The 'unseen' value of precision data

Two decades after precision medicine began, critical insights are still trapped in low-resolution data — a bottleneck that Verily, UCHealth, the University of Colorado Anschutz, and RefinedScience are now overcoming. Terabytes of valuable clinical data, the natural byproduct of patient care, are largely unfit for AI and machine learning because they were never collected as research-grade evidence. This struggle with low-resolution data spans nearly every disease and demands a fundamental re-evaluation of data governance and culture. To achieve high-fidelity evidence at scale, governance must now integrate deep clinical expertise — and new infrastructure must be built to support data transformation. With Verily Refinery, a powerful curation engine built on Pre, disparate and multimodal data is transformed into a clinically informed, FHIR native model. Following positive results in a test case study with the University of Colorado Anschutz and RefinedScience, the platform is now providing the "resolving power" to make high-resolution clinical data actionable and unlock its latent potential.

Acute myeloid leukemia (AML) is a deadly blood cancer with survival rates significantly lower than common cancers like breast cancer (around 91–92%) and even lower among older patients. CU's hypothesis is that by leveraging advanced molecular and clinical data, including next-generation sequencing techniques (like CITE-seq), they can discover new, clinically important subclasses of AML. Using the data, they identified the anti-CD70 antibody drug cusatuzumab, which had been deprioritized by Janssen after a conventional Phase 2 trial. A subsequent analysis by RefinedScience suggests a complex precision biomarker can now identify a specific subclass of AML patients who are likely to benefit from the drug, creating a major opportunity for drug rescue. Their joint venture, OncoVerity, currently has a new Phase II study underway to prove this. Ultimately, Verily's physician-scientists and data technologists are building the necessary AI "lenses" to accelerate this new understanding of disease and research, establishing a scalable design pattern for precision health.

The recent project with UC Health, CU Anschutz and RefinedScience confirms the promise of their approach but also reveals the technical gaps they need to cross. This work leveraged Verily Refinery, powered by Pre, to take clinical data, predominantly molecular diagnostic reports from the electronic health record, and turn it into structured labels (a matrix of patient rows and abstracted label columns). Historically, this abstraction has been a time-consuming, "artisanal" process.

Their team then uses this matrix (X) to discover new AML classes by trying to find a machine learning function f that reliably predicts a clinical outcome (Y, such as 5-year survival) using the formula Y = f(X) + E. Components of X that are highly predictive can become companion diagnostics.

The results showed that Verily could automate the creation of these label columns using modern, scalable methods. Verily deployed targeted AI models, including LLMs capable of processing TIFFs while preserving tabular formatting. The results were encouraging, indicating that automated systems could match and even exceed manual curation quality:

• AI models matched the accuracy of human expertise, achieving greater than 95% accuracy on highly complex variables such as Karyotype strings, Next-Generation Sequencing (NGS) mutations, and pathology blast percentages.¹
• The program is estimated to have led to a 97% reduction in researcher time, compressing a projected 1,200 hours of manual abstraction into just 40 hours.²

However, our work revealed a greater opportunity for Verily: we can greatly enhance the model's resolving power by utilizing a clinical knowledge-augmented approach that injects deep clinical expertise into the data processing pipeline, establishing a clinically informed, FHIR native data model that is AI and analytics-ready.

Evidence: When developing an extraction for myeloblast percentage (a key AML metric), a simple, "zero-shot" AI approach achieved 72% accuracy. When clinical knowledge was incorporated into the extraction, the accuracy rose to 95%. Expertise matters.

Here are three key opportunities to extend their accomplishments:

1. Sharpening clinical variables: Current data labels often combine multiple biological concepts into overly broad categories, which reduces data resolution. Labels must be abstracted to define a single, distinct clinical concept.
2. Separating clinical logic from raw data: RefinedScience often "overloads" clinical logic into their data labels. For example, a classification like "AML with Recurrent Genetic Abnormalities" should be treated as a separate derivation (or "enrichment") from the specific genetic abnormalities that define it.
3. Sharpening data resolution via ‘Knowledge graphs’: This is where the complex clinical logic should be stored. A knowledge graph can represent the sequencing and hierarchy of information.
• Precedence: It can handle rules of precedence, such as how the percentage of "blasts" (immature blood cells) from a bone marrow aspirate usually holds more weight than blasts from a peripheral blood smear, but this precedence can flip under certain conditions.
• Ontology: It can also map the hierarchy of disease, like how AML is broken down into "AML with Recurrent Genetic Abnormalities" and then into more specific classifications.
• LLM Performance: Evidence from companies like Palantir and Microsoft shows that using knowledge graphs to ground Large Language Models (LLMs) significantly improves performance for tasks like retrieval and answering questions compared to standard methods. A GraphRAG method increased token efficiency by 9X to 43X.

In clinical medicine, some tasks must be deterministic (e.g., treating severe bleeding before a benign mole), and there is little room for AI "hallucinations." A knowledge graph, which represents facts ("entities") connected by their relationships ("semantics"), provides the structure needed to enforce this clinical logic.

The results of our recent project with UC Health, CU Anschutz and RefinedScience (see full case study) — demonstrated by high-value clinical use cases (e.g., the potential to rescue cusatuzumab) combined with an evolving tech stack of data refinement and knowledge graphs — establishes a scalable design pattern for precision health. Following the initial program's success, CU Anschutz has committed to a multi-year expansion, scaling the solution to population-level datasets across new therapeutic areas, including cardiovascular disease, neuroscience, and ophthalmology.

The promise of AI to transform health is historic, harkening back to past technological leaps. We’re now positioned to use data, knowledge, and technology to provide new, high-resolution "lenses" on human disease. Our modern applications help partners resolve new understandings of disease and identify value where it was previously unrecognized. The full benefits of AI for precision health will emerge through an intentional build of our tools that brings together technical capabilities, clinical expertise, with health system data. By focusing on patient safety and impact, this approach creates the structural knowledge network necessary to accelerate precision health for all.

¹ Data on file
² Based on Verily estimate - Data on file