Verily Health said on Sept. 21, 2026 that a sponsored pilot with UCHealth, the University of Colorado Anschutz and RefinedScience tested whether AI could turn complex, unstructured clinical data into structured, research-ready variables. The project targeted clinical data abstraction, a bottleneck in biomedical research because patient information is often scattered across physician notes, pathology reports, genomic reports, scanned PDFs and other clinical documents. Verily said accuracy, privacy and security had remained barriers to using AI for this work.

Healthcare organizations hold large amounts of clinical data, but Verily said much of it is fragmented and difficult to extract for advanced analytic modeling. Researchers and other experts often become “data wranglers,” spending time preparing messy records instead of generating scientific insights. Kathryn Twyman, director of AI and Data Science at Verily Health, said clinical data resolution is needed to identify new biomarkers and therapeutic opportunities, and that automating abstraction in a trusted, expert-led way could shorten the path from data to insight.

Verily said standard AI models have underdelivered because clinical interpretation requires context, medical terminology and relationships across multiple records. Twyman said standard models can treat notes as flat text and lack a structural “map” of medical relationships, creating reasoning drift (mixing patient anecdotes with clinical facts). The article said zero-shot AI (task attempted without special adaptation) struggled in the pilot; in one test case, accuracy was only around 72% because models did not understand that bone marrow aspirate takes precedence over peripheral blood smear when blast percentages differ.

The pilot used acute myeloid leukemia, or AML, because it includes heterogeneous patient populations, complex genomic information, multiple report types and nuanced clinical interpretation. Steve Hess, CIO of UCHealth, said the team gave Verily the “hardest of the hard” problem on purpose and called the pilot a success. Verily said its clinical knowledge-augmented approach acted as a “clinical foundation layer” by grounding AI extractions in established medical ontologies (formal maps of medical concepts), helping the system check whether a relationship was medically plausible before clinician confirmation.

Verily said the clinical knowledge approach raised accuracy from 72% for simple zero-shot extraction of myeloblast percentage, a key AML metric, to more than 95%. The pilot also projected that about 1,200 hours of manual clinical data abstraction could fall to 40 hours. Twyman said RefinedScience previously used a manually curated dataset to identify a complex precision biomarker for a subgroup of AML patients who may respond to cusatuzumab, an anti-CD70 antibody drug that had been deprioritized after a Phase II study. Verily said faster abstractions could let researchers spend less time preparing data and more time searching larger datasets for similar insights.

Verily Health said on Sept. 21, 2026 that a sponsored pilot with UCHealth, the University of Colorado Anschutz and RefinedScience tested whether AI could turn complex, unstructured clinical data into structured, research-ready variables. The project targeted clinical data abstraction, a bottleneck in biomedical research because patient information is often scattered across physician notes, pathology reports, genomic reports, scanned PDFs and other clinical documents. Verily said accuracy, privacy and security had remained barriers to using AI for this work.

Healthcare organizations hold large amounts of clinical data, but Verily said much of it is fragmented and difficult to extract for advanced analytic modeling. Researchers and other experts often become “data wranglers,” spending time preparing messy records instead of generating scientific insights. Kathryn Twyman, director of AI and Data Science at Verily Health, said clinical data resolution is needed to identify new biomarkers and therapeutic opportunities, and that automating abstraction in a trusted, expert-led way could shorten the path from data to insight.

Verily said standard AI models have underdelivered because clinical interpretation requires context, medical terminology and relationships across multiple records. Twyman said standard models can treat notes as flat text and lack a structural “map” of medical relationships, creating reasoning drift (mixing patient anecdotes with clinical facts). The article said zero-shot AI (task attempted without special adaptation) struggled in the pilot; in one test case, accuracy was only around 72% because models did not understand that bone marrow aspirate takes precedence over peripheral blood smear when blast percentages differ.

The pilot used acute myeloid leukemia, or AML, because it includes heterogeneous patient populations, complex genomic information, multiple report types and nuanced clinical interpretation. Steve Hess, CIO of UCHealth, said the team gave Verily the “hardest of the hard” problem on purpose and called the pilot a success. Verily said its clinical knowledge-augmented approach acted as a “clinical foundation layer” by grounding AI extractions in established medical ontologies (formal maps of medical concepts), helping the system check whether a relationship was medically plausible before clinician confirmation.

Verily said the clinical knowledge approach raised accuracy from 72% for simple zero-shot extraction of myeloblast percentage, a key AML metric, to more than 95%. The pilot also projected that about 1,200 hours of manual clinical data abstraction could fall to 40 hours. Twyman said RefinedScience previously used a manually curated dataset to identify a complex precision biomarker for a subgroup of AML patients who may respond to cusatuzumab, an anti-CD70 antibody drug that had been deprioritized after a Phase II study. Verily said faster abstractions could let researchers spend less time preparing data and more time searching larger datasets for similar insights.