a dating model carries whatever reference it learned: this one was developed on data from the United States and Zambia and validated on 111 participants at four US sites, while the company markets it for global health settings. The widest limits of agreement in the summary fall in the 28 to 37 week window.

A published error rate with its method and denominator is on the record, and nothing stands behind it contractually. The 510(k) summary for the Butterfly Gestational Age Tool (K252148, decision 27 March 2026) is Butterfly's own filing: a prospective evaluation of 111 pregnant participants at four US sites, against gestational age from the last menstrual period, reporting mean bias of -3.1 to 2.6 days across gestational age windows and the two probes, with 95 percent limits of agreement as wide as -20.0 to 19.2 days in the 28 to 37 week window.

It is a result on a test set for the cleared indication (a singleton pregnancy at 16 to 37 weeks, adjunctive, not for prenatal management or delivery planning), not performance in use. It covers the gestational age tool only: not the bladder volume tool, Compass AI or partner applications, whose performance sits with their own developers. No warranty covering model output, accuracy commitment, indemnity or remediation route is published. The purchase channel matters here: individual clinicians can buy probes through the online store, where standard terms rather than a negotiated agreement govern.

Integration is claimed through a separate workflow product, with no record system named. Getting studies from a handheld probe into the hospital record is the known weak point of point of care ultrasound: an exam that never reaches the record is hard to bill, hard to audit and unavailable to the next clinician.

Butterfly sells Compass AI to organizations as enterprise workflow software for point of care programs, and describes it as not intended to provide information for diagnostic or therapeutic decisions. No EHR, PACS vendor, interface standard in production or marketplace listing is named on the company's main pages, and nothing describes whether orders and results flow both ways or how an exam reconciles to an order.

One hosted option is stated, and where it runs is not. Acquisition runs on a phone or tablet, studies move to Butterfly's cloud, and new capability reaches fielded devices through cloud delivered software. No on premises or customer hosted option is described, and the security page names no infrastructure provider or region.

The company sells across the Americas, Europe and Asia, holds a European conformity mark and states a global privacy program that meets the GDPR, so where a given customer's images rest is a live question. Delivery by software cuts both ways: a probe gains capability without a hardware change, and its behavior is set by a remote software state the clinician does not control.

A device price and a route to buy online are published; the full membership ladder is not on the home page. The home page lists the Butterfly iQ+ from $2,699 and shows financing (a $3,119 purchase at about $87 a month over 36 months at 0 percent APR), and the site footer links a pricing page and an online store, so an individual buyer can price and buy a probe without a sales conversation. Compass AI, the enterprise workflow product, is sold to organizations with no rate published on the home page.

Coverage is named broadly, and validation stands behind one part of it. A single semiconductor probe images across depths that usually take several transducers, so abdominal, lung, cardiac, musculoskeletal, vascular and obstetric imaging run on one device, and the company sells into hospitals, medical schools, individual practices, veterinary practices, global health programs and home and community care.

The part with validation behind it is the gestational age tool: a singleton pregnancy at 16 to 37 weeks, evaluated prospectively at four US sites with physicians and sonographers as operators (K252148). The company targets that tool at settings with no sonographer, including global health programs, while the clearance study was run in the United States with trained operators.