Clinical AI & imaging · Opportunistic screening
Bunkerhill Health
Nine FDA-cleared imaging and ECG algorithms, most built inside academic medical centers, wrapped in a platform that acts on what they find. The algorithms are cleared devices. The agents around them are not.
The algorithm that worked and went nowhere
In 2017 Nishith Khandwala was a computer science graduate student at Stanford, in AI labs that physicians from the medical school kept wandering into with ideas. One ran preventive cardiology. His job was to stop the second heart attack, and reading back through a new patient’s chart he kept finding the same thing: years earlier that patient had come through the hospital for a car accident or a pneumonia, got a CT scan, and on it was calcium in the coronary arteries. Nobody acted. The patient had to have a heart attack before anyone sent them to a cardiologist.
The lab built the algorithm. It worked. They published. “We made some incredibly bold claims about how this is gonna save hundreds of thousands of lives,” Khandwala later told Bill Russell on This Week Health. Two years passed. “We had yet to impact a single patient, to touch a single patient, let alone save lives, and it was very embarrassing.” He was ready to leave healthcare. Then in 2020 his father had a heart attack, and in the emergency room a cardiologist began explaining what coronary calcium meant.
Why the scan just sits there
The clinical idea is opportunistic screening: reading a scan ordered for one reason to answer a second question it happens to contain. Millions of chest CTs are done in the United States every year, and each already holds information about the heart that nobody went looking for. No radiology department has the staff to review them all by hand, and once a finding is surfaced somebody still has to chase the patient. At the University of Texas Medical Branch, early results reported by the health system found 11.7 percent of routine chest CT patients had unaddressed, clinically significant coronary artery calcium. Khandwala puts the second half in arithmetic: a system running 100,000 scans might flag 9,000 at-risk patients, and handing a care team 9,000 names solves nothing.
Two engineers and a hospital that never existed
Khandwala co-founded the company with David Eng, a fellow Stanford computer scientist and now its chief technology officer. Both came out of the Stanford Center for Artificial Intelligence in Medicine and Imaging, and before the company existed they helped organize one of the first prospective multi-institution randomized trials of a medical AI model, a 1,903-patient skeletal age study registered by Stanford as NCT03530098.
The early business was not a product but a consortium. A 2020 note from the University of Kentucky College of Medicine describes the offer: a radiologist with an algorithm could validate it on outside data through Bunkerhill, publish, and decide later whether to commercialize. “The BunkerHill platform and all of its services are free of cost for all R&D partners,” it says. Advisers included Curtis Langlotz, Matt Lungren and Ross Filice. The consortium is still visible in the literature, with Bunkerhill-affiliated authors on a 2026 cardiac MRI paper in Nature Biomedical Engineering.
The name comes from a television show. “We are not a Boston company, even though people might think or find that as a reference,” Khandwala said. Pure Genius was cancelled midway through its first season, but it was set in a fictional hospital called Bunker Hill where the morning’s research reached patients in the afternoon. That speed was the point.
From one algorithm to nine, and then to agents
The first clearance landed in April 2023, a coronary calcium device called iCAC, filed as Bunkerhill, Inc. Eight followed as BunkerHill Health: gated coronary calcium, aortic valve calcification, bone mineral density, abdominal aortic quantification, ECG-based ejection fraction, mitral annular calcification, and contrast-enhanced calcium in March 2026. Several were built by academic teams and validated on consortium data before Bunkerhill filed.
Around that portfolio the company built Carebricks. A large language model reasons over the chart, the literature and a hospital’s own protocols, calls the cleared algorithms as tools, then acts: places a referral, messages a clinician, works a prior authorization. In April 2026 a company-led submission produced a national Medicare billing code, HCPCS G0680, for algorithmic analysis of coronary and aortic valve calcium on chest CT, paid at $15.50.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| Nine 510(k) clearances | K230223 (Apr 25, 2023), K240369 (Sep 27, 2024), K243229 (Jan 27, 2025), K242295 (Apr 8, 2025), K243779 (Jul 1, 2025), K250649 (Sep 19, 2025), K250662 (Oct 3, 2025), K260166 and K260167 (both Mar 6, 2026). All Class II, all substantially equivalent. | Public record |
| Who holds them | Eight belong to BunkerHill Health of San Francisco. The first, K230223, belongs to Bunkerhill, Inc. of Palo Alto. Both used the same outside counsel at Hogan Lovells, and K230223 is the iCAC product still sold. A search on the current name alone returns eight. | Public record |
| What the clearances permit | Each indications for use statement reviewed carries the same limit: “Results of the subject device are not intended to be used on a stand-alone basis and are solely intended to aid and provide information to the physician.” | Public record |
