Clinical AI · Emergency department diagnostics
Biocogniv
A Vermont company that ran machine learning over routine blood panels to flag COVID-19 and sepsis at the emergency department door. In April 2026 an electric aircraft maker bought it, for the software tooling it built along the way.
The patients who look fine
Sepsis is easy to name afterward and hard to see while it happens. The classic bedside trigger, the Systemic Inflammatory Response Syndrome criteria, is a checklist of temperature, heart rate, breathing rate and white cell count. Plenty of people about to crash do not trip it on arrival.
In 2021 Biocogniv put a number on that gap. With Frank Peacock of Baylor College of Medicine, it trained a gradient boosting model on 679,940 emergency department encounters from 62 US hospitals, then tested it on 361,305 encounters at 31 hospitals it had never seen. Inside that test set sat 2,681 encounters that ended in death, a long ICU stay or a sepsis discharge code, and that showed none of the classic signs at the door. The model flagged 1,364 at presentation: 50.9 percent of the patients nobody was alarmed about yet.
The inputs were not exotic: a complete blood count, a metabolic panel, vitals. Tests a hospital already runs, that come back in under an hour.
The result was published as a conference abstract. It has never appeared as a full paper.
Why the bar was low, and why that mattered
Biocogniv was building into a market where the incumbent had just been embarrassed. In June 2021 a University of Michigan team externally validated Epic’s sepsis model, then running at hundreds of US hospitals. Against a sepsis definition closer to the one Medicare and the CDC use, it sorted patients correctly 63 percent of the time, against the 76 to 83 percent on Epic’s own information sheet.
A 2024 validation of that model in two county emergency departments found a sensitivity of 14.7 percent within a six hour window and a median lead time of zero minutes. A 2025 review of 91 sepsis prediction studies found median discrimination falling to 0.783 under full-window external validation.
Against that backdrop, 0.94 on an external hospital set was a real claim. It was also retrospective, measured against a composite proxy endpoint rather than an adjudicated diagnosis, on records rather than patients.
The founder
Artur Adib took his PhD at Brown and was a faculty member at the National Institutes of Health. He then went commercial: pre-IPO engineer at Twitter, principal engineer at Magic Leap, and founding team member and head of software at BETA Technologies, the Vermont electric aircraft company.
He started Biocogniv in 2019 and took it through Y Combinator’s Summer 2020 batch. The team stayed small and local: University of Vermont data scientists, an ICU surgeon with a computational biochemistry PhD, a former Goldman Sachs technology VP as CTO. Y Combinator lists six people. The advisory bench was heavier than the payroll, including two former Labcorp executives.
COVID first, then sepsis, then aircraft
The first product was AI-COVID. In December 2020 the Journal of Medical Internet Research published a model built on 192,779 patients from 66 US hospitals that scored COVID-19 probability from routine bloods alone, area under the curve 0.91. The point was throughput: PCR turnaround in emergency departments ran above 24 hours.
By 2022 the platform had a second life as plumbing. The University of Texas Medical Branch used interpAssist to attach expert commentary to COVID results at scale, and won a Gage Award. In March 2022 the company closed a $6 million seed led by Breyer Capital, with Y Combinator, 1984 Ventures, Inflect Health, Promus Ventures and The Fund at Hula, plus angels including Boston Scientific co-founder John Abele and former IBM chief executive Sam Palmisano. Pilots ran at North Country Hospital, San Joaquin Hospital and UTMB.
In April 2023 came the announced FDA Breakthrough Device Designation for Sepsis aiMarker. That October the company certified to ISO 13485:2016 and filed its first Form D, reporting $10.74 million sold. Then the record thins: no further clinical results, no FDA submission, no new customers.
