SPEAKERS:
- Briana Belford, Practice Leader, Integrated Intelligence, Real Chemistry (Moderator)
- David Ginkel, Head of Publications (Biologic Division), AstraZeneca
- Janine Gaiha-Rohrbach, VP, Head, Global Medical Immunology, Biogen
- Jeremy Pincus, VP, Medical Affairs Digital and Tech, GSK
- Rachel Adams, Head USA Medical Communications and Scientific Training, Takeda
KEY TAKEAWAYS:
- Past-data segmentation creates scientific echo chambers that project yesterday's questions onto tomorrow's clinical needs
- LLM platforms have inverted the medical information hierarchy; digital is now the first touchpoint, not the last
- Guideline-directed therapy - the only outcome metric that matters - takes a year to observe, leaving course correction nearly impossible
- Generative Engine Optimization is an evidence strategy, not a communications tactic, and must begin at study design
- The highest-value personalization target is decision stage - what clinical question a physician is currently navigating - not content preference
Seven seconds. That is how long a physician now spends with a publication, manuscript, abstract, or poster. "You'd be lucky to maybe read the headline and one line out of the results," said David Ginkel, Head of Publications at AstraZeneca's biologic division, at Pharma USA 2026. "That's how brief it's become."
The compression of attention is precisely why personalization has become medical affairs' dominant strategic conversation. But the industry may be solving the delivery problem while the deeper problem goes unaddressed. "At a therapy area, is personalization at an individual?" asked Jeremy Pincus, VP of Medical Affairs Digital and Tech at GSK. "It's personalization at a stage of a customer journey. I'm not sure we're there yet as an industry." Briana Belford, Practice Leader at Real Chemistry, framed the session's organizing challenge: "How do you know when something requires MSL outreach versus a medical information response versus a digital medical experience?" Getting that triage right assumes you understand what the physician actually needs - and that assumption deserves scrutiny.
Janine Gaiha-Rohrbach, VP and Global Medical Head of Immunology at Biogen, reframed the entire premise. "The personalized scientific exchange for me is not predominantly about targeting the provider," she said. "I think it is about understanding what is the clinical decision that they're trying to make, and what is the evidence that we need to support so that we are equipping them for making the best decision." That distinction, delivery target versus decision support, is the fault line running through every conversation medical affairs is having about personalization right now. The industry has built sophisticated machinery for the former. It is only beginning to ask whether the latter is even measurable.
Past data builds tomorrow's blind spots
Belford opened the segmentation discussion with a question that contained its own answer: "If we're using past data to inform future engagements, how do we avoid reinforcing those same issues or essentially serving our audiences the content or experiences that they've already liked?" Most current segmentation models don't avoid it at all.
Gaiha-Rohrbach named the structural risk directly: "There's a risk of, if we're taking just the information from the past, that we're kind of creating almost like a scientific echo chamber." If algorithms ingest previous downloads and click patterns, they project yesterday's clinical questions onto tomorrow's information needs. The model learns what a physician has already engaged with, not what they're about to need. This is the personalization equivalent of the streetlight effect: looking where the data is, not where the insight gap actually lives.
Breaking the echo chamber requires building segmentation models on signals that look forward, not backward. Gaiha-Rohrbach described three distinct layers: a scientific signal drawn from emerging publications and congress conversations; a clinical context signal that tracks how physicians understand patient complexity, treatment paradigm shifts, and healthcare system variation; and a behavioral signal capturing real-time content engagement. The first two signals are forward-facing; only the third is retrospective. Most current models over-index on the third layer, which is also the easiest to collect and the least predictive of emerging need.
Rachel Adams, Head of US Medical Communications and Scientific Training at Takeda, added a dimension that data science models typically miss entirely. Physicians identify with professional communities, "tribes," in her framing, as much as with clinical subspecialties. An advanced practice provider specializing in sleep medicine still attends APP conferences, not just sleep medicine meetings, and their information needs and influence networks reflect both identities. Segmentation that only tracks disease-state interest misses the channels through which clinical behavior actually changes. "We think we know the need," Adams said, "but what we don't know is the need behind the need." The surface question often obscures the underlying driver: a competence gap, a confidence deficit, a specific patient complexity they haven't encountered before.
Pincus added an operational guardrail that prevents a common shortcut. Medical affairs' customer journey: awareness, assessment, adoption, advocacy, must be tracked independently from commercial. A physician expert in biologics may be prescribing a competitor's brand, making a commercial "adopt stage" data point actively misleading for medical segmentation. "The data point on that customer journey will be very different between commercial and medical," Pincus said. The data architecture that works for market access doesn't transfer to scientific exchange, and conflating the two produces confident-looking models with fundamentally wrong inputs.
The channel nobody planned to lead with
Open Evidence, a retrieval-augmented LLM platform built on peer-reviewed medical literature, logged one million unique physician queries in a single day. "The way that HCPs are getting their information is changing at just incredibly rapid pace," Adams observed. "The growth that one platform alone has experienced is truly exponential." The traditional triage model assigned digital the lowest-value role: static websites, general background information, the destination for physicians who wanted to explore a topic further. That model is now structurally obsolete.
