Clinical software · Hospital deterioration

AgileMD

An early warning score grown out of a decade of University of Chicago research, cleared by the FDA in June 2024 and now running inside the electronic health record at Yale New Haven Health, UNC Health and others.

One hour on a ward

Dana Edelson was in her second year of residency at the University of Chicago Medical Center, the most senior doctor on a slammed floor. One man was bleeding heavily. Another patient was an eighteen-year-old who looked fine, still joking.

She worried about the bleeding. When she got back to the teenager he was breathing 40 times a minute and needed a ventilator. Before he could be ventilated he went into cardiac arrest and died. Septic shock had taken about an hour.

“There was data that could have predicted it; I just didn’t see it,” Edelson told Crain’s Chicago Business in 2024. “I didn’t know what to look for.”

AgileMD exists to make that sentence untrue. Fifteen years on it has what most clinical AI vendors do not: an FDA clearance, a comparison it did not run itself, and a health system willing to put a mortality number on the record.

Why the alarms do not help

Deterioration on a ward is common: up to 5% of hospitalized patients by the peer-reviewed paper behind the model, approximately ten percent by AgileMD’s own announcement. Delay costs. The company cites a 3% rise in the odds of hospital death for every hour an ICU transfer is late.

Sepsis is the biggest single driver: more than 1.7 million adults a year in US hospital care, over 250,000 deaths, $40 billion in Medicare spending. Six medical societies argued in 2023 that the federal sepsis bundle, SEP-1, should be retired rather than tied to payment, because its adoption raised antibiotic use without lowering mortality.

So hospitals bought sepsis alerts, and the alerts fired constantly. Most watch fixed thresholds on single values, so a patient who runs a low blood pressure trips them all day.

“Most tools are set to fire around fixed parameters, so they over-alert on stable patients who just run low or high,” Deborah Rhodes, chief quality officer at Yale New Haven Health, told Becker’s in 2025. Alarm fatigue follows, which is why accuracy matters more than volume.

Chocolate and peanut butter

Borna Safabakhsh came at it from the business school. As an MBA student at Chicago Booth he was working on clinical pathways, the guidance that usually lives on a laminated sheet at a nursing station and goes stale. In 2011 the team won the university’s Edward L. Kaplan New Venture Challenge, went through Y Combinator that summer as Agile Diagnosis, then Rock Health’s second batch. TechCrunch covered the beta in April 2012: a tablet app of visual guidelines, carrying the $150,000 convertible note SV Angel and Start Fund offered the whole batch. A third co-founder, Dr. Scott Stern, was assistant dean of technology and innovation at UChicago’s Pritzker School of Medicine.

Edelson was on a separate track in the same university, building eCART, the electronic Cardiac Arrest Risk Triage score, out of health record data. Around 2014 the medical center put the first version into quiet use in the background of an iPad. It crashed the servers on day one.

That same day she sat in a meeting where Safabakhsh was pitching his pathways tool. Her score found the patient; his product told the team what to do next. “It was the chocolate to my peanut butter,” she said. Her company, Quant HC, merged with AgileMD in October 2019.

From app to regulated device

COVID paid for the regulatory work. In May 2021 BARDA, the federal biodefense agency, put $750,000 into eCART through its DRIVe program, money that would let the company “finalize regulatory work, pilot in 10 hospitals and rapidly make available the solution to many others,” Edelson said. Four months later the National Institute of Biomedical Imaging and Bioengineering added $1,995,966.

AgileMD filed its 510(k) in September 2023. The FDA cleared the eCARTv5 Clinical Deterioration Suite on June 21, 2024. Four months later Yale published the comparison that mattered most.

What is proven, and what is still claimed

Read plainly: this is one of the better evidenced products in clinical AI, and the evidence still does not settle the argument by itself. The model wins a large multicenter accuracy test, but the strongest outcome data comes from a pre and post study and from one health system talking to a trade publication. The co-founders are authors on the studies that judge their product, which they disclose. The trial that would settle it finishes at the end of 2026.

What to watch

• Results from NCT05893420, with primary completion estimated for December 31, 2026.
• Whether the Yale mortality result reaches peer review, and whether a second health system publishes one.
• A pediatric clearance. The label stops at adults while hospitals run the tool on children under their own authority.
• A replacement for the NIH award that ended in March 2025, and any first Form D.

In their words

AgileMD funding profile on HVCFHealthcare Venture Capital Fund

Related companies

• Medisas: Also in Healthcare IT. A hospital doctor who could program set out to replace the scrap of paper that care teams pass between shifts
• Sepsis Scout: Also in Healthcare IT. A Y Combinator company that put a deep learning sepsis risk score on an off-the-shelf wearable and aimed it at cancer patients recovering at home
• Ribbon Health: Also in Healthcare IT. An API built to answer a question the American health system keeps getting wrong: is this doctor real, in network, and reachable at the number listed
• Comprehend Systems: Also in Healthcare IT. A Y Combinator company that sold software for watching clinical trials while they were still running

Sources

1. Public recordFDA 510(k) K233253, eCARTv5 Clinical Deterioration SuiteFDA · Jun 21, 2024
2. Public recordMulticenter Development and Prospective Validation of eCARTv5Crit Care Explor · Apr 2025
3. Public recordImpact of a Machine Learning Early Warning Score on Hospital MortalityCrit Care Med · Sep 2022
4. Public recordEarly Warning Scores With and Without Artificial IntelligenceJAMA Netw Open · Oct 2024
5. Public recordClinicalTrials.gov NCT05893420 and NCT06771830NLM
6. Public recordNIH RePORTER award 1R44EB030955-01A1NIBIB · FY2021
7. Public recordBARDA and AgileMD partner on an EHR integrated toolmedicalcountermeasures.gov · May 5, 2021
8. IndependentUChicago-born early warning device for sepsis gets FDA nodCrain’s Chicago Business · Jul 8, 2024
9. IndependentHow Yale New Haven Health uses eCARTBecker’s Hospital Review · Aug 19, 2025
10. CompanyAgileMD receives FDA clearance for eCARTBusiness Wire · Jun 25, 2024