What surfaced next, in January 2026, was a post about software engineering. Biocogniv had built a multi agent system it called SAGE that produced the requirements, hazard analyses, tests and traceability matrices IEC 62304 and ISO 13485 demand. Seventeen agents, two engineers, artifacts that survived a certification audit, and a timeline the company put at weeks instead of a projected 12 to 18 months. Three months later BETA bought the company.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| Peer-reviewed COVID paper | JMIR, Dec 2, 2020. 192,779 patients, 66 US hospitals, AUROC 0.91. At a cutoff of 1.0, sensitivity 95.9% and specificity 41.7%; at 2.0, 92.6% and 59.9%. Two authors disclose ownership. The paper names its own weakness: prepandemic negative controls, so time period bias cannot be excluded. | Public record |
| Sepsis evidence | One conference abstract, Annals of Emergency Medicine, Oct 2021. C-statistic 0.94 externally, against 0.70 for qSOFA and 0.84 for SIRS. Endpoint: a composite of mortality, ICU stay over two days, or sepsis discharge codes. No full paper found. | Public record |
| Trials and clearances | Zero studies on ClinicalTrials.gov. No 510(k), PMA, device registration or recall in openFDA. Searched Sep 23, 2026. | Not found |
| Breakthrough designation | Announced Apr 7, 2023 for Sepsis aiMarker. FDA publishes no list of designated devices, so it cannot be checked. The company’s own Sep 2024 post dates the same designation to 2022. | Company dates differ |
| Prediction window | The site said 72 hours in advance in December 2024 and 24 hours in May 2026. Nothing published explains the change. | Claim changed |
| Federal grants | NSF SBIR Phase I 2014934, $209,881, and NIH R43 HL154941, $299,749. Total $509,630. Both funded a pulmonary embolism test using AI on blood spectra. No Phase II at either agency, and no such product appeared. | Public record |
| Deployments | UTMB Health ran more than 325,000 test interpretations over 13 months, corroborated by Fast Company in 2022. No customer named since. | Independent |
| Quality system | ISO 13485:2016 certified Oct 2023, maintained Jul 2024. Certifying body not named. | Company-stated |
| Outcome | BETA Technologies Form 10-Q, quarter ended Jun 30, 2026: an asset acquisition of stock valued at $10,797,000, net cash of $2,561,000 and a $1,645,000 holdback. The full $15,003,000 was expensed as in-process research and development. | Public record |
| Website | biocogniv.com does not resolve as of Sep 23, 2026. Registration runs to May 2028, but the zone returns SERVFAIL. Last archive capture May 18, 2026. | Public record |
Read plainly: one solid peer-reviewed result, on COVID, in 2020, and one striking sepsis result that never left abstract form. No FDA filing, no registered trial, no paying hospital named after 2022. What a buyer paid $15 million for was machinery for producing regulated software evidence fast. The diagnostics were why that machinery existed. They were not what sold.
What to watch
- Whether the sepsis work is ever published in full. It has sat in an abstract since 2021.
- Where the Sepsis aiMarker regulatory file goes. BETA expensed the purchase as in-process research and development with no alternative future use, not how a buyer accounts for finishing a device.
- BETA’s 10-K for 2026, and whether the $1,645,000 holdback is paid on schedule.
- Whether SAGE shows up in aircraft certification work. That is the thesis BETA bought.
In their words
“Nine months into this pandemic, we now have a better understanding of how to care for patients with COVID-19”
Timothy Plante, MD, University of Vermont, Vermont Business Magazine, 2020 · Independent
“In hospital settings, every hour of delayed care can lead to long-term morbidity and, in some cases, death.”
Artur Adib, PhD, founder and CEO, seed release, 2022 · Company release
“We can detect if a patient may be highly likely to deteriorate into sepsis before it can currently be detected,”
Jabez Boyd, COO, Burlington Free Press, 2022 · Independent
“With Biocogniv’s platform, we were able to scale interpretation and guidance for COVID and coordinate everyone during this time of unprecedented uncertainty.”
Peter McCaffrey, MD, UTMB, Fast Company, 2022 · Independent
“Sepsis is something we can know occurs with certainty after the fact, but when it’s unfolding, it’s often unclear whether a patient has sepsis or not,”
Karandeep Singh, MD, University of Michigan, Michigan Medicine, 2021 · Independent
“Our compliance and audit success validates the core hypothesis: agentic AI can generate quality system evidence faster and more consistently than human teams alone.”
Artur Adib, PhD, CEO, company post, January 2026 · Company post
Related companies
Sources
- Public recordBETA Technologies Form 10-Q, Note 3 and Item 2
- Public recordForm D, Biocogniv Inc., CIK 0001997960
- Public recordNSF award 2014934 and NIH RePORTER 1R43HL154941-01
- Public recordopenFDA device and ClinicalTrials.gov queries, no records
- Public recordRDAP and DNS check, biocogniv.com
- Public recordMachine Learning Tool to Rule Out COVID-19 Using Routine Blood Tests
- Public recordAI Occult Sepsis Detection in the Emergency Department
- IndependentBiocogniv, UVM partner on AI to help labs rule-out COVID-19
- IndependentA Burlington company using AI to predict sepsis just raised $6 million
- IndependentMeet the startups using AI to help doctors fight burnout
- IndependentPopular sepsis prediction tool less accurate than claimed
- IndependentExternal validation of the Epic sepsis predictive model
- IndependentSystematic review of sepsis real-time prediction models
- IndependentCompany listing, batch S20, status Acquired
- IndependentPortfolio page naming BETA Technologies as buyer
- CompanyBiocogniv Closes $6M in Seed Funding Led By Breyer Capital
- CompanyBiocogniv Receives FDA Breakthrough Designation for Sepsis aiMarker
- CompanyBiocogniv Powers UTMB’s Award-Winning COVID-19 Program
- CompanyBiocogniv Achieves and Maintains ISO 13485:2016 Certification
- CompanyHow Biocogniv’s Agentic AI Process Increased Development Velocity
- CompanyHome and about pages, archived
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
