Joe Corkery, MD

All the talk at JPM Healthcare this year centered on announcements like OpenAI ChatGPT Health, Anthropic Health, and others that signaled the arrival of the empowered patient. My go-to question for providers was simple: “Are you concerned that patients can now feed their entire longitudinal medical record into a frontier model and have answers before they see you?” The answer was remarkably consistent: “Yes, but not because patients are being proactive. It’s because today our infrastructure doesn’t give us access to the same context.” That imbalance will quietly erode trust between patients and physicians. Not because physicians aren’t important or valuable but because the environments they work in don’t provide the context required to fully leverage their superpowers. Healthcare systems must radically retool their infrastructure to stay relevant in the era of the empowered patient. Consumer-grade health alone is not enough. Patients need,and deserve, clinical-grade care. As patient-generated data from wearables becomes pervasive, that context must flow seamlessly between the consumer and clinical worlds. At Tempus AI, our mission has always been to aggregate multimodal patient data to increase contextual awareness at the most critical decision moments. That data is increasingly created outside the health system. ŌURA Apple WHOOP Fitbit (now part of Google)

190

7 Comments

Thinking about the 3 major US health system EMR vendors who have 80%+ share of the Canadian market and their announced AI strategies. Trying to imagine what their Canadian rollouts might look like and to ask some key questions for Cdn health systems. Epic MEDITECH and Oracle (previously Cerner Corporation) have quite different strategies that are in keeping with their corporate histories. Epic is developing its Comet model based on a walled garden approach and historic client data. Meditech is partnering with Google, Suki, and others to bring capabilities into its EMR. Oracle is re-platforming entirely and is promising no more bolt-ons. While the 3 strategies are quite different, there are common questions across all platforms. I will ask some below. I am actively looking for resources and opinions; comments or by DM. 1. Most basic is price. How much will the new AI features cost? Early ballpark estimates are important to evaluate make/buy options. Can I not buy and stay on old product? Product upgrades of this type have been in the tens of millions in the past this is a CEO and Board discussion. 2.. When will the Canadian product be ready? Most vendors are focused on the US market for 2026. This is understandable. When will a full Canadian rollout begin? During the meaningful use product updates, Cdn product lagged by several years. Are we going to wait until 2030? 3. How will AI models be trained for Ex-US markets? US product features, including revenue cycle, prior authorization, formulary, semantics, and a number of other features, are important to American clients but not to Cdn & global clients. Worse, the business/clinical rationale implicit in these “features” and the Ai can be actively negative for patient care. How will US biases be managed? 4. Will application programming interfaces (APIs) be open to Canadian competitors for scribing, best practice clinical decision support(CDS), past practice CDS (I.e. clinical data lakes), and other features. Or will these be closed to the U.S. source systems and partnerships? 5. Are Canadian scribes, CDS vendors and other innovators going to be allowed to compete on a level playing field in the health system market of their own country? 6. Will the AI features be trained on Canadian data at all? And if trained on Canadian data, will that Canadian data have to be moved to the U.S. for training purposes, or will the training happen in Canada? Fundamentally, will the AI even see and use a hospital or health system's data? How? 7. Do you have other questions? Opinions and feedback very welcome… Epic Comet model details: https://lnkd.in/gdJzzr6Y Meditech and Google https://lnkd.in/gy-B6J84 Cerner replatforms https://lnkd.in/gmg7bSNW

31

11 Comments

Thoroughly enjoyed this article from my former colleague, Fawad Butt. The content should be required reading for organizations looking to go all in on AI but still struggling to articulate a specific strategy. At a time when LLMs appear to be experiencing some limited returns relative to the giant leaps and bounds of previous releases, small language models have the potential to unlock real value for meaningful use cases. Kudos to Fawad for not even directly mentioning agentic AI until his penultimate sentence, since the entirety of the article makes such a strong case for agentic architecture powered by SLMs. ChatGPT et al will continue to excite consumers when they consider the potential, but SLMs will prove more valuable in the near term for customers. 

11

The "Wild West" phase of Healthcare AI is officially over. We’ve moved past the hype and into the era of CPT codes, FDA change control plans, and multi-million dollar liability questions. I’m sitting down with a powerhouse panel to answer the one question every health system and investor is asking: When the AI makes the diagnosis, who is responsible when it breaks? We’re diving deep into: ⚖️ Liability: If we move from "Augmentation" to "Autonomous" does the malpractice shield shift from the doctor to the dev team? 💰 Reimbursement: Breaking down the 2025 CPT code updates. Can "Autonomous AI" finally scale without a human in the loop? Regulatory: How "Predetermined Change Control Plans" (PCCPs) are finally allowing AI to evolve in real-time without the FDA hitting the "pause" button. The Panel: Aakash Saraiya - the clinician and investor lens Hugh Barrigan – the technical builder lens Christopher Page – the clinician and investor lens If an AI is 10% more accurate than a human, but we can't explain how it reached the diagnosis—should we be allowed to use it? 💬 #DigitalHealth #HealthTech #AIinHealthcare

9

What the model gets to see may matter more than how the model was trained. 1️⃣ Context engineering means deliberately designing what information an LLM receives, not just asking it better questions. 2️⃣ Real clinical notes are noisy, repetitive, and temporally inconsistent, raw electronic health records fed directly to LLMs degrade performance. 3️⃣ As context length increases without filtering, LLM performance collapses; more relevant, dense information improves it. 4️⃣ Even a single irrelevant sentence added to a prompt can more than double the chance of a wrong answer. 5️⃣ Three core components are proposed: context retrieval (pulling the right data), context processing (distilling and structuring it), and context management (maintaining relevant memory across a clinical interaction). 6️⃣ These components can feed into agent systems using retrieval-augmented generation, knowledge graphs, or the Model Context Protocol to reduce reliance on probabilistic generation alone. 7️⃣ Over-filtering carries its own risks: relevant information can be discarded, aggressive summarization can introduce errors, and poorly built retrieval pipelines can surface outdated data. 8️⃣ LLMs' confident tone may increase over-reliance in junior clinicians and add a verification burden for experienced ones. 9️⃣ Who is accountable when AI-curated context causes a clinical error remains an open governance question. 🔟 The authors argue physicians should act as mediators of context, governing what AI models receive and what they infer. ✍🏻 David JH W., Jonathan H. Chen, Ethan Goh, MD. Context is key: context engineering as the next frontier of medical AI. BMJ Digital Health and AI. 2026. DOI: 10.1136/bmjdh-2026-000032 | Open Access

50

15 Comments

AI World Models coming to Healthcare, again. Current focus is on LLMs, but World Model neural nets are debuting next year for healthcare at scale. https://lnkd.in/gNJYrE69

3

With OpenAI is paving the path to a first-level triage / PCP layer + empowered patients, and Anthropic now building context-rich interoperability tools that link fragmented workflows and remove unnecessary admin….we are entering a brand new era in healthcare. Innovation can now focus on providing better care and experiences vs. solving infra problems with point solutions.

29

1 Comment

A CMIO told me a common pitfall with new EHR features: “They demo it on the simplest patient - one comorbidity, one medication. When we use it on real patients with real complexity, it falls apart.” I heard a similar story about a new AI note summarization feature released from an EHR. The health system was told how notes from dozens of patient encounters could be summarized for a patient. When the health system tried it for complex patient care such as oncology? Couldn’t summarize more than four notes. Stories like this are humbling because they remind us that simply having Tech features on a checklist doesn’t mean we can solve the complexity of medicine. I was thinking about all of this when a health system recently asked me to explain why they should use SeamlessMD when they have this alternative product that has a similar feature set on paper. This other product also had “reminders”, “tasks”, “questionnaires”, and “education” to guide patients pre/post-surgery, oncology, chronic care, etc. At the same time, the health system acknowledged the other product has had very little success in improving clinical outcomes like SeamlessMD. So they were wondering… what the heck was so different about our product? Then it struck me. It was our clinical content… but not in the way you might expect. One of the biggest benefits about developing our own clinical content pathways is that it forced us to focus to deeply understand the complexity of patient journeys and clinician workflows, and therefore uncover the complex features and workflows truly required to digitize complex patient journeys. Yes “reminders”, “tasks”, “education”, and “symptom monitoring” features are needed for any of these journeys… but that’s table stakes. The nuance to how those features must be designed for different journeys is critical for success, and you don’t develop those insights unless you have to develop the clinical pathways yourself. If a patient undergoing heart surgery is on specific medications, smokes, and has diabetes, their digital journey should be very different from a young patient having ankle surgery - which requires more sophisticated workflows in the underlying solution. Yet that complexity never shows up in a feature checklist and clinicians only realize this level of sophistication matters when actually implementing the product. It would be nice if every patient was young and relatively healthy with one comorbidity and few active issues - Digital Health would be really easy! But it’s NOT - which means Digital Health that actually works are the products that provide both the clinical content/workflows AND the Tech because that’s the one way to know the R&D was built right.

24

2 Comments

The healthcare technology landscape is undergoing significant shifts. Epic, a dominant player with 42% of the US market, faces legal challenges including an antitrust lawsuit from the Texas Attorney General over market monopolization and data access restrictions. Concