Knit Health has launched with $11.6 million in seed funding to build the first large clinical behavior model, a healthcare AI trained not on medical literature but on the actual decisions of doctors and care teams. The model is being trained on anonymized electronic health records from more than 130 million patients across 30 major U.S. health systems, roughly a third of the American population.
Training on human decisions, not textbooks
The company was founded by UC Berkeley researchers Jonathan Kolstad, Maya Petersen, Jonas Knecht, and Ted Robertson. It grew out of a six-year effort to unlock behavioral signals hidden in the audit logs of electronic medical records-millisecond-level records of every action taken by clinicians. Kolstad, an economist whose work examines how healthcare systems function, saw those logs as a record of clinical intelligence that exists nowhere else.
"This form of intelligence is a fundamentally different way to build AI," said Kolstad, who serves as CEO. "What we are saying is that if a clinician or provider team knew everything about a patient up to this point, what might they do next?"
While many AI models rely on medical literature and guidelines, AI for Healthcare built this way learns from behavior. An early test showed the model's potential: after seeing a chest CT scan followed by a brain scan for a patient who appeared to be having a heart attack, the model began assigning higher probabilities to stroke-related conditions. It had picked up a clinical pattern without being told what to look for.
Adding causal reasoning to clinical behavior
A chance meeting at a Berkeley brew pub with Dr. Maya Petersen, an MD/PhD and expert in causal inference, shifted the model's direction. Petersen pointed out that learning what doctors do is not the same as knowing what is right. Her insight led to a framework that could compare a typical doctor's decisions to those made at the best hospitals, isolating what the top performers do differently. Petersen became a co-founder and chief scientist.
Ted Robertson, the company's COO and executive director of the Center for Healthcare Marketplace Innovation, rounded out the founding team. "Knit is a proof point that there is groundbreaking research at Cal that can and should make the world better," he said.
Pilots and real-world performance
Data from Providence Health, one of the nation's largest health systems, showed the model could predict 60% to 80% earlier in an emergency department visit whether a patient would be admitted. The model performed at the level of clinicians themselves, but faster. This summer, the model is running at several pilot sites. One will determine if emergency department patients need hospital admission or can be treated elsewhere. Another will handle specialty referrals, routing patients to the right specialist and directing others back to primary care.
The research behind the model is detailed in a paper titled "Deep Causal Behavioral Policy Learning: Applications to Healthcare," reflecting the AI for Science & Research work at UC Berkeley. Knit Health was incubated at the Center for Healthcare Marketplace Innovation, a joint center spanning Haas and the College of Computing, Data Science, and Society, and is the first startup to emerge from it.
Why this matters for healthcare professionals
The model runs invisibly inside existing workflows, asking of every patient what should happen next given everything known up to that moment. For clinicians, that means decision support that does not require reading new guidelines or sifting through literature. It can reduce the cognitive load of routine decisions, free beds in emergency departments, and ensure patients see the right specialist sooner. The team's longer-term goal is to make the expertise of the country's most skilled doctors available to all patients, regardless of location or timing.
Your membership also unlocks:
