UC Berkeley researchers and entrepreneurs have built an artificial intelligence system that is trained not on textbooks or papers but on the real-time actions of clinicians, a move they say could embed what they call clinical intelligence into health systems.
From audit logs to clinical behavior model
Experts across health economics, computational health, biostatistics and behavioral design at the Haas School of Business and the university’s public health department pooled their work and data to create what they describe as the world’s first large clinical behavior model (LCBM). The new AI was developed by a start-up called Knit Health, founded by Professor Jonathan Kolstad, doctoral researcher Jonas Knecht, entrepreneur Ted Robertson and epidemiologist Dr. Maya Petersen.
The team based the model on extensive audit logs from electronic medical records — detailed, millisecond-by-millisecond records that document every action taken by physicians, nurses and care teams. Those logs, the researchers argue, encode clinicians’ reasoning in a way that is absent from formal clinical literature and conventional AI training sets.
“when a doctor makes a decision, they are revealing what’s in their brain. That behavioral record contains a form of clinical intelligence that exists nowhere else,”
The founders say this approach lets the model learn how clinicians actually practice medicine, rather than how medicine is described in textbooks or journals.
How the model differs from other healthcare AI
Most existing healthcare AI has been trained on clinical notes, imaging, structured data or the corpus of medical research. The Knit Health team emphasizes that the LCBM's training data reflect behavior — sequences of clicks, orders, and other interactions within electronic health record (EHR) systems — which they believe capture decision-making patterns and tacit knowledge applied in day-to-day care.
Proponents say this could improve tools that guide clinicians and patients to the right care at the right time by aligning recommendations with how teams actually behave in real settings. The group began collaborating after conversations in October 2023 between Kolstad and Knecht, and less than three years later the company and model were launched.
Founding team and roles
Key participants in the effort bring distinct disciplinary perspectives, combining health economics, clinical epidemiology and computational methods. The public presentation of the team highlights cross-disciplinary work as a central element of the project's design and validation strategy.
| Person | Role/Affiliation |
|---|---|
| Jonathan Kolstad | Professor, Haas School of Business; health economist |
| Jonas Knecht | Doctoral researcher, computational decision-making |
| Ted Robertson | Co‑founder, Knit Health (entrepreneur) |
| Dr. Maya Petersen | Professor, UC Berkeley School of Public Health; co‑founder |
Potential benefits and unanswered questions
The team frames the LCBM as a tool to inject contextual clinical knowledge into decision support, which could improve guidance around testing, referrals and treatments by mirroring how clinicians actually act. Advocates say this may reduce mismatches between AI recommendations and clinical workflow, a common barrier to adoption.
However, training on clinician behavior raises important questions that the team will need to address publicly as the model is evaluated and deployed, including:
- Whether patterns learned from historical behavior encode biases or suboptimal practices.
- How the model’s recommendations will be validated for safety and effectiveness across diverse care settings.
- Data governance and privacy measures for using detailed EHR audit logs.
Those issues are central to broader debates about clinical AI: models can reflect and amplify existing practice patterns as readily as they can generalize best practices. The researchers’ choice to use behavior as a training signal is novel, but its downstream impact will depend on rigorous evaluation and transparent governance.
The startup’s public materials stress that the LCBM is meant to complement clinician judgment by surfacing decision patterns from large volumes of operational data. As healthcare systems and regulators weigh the benefits and risks of AI-driven guidance, the Knit Health approach adds a new direction to how such systems might be trained and integrated into clinical care.
Further details about model performance, external validation and deployment plans were not included in the university announcement. Observers say those forthcoming data will be critical to assessing whether clinician-trained models can deliver safer, more usable decision support at scale.
