James Maitland brings a wealth of experience to the intersection of robotics and healthcare technology. As an expert in health informatics, he has spent years navigating the complex landscapes of IoT and clinical data science. Today, we sit down to discuss a transformative shift in how we handle the massive amounts of unstructured data—from pathology reports to messy physician notes—that have historically slowed the pace of medical discovery to a crawl. Our conversation explores how a healthcare-native AI approach is finally cracking the code on data abstraction, turning weeks of manual labor into hours of automated, high-precision insight. We delve into the failures of traditional “zero-shot” models, the specific breakthroughs achieved in acute myeloid leukemia research, and the potential for these technologies to revitalize drug discovery for patients who have run out of options.
For years, the brightest minds in clinical research have been bogged down by the sheer volume of fragmented data locked in PDFs and handwritten notes. How is this “data wrangling” affecting the speed of innovation, and what does it feel like for a researcher to finally break free from that manual labor?
It is incredibly frustrating for a top-tier scientist to realize that the majority of their day is spent acting as a “data wrangler” rather than an innovator. We are seeing healthcare organizations that are rich in data but poor in actionable insights because that information is buried in scanned PDFs and genomic reports that a computer cannot naturally read. In recent pilot projects, we’ve observed that a staggering 1,200 hours of manual abstraction can be reduced to just 40 hours when using a clinical knowledge-augmented AI. This shift is deeply emotional for researchers; it’s the difference between staring at a messy spreadsheet of blast percentages and actually uncovering a therapeutic opportunity that was hidden in plain sight. By automating this “essential first step,” we are finally allowing these experts to focus on the patient journey and the discovery of new biomarkers that could save lives.
Many early attempts to use AI for clinical data abstraction fell short of expectations, often failing to grasp the nuance of a doctor’s notes. Why have traditional models struggled to provide the accuracy needed for high-stakes medical research?
Traditional AI models often treat clinical notes as flat, unstructured text, which leads to a dangerous “reasoning drift” where the software cannot distinguish between a doctor’s hypothesis and a biological fact. When we look at “zero-shot” models—those not specifically trained for clinical environments—they often hit a ceiling with an accuracy of only about 72%. For instance, in complex cases like acute myeloid leukemia, these models might fail to realize that a bone marrow aspirate holds precedence over a peripheral blood smear when blast percentages differ. Without a structural “map” of medical relationships, standard AI is prone to guessing based on anecdotal patterns rather than universal clinical truths. It takes a healthcare-native approach to ensure that the technology isn’t just processing text, but actually understanding the clinical context and relationships across multiple records.
The recent collaboration between Verily Health and UCHealth tackled what many consider the “hardest of the hard” problems in clinical data. Could you walk us through the methodology that allowed them to jump from 72% to over 95% accuracy in data extraction?
The success of the pilot hinged on what we call a “clinical foundation layer,” which forces the AI to ground its extractions in established medical ontologies. Instead of letting the AI guess based on word frequency, this healthcare-native architecture requires the system to independently verify if a relationship is medically plausible before it is confirmed by a clinician. By incorporating this clinical knowledge into the extraction pipeline, the team saw accuracy for critical metrics, like myeloblast percentages, soar to more than 95%. This isn’t just a slight improvement; it represents a fundamental bridge in the accuracy gap that has historically prevented AI from being used in high-stakes precision medicine. Steve Hess at UCHealth purposely chose acute myeloid leukemia for this test because its heterogeneous patient populations and complex genomic data provided the ultimate stress test for the technology’s reasoning capabilities.
Beyond just saving time, how does this high-precision data abstraction change the game for precision medicine and the development of new therapies?
When we solve the abstraction bottleneck, we aren’t just moving faster; we are redefining what is possible in the search for targeted treatments. For example, by using a manually curated dataset, researchers at RefinedScience were able to identify a complex precision biomarker for a subgroup of patients who might respond to cusatuzumab, an anti-CD70 antibody drug that had previously been deprioritized after a Phase II study. This kind of discovery usually takes years of painstaking manual work, but with scalable AI abstractions, we can now hunt for these hidden therapeutic opportunities across much larger datasets. It allows us to unlock the latent value within clinical datasets that would otherwise remain invisible or “latent” to human researchers who are overwhelmed by paper files. We are moving toward a reality where clinical data resolution is the engine that drives the next generation of life-saving interventions by finding the right drug for the right patient subgroup.
What is your forecast for the future of clinical data science over the next few years?
I believe we are entering an era where the boundary between “raw data” and “medical insight” will virtually disappear, as AI becomes a trusted, expert-led partner in every lab. In the coming years, the standard for clinical research will shift from manual “data wrangling” to automated, knowledge-augmented reasoning that identifies patient subgroups with surgical precision. We will see a massive influx of previously “failed” drugs being rediscovered and repurposed because we finally have the tools to see exactly which patients will benefit from them. As accuracy continues to exceed that 95% threshold across diverse disease states, the bottleneck of the patient journey will dissolve, leading to a surge in precision medicine breakthroughs that were once thought impossible. The goal is to reach a point where no valuable piece of data is left behind simply because it was trapped in a PDF.