| What is not cleared | No 510(k) found covers the Carebricks platform, its language model reasoning layer, or the agents that place orders and contact patients. The cleared devices are the individual algorithms, while marketing describes agents that automatically initiate cardiology referrals. | Claim broader than the clearances |
| Randomized evidence | NOTIFY-PICTURE, Circulation, Jan 20, 2026, with Eng and Khandwala as co-authors. The trial, NCT05588895, enrolled 202 participants under Stanford sponsorship. Fierce Healthcare separately reports a 44-percentage-point rise in statin prescribing in the earlier NOTIFY-1 study, NCT04789278, 173 participants. | Public record |
| Algorithm validation | Abdominal Radiology, 2026: multicenter validation of the bone density model in 504 participants. Area under the curve 0.96 internally, 0.82 externally. Four Bunkerhill-affiliated authors. | Public record |
| Not on the record | No ClinicalTrials.gov study lists Bunkerhill as sponsor or collaborator; every relevant trial is Stanford-registered. No NIH or NSF award under the company name. No FDA recalls and no MAUDE adverse event reports as of Sep 22, 2026. | Not found |
| Independent distribution | HeartSciences Inc. (Nasdaq: HSCS) told the SEC in its Form 10-K filed Jul 23, 2026 that MyoVista Insights version 1.3 opened an AI-ECG marketplace “beginning with an FDA-cleared AI-ECG model from Bunkerhill Health.” | Public record |
| Named deployments | Stamford Health, reported by CT Insider in Oct 2024 with more than 2,000 chest CTs run. Cleveland Clinic, Mayo Clinic, Intermountain, UTMB, Ballad, Sentara and Endeavor are company-named. | Mixed |
Read plainly: the regulatory record is unusually deep for a company this size, and it is real. What it does not contain is a trial of the agents. Every randomized result belongs to a Stanford-sponsored study of notification, not of an autonomous workflow, and the cleared labels say the outputs are assistive and the action is the physician’s call. That distance is what the public record has not tested. One repeated detail runs the other way: CT Insider called the software “FDA-approved,” but these are 510(k) clearances, a lower bar.
What to watch
- Any prospective study of a Carebricks agent, rather than a notification algorithm, registered with Bunkerhill as sponsor.
- Uptake of HCPCS G0680. At $15.50 per assessment it covers cost rather than creating margin, which the company says was the intent.
- Whether ECG-EF earns revenue through the HeartSciences marketplace, the one channel with quarterly public disclosure.
- How the FDA treats a language model layer that calls cleared devices as tools. No clearance found covers it today.
In their words
“We had yet to impact a single patient, to touch a single patient, let alone save lives, and it was very embarrassing.”
Nishith Khandwala, co-founder and CEO, This Week Health, 2025 · Interview
“Great ideas are everywhere, we just don’t translate them into reality.”
Nishith Khandwala, to Fortune, July 2026 · Independent
“So we license AI algorithms that researchers have created at these health systems and bring them all to our platform.”
Nishith Khandwala, This Week Health, 2025 · Interview
“You leave the AI aside, you leave the dollars aside, this is such a wonderful application of new ideas that are just helping patients live healthier lives.”
Nishith Khandwala, to Fierce Healthcare, May 2026 · Independent
“We don’t need superintelligence to solve our biggest problems. We need average intelligence.”
Peter McCaffrey, MD, chief AI officer, UTMB, to Fortune, July 2026 · Independent
“The ability to assess bone health without additional scans, radiation exposure to patients, and cumbersome manual analysis needed from radiologists represents a paradigm shift in how we think about preventative care.”
Akshay Chaudhari, Stanford Radiology, to Diagnostic Imaging, April 2025 · Independent
“We have a way to go, but the initial results are exciting.”
David Hsi, MD, chief of cardiology, Stamford Health, to CT Insider, October 2024 · Independent
Related companies
Sources
- Public recordFDA 510(k) database, nine clearances to Bunkerhill, Inc. and BunkerHill Health
- Public recordFDA clearance letters and indications for use, K230223, K250662, K260166
- Public recordFDA recall, MAUDE and device listing searches
- Public recordThe NOTIFY-PICTURE Trial
- Public recordMulticenter validation of a deep learning osteoporosis algorithm on abdominal CT
- Public recordClinicalTrials.gov NCT03530098, NCT04789278, NCT05588895
- Public recordHeartSciences Inc. Form 10-K, CIK 0001468492
- Public recordA generalizable deep learning system for cardiac MRI
- IndependentFDA Clears AI-Powered Bone Mineral Density Assessment from Abdominal CT Scans
- IndependentBunkerhill Health raises $55 million to put AI agents to work inside hospitals
- IndependentHealth systems are exploring AI-powered cardiac risk screening
- IndependentThis hospital is CT’s first to use AI to detect coronary artery disease
- InterviewShrinking the Gap Between Idea and Reality
- IndependentWang Arranges Agreement with AI Company BunkerHill
- CompanySeries B announcement and the AVC, BMD and AAQ releases
- PartnerHeartSciences Launches MyoVista Insights Version 1.3
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
