SPEAKER:
Jonathan Kish, Vice President & Head of Research Sciences, Flatiron Health
KEY TAKEAWAYS:
- Claims data reveals prescribing patterns; unstructured clinical notes reveal the deeper reasoning behind treatment decision-making.
- Local care infrastructure, in addition to physician preferences, may be an important determinant of treatment choice.
- The advent of LLM-powered extraction has recently made physician reasoning available at scale.
- Quality governance, not technical perfection, is the real adoption gate for EHR-native intelligence.
- Pre-launch TPP benchmarking and six-month post-launch barrier diagnosis are the highest-value use cases.
The 'why' behind every treatment decision has always existed — just not in claims
A physician documents that a patient's "advanced age and severe cardiac comorbidities" make systemic chemotherapy too risky. Another note explains that a patient who was intended for CAR-T therapy never received it because the cell manufacturing process failed. In a claims dataset, both are recorded as a treatment switch. From a commercial perspective, one patient requires better patient selection guidance and the other requires a logistics support program. While insurance claims can tell you what happened, only an understanding of the why can give commercial teams the context required to respond accurately and efficiently.
Jonathan Kish, VP and Head of Research Sciences at Flatiron Health, opened his Pharma USA 2026 session by naming a structural gap that oncology commercial teams have operated around for decades: the absence of physician reasoning from the commercial intelligence stack. "Can we understand the 'why' behind treatment decision making?" he proposed.
Physician reasoning at scale: a capability that didn't exist two years ago
The pace of oncology innovation compounds the need to unlock decision-making insights from the point of care. New genomic testing protocols, updated treatment guidelines, novel market entrants, and shifting reimbursement dynamics are all moving faster than the research cycles designed to track them.
What Kish presented was a fundamental limitation in how pharma organizations understand physician behavior, along with the claim that large language models have unlocked these insights at scale for the first time.
What claims data was never designed to tell you
The pharma industry's commercial intelligence architecture rests on two pillars that were not designed to talk to each other. Claims data provides prescribing behavior at scale: treatment sequences, line-of-therapy transitions, market share shifts. Primary market research captures physician preferences through advisory board input, survey responses, and qualitative interviews. The assumption embedded in most commercial strategy is that combining these two streams produces a sufficient picture of what drives treatment decisions.
"Claims datasets can only show a broad picture of what's happening in different patient populations," Kish said, adding that the clinical depth from an EHR can improve understanding of patient journeys and the key drivers of certain behaviors and treatment patterns with physicians. Claims record what a physician prescribed; they cannot record why a physician hesitated, what clinical factors tipped the decision, or whether a logistical barrier intervened between intent and action.
How the clinical record informs commercial strategy
The clinical note examples made this concrete. The cardiac comorbidity note is a clinical preference signal: the physician's reasoning is explicit, the barrier is identifiable, and a commercial team equipped with that signal knows exactly what to address and for which patient profile. The manufacturing failure note was a supply chain and logistics signal, and the resolution for that must be different as well.
What Kish didn't state explicitly, but his evidence supports, is that this gap has existed for as long as oncology commercial teams have operated, and it shapes how commercial teams segment the market, deploy field resources, and ultimately how patients access treatment. Kish's HCP segmentation example makes that concrete.
The point-of-care missing link in HCP segmentation and field strategy
Conventional HCP segmentation organizes field strategy around physician-level attributes: prescribing volume, specialty tier, NPI affiliation, formulary access, and consequently physician characteristics are treated as the primary driver of treatment selection.
Kish explained how clinical depth extracted from unstructured physician notes can unlock more relevant patient-centric segmentation variables. Using Georgia as an illustrative example, in Atlanta, community oncologists operate within a mature CAR-T referral network anchored by Emory's Winship Cancer Institute, featuring transplant evaluation pathways, patient travel support, and coordination protocols. Ninety miles south in the Warner Robins and Macon corridor, providers treat clinically identical patients with BiTE therapies, which offer no cell extraction, no manufacturing wait, and no travel requirement. "That 90-mile difference might make all the difference in the world for the patient to be able to get on therapy quickly," Kish noted.
The commercial implication is significant. Atlanta-area field strategy can focus on reinforcing existing referral patterns and ensuring community oncologists identify strong CAR-T candidates. The Warner Robins corridor requires a different deployment: patient assistance programs, travel reimbursement, manufacturing timeline education, and a case for why the logistical burden is worth it for the right patient. No claims dataset reveals this distinction with geographic precision. No advisory board question captures it at the territory level.
The implication Kish left unspoken is that, if local care infrastructure is an important segmentation variable, there is a meaningful opportunity to strengthen pharma field force alignment models by incorporating it more directly. This is not oncology-specific; any disease where treatment logistics mediate between physician intent and patient access faces the same problem, whether it's infusion centers for immunology, specialty pharmacies for rare disease, or transplant evaluation for cell therapies. The oncology EHR intelligence layer makes details like access barriers visible and, for commercial teams willing to act on it, that's a significant competitive edge.
Discover more on this topic at Pharma Customer Engagement USA 2026 (October 27-28, Philadelphia) - where commercial, marketing, medical, data and AI pioneers converge. Explore the agenda here.
The commercial teams that act first are the hardest to catch
Extracting structured insights from unstructured clinical notes at scale has been technically aspirational for years. Rule-based NLP systems could identify specific terms; early machine learning models could classify certain patterns. Neither could process the contextual, inferential reasoning embedded in a physician's narrative note, such as the weighing of competing clinical factors, the acknowledgment of patient preference, or the documentation of a logistics failure. "This type of work has only become available in the last year and a half with the advent of LLM technology. Now, with the LLM approach, we can run at scale to all of the corpus of unstructured data that exists within Flatiron," Kish said.
The competitive clock this sets is real. First-mover advantage in commercial intelligence is not permanent, but it is durable given how treatment paradigms can shift quickly.
Building trust into LLM-powered clinical intelligence
For organizations evaluating EHR-native intelligence, the technical question is largely settled, but quality governance deserves scrutiny now. Kish described Flatiron's VALID framework for variable-level accuracy testing, outcomes validation against expected population benchmarks, and cross-dataset verification as the quality control architecture applied to both structured data extraction and LLM-based physician reasoning analysis. He was direct about its limits: "Is it perfect? No one can say that any LLM is necessarily perfect in the extraction, but we don't let things go out the door that we don't believe have the quality to support the specific use case."
Kish's candor about LLM limitations is itself a signal: Flatiron has built a standardized, repeatable system for extracting clinical insights, not a one-off research project, and for commercial oncology teams evaluating scalability across tumor types and territories, that distinction is the one that matters.
Turning clinical intelligence into commercial advantage: where to start
Kish's presentation makes it abundantly clear that insights into physician reasoning can significantly change how commercial teams act. The question for commercial strategy leaders is not whether to build the oncology intelligence layer around this data — it's where in the drug lifecycle it delivers the highest return. Kish's answer is specific: pre-launch TPP benchmarking and six-month post-launch barrier diagnosis are the high-value entry points where the gap between claims-only intelligence and EHR-derived insights is widest, and where the cost of operating with incomplete information is hardest to recover from.
To get you highlights of Pharma USA 2026 faster, we are using generative AI technology to summarise the transcripts of the sessions. If you have any feedback about the summary, please contact lucy.fisher@thomsonreuters.com.
Discover more on this topic at Pharma Customer Engagement USA 2026 (October 27-28, Philadelphia) - where commercial, marketing, medical, data and AI pioneers converge. Explore the agenda here.