Gaiha-Rohrbach captured the new information hierarchy cleanly: "Digital can scale information, MSLs scale understanding." The distinction is useful but also dangerous if it becomes a justification for ceding the information layer entirely to platforms pharma doesn't control. Gemini has ingested PubMed and ClinicalTrials.gov. Physicians are querying these systems for drug-specific evidence, dosing guidance, and competitive comparisons - and the answers they receive are shaped by what is machine-readable, consistently formatted, and frequently cited, not by what a medical affairs team believes is most accurate.
This is the strategic imperative Adams identified as Generative Engine Optimization. "It is imperative for all of us to really think about how we move forward now to address this GEO strategy," she said. The mechanics are specific: publish frequently, format for machine readability rather than PDF download, maintain decimal-level consistency across commercial and medical outputs, structure content with atomic statements tied to single references. LLMs reward consistency across sources and penalize fragmentation. An organization whose most important evidence lives in a gated PDF is effectively invisible to the platforms physicians now consult first.
Pincus drew a near-term boundary: "I don't believe at least in the next couple years we'll see public-facing pharma websites with a chat capability... the chat will still come back with approved materials, not generative because it still needs to go through a copy approval team." The regulatory constraint is real. But the liability position medical affairs leaders should resist conflating with strategic comfort: "When the data's wrong in Open Evidence, Open Evidence is going to get the warning letter, not the pharma company. I think we need to keep it that way." Legal clarity doesn't resolve the strategic problem. If an LLM misrepresents a drug's evidence profile to a million physicians in a day, the regulatory accountability lands elsewhere - but the clinical impact lands on patients.
Ginkel connected the LLM challenge directly to evidence generation strategy. "Instead of doing 50 studies, let's do three that really matter and that are the questions they want to know," he said. "Because we didn't answer this question in the clinical study, the phase three did not address this population or this question." GEO demands evidence generation that anticipates the questions LLMs will be asked - which means the study design conversation and the digital strategy conversation need to happen in the same room.
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 metric that arrives too late
"It's not clicks, it's not engagement rates, it's not scripts," Belford said. "It's improving the quality of decisions that our customers, that clinicians are making." The statement defines medical affairs' mission accurately. It also describes a metric that almost no organization can currently operationalize.
Ginkel articulated AstraZeneca's North Star: "We talk about do we increase the rates or the number of patients who are on guideline-directed therapy." The metric is genuinely patient-centered, not brand prescriptions, but adherence to clinical guidelines regardless of product. A lupus patient remaining on high-dose oral corticosteroids when they meet criteria for a biologic represents a failure of scientific exchange, not a commercial loss. Framing the outcome this way reorients the entire measurement conversation.
Pincus named the temporal trap that makes this North Star difficult to navigate by. "We will measure a reduction in unmet medical need. It takes like a year. And all of a sudden everything you've done up to a year could have been wrong." This is not a reporting inconvenience. It means segmentation models, content decisions, and channel investments are being made today based on outcome data that won't validate or invalidate those decisions for twelve months. Leadership wanting weekly updates on year-long outcomes is a structural mismatch, not a management failure.
Ginkel noted that intermediate metrics compound the problem rather than solving it. Altmetric scores are contextless without comparison points - a score of 42 is meaningless until ranked against MSL request frequency, medical information query volume, and congress engagement patterns. "What does 42 mean? It means nothing," Ginkel said. "But if you put it in perspective - this was the third most downloaded, requested data poster - then that's got value." Identifying proxy metrics for non-KOL physicians, the practitioners who treat the majority of patients, remains an open problem.
The decision layer no one is personalizing
Three speakers arrived at the same place from different directions. Gaiha-Rohrbach put it plainly in her opening remarks: personalized scientific exchange is about understanding the clinical decision a physician is trying to make, not targeting the provider. Adams called it the need behind the need - the gap between the surface question and its actual driver, whether a competence gap, a confidence deficit, or a patient complexity the physician hasn't encountered before. Pincus located the right unit of personalization at journey stage, not the individual.
Current segmentation models don't reach any of this. They record what physicians have already accessed and project it forward, which is why Gaiha-Rohrbach's echo chamber warning lands as a design critique rather than a data science one. The behavioral layer is the easiest signal to collect and the least predictive of emerging need. Gaiha-Rohrbach named the risk of over-indexing on it: "if we're going out there and come in the surveillance mode... we're losing that credibility of really having that deep scientific dialogue, understanding what are the needs, what are the questions they're trying to address today." Clinical intent and behavioral tracking are pulling in opposite directions, and most current models are built on the wrong one.
Decision stage is the personalization target that would change that calculus. A physician navigating biologic eligibility criteria needs different evidence than one troubleshooting treatment failure in a patient already on therapy. Knowing which question they're currently working through creates a hypothesis worth testing. Knowing which publications they downloaded last month does not.
Medical affairs leaders evaluating their current programs should pressure-test three things. First: what percentage of your segmentation model is backward-looking behavioral data versus forward-looking clinical and scientific signals? If the behavioral layer dominates, your model is more likely to reinforce existing gaps than close emerging ones. Second: can an LLM find your most important evidence in machine-readable format within 48 hours of congress presentation? If not, the channel that is becoming physicians' first touchpoint is working without your input. Third: for your highest-priority targets, can your MSLs identify which clinical decision each physician is currently navigating - not which content they last accessed? If the answer is no, the personalization infrastructure and the scientific exchange mission are not yet connected.
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.
