Berkeley Spinout Raises $11.6M to Build AI That Learns From Real Doctors, Not Medical Textbooks

A new healthcare AI company emerging from UC Berkeley is taking a fundamentally different approach to clinical intelligence: instead of training algorithms on medical literature and research papers, it's learning directly from how doctors actually practice medicine.

Knit Health launched in May 2024 with $11.6 million in seed funding to develop what it calls the world's first Large Clinical Behavior Model (LCBM)—an AI system trained on anonymized electronic health records from 130 million patients across 30 major U.S. health systems, representing approximately one-third of the American population.

The distinction matters because clinical practice often diverges significantly from published guidelines. By analyzing audit logs that capture millisecond-by-millisecond records of every action taken by doctors, nurses, and care teams, Knit Health's model learns the patterns of actual clinical decision-making rather than idealized protocols.

From Academic Research to Commercial Deployment

The company was founded by UC Berkeley Haas healthcare economist Professor Jonathan Kolstad (CEO), doctoral student Jonas Knecht, behavioral economics expert Ted Robertson (COO), and Dr. Maya Petersen, an MD/PhD professor of epidemiology and biostatistics at UC Berkeley Public Health who serves as Chief Scientist.

Knit Health is the first startup to emerge from the Center for Healthcare Marketplace Innovation (CHMI), a joint initiative spanning UC Berkeley's Haas School of Business and College of Computing, Data Science, and Society. The seed round was co-led by Uncork Capital and Frist Cressey Ventures, with earlier pre-seed backing from Moxxie Ventures and Coalition Operators.

The founding team secured critical data partnerships with UC San Francisco and Providence Health, with the latter providing emergency room data for model validation. Development moved rapidly: initial concept discussions began in October 2023, the company launched seven months later, and pilot deployments commenced in summer 2024.

Early Performance Signals Clinical-Level Accuracy

In validation testing using Providence Health emergency department data, Knit Health's model predicted 60-80% earlier in patient visits whether ED patients would require hospital admission—performing at clinician-level accuracy while providing substantially earlier signals.

This early prediction capability could transform hospital operations by optimizing patient routing and resource allocation. Rather than waiting until late in an ED visit to determine admission needs, hospitals could proactively prepare beds, coordinate specialist consultations, and reduce bottlenecks that contribute to emergency department crowding.

The model's causal inference layer, contributed by Dr. Petersen, aims to identify not just what clinicians typically do, but what distinguishes the decisions of top-performing clinicians—potentially elevating care quality across entire health systems.

Defining a New Category in Healthcare AI

Knit Health enters a crowded healthcare AI market with distinctive positioning. While most competitors focus on back-office automation or literature-based clinical decision support, the LCBM approach creates a new category: behavioral clinical intelligence that learns from practice patterns rather than published research.

The scale of training data—130 million patients—creates significant barriers to entry for competitors attempting similar behavioral models. As healthcare systems increasingly seek AI tools that integrate seamlessly with existing clinical workflows, Knit Health's approach of learning from actual practice patterns may prove more readily adoptable than systems requiring clinicians to change established behaviors.

The company's challenge now lies in demonstrating that predicting clinical decisions translates to measurably better patient outcomes and operational efficiency across diverse healthcare settings.

Sources

1. UC Berkeley researchers, entrepreneurs create first healthcare AI model trained to understand clinical decisions — Haas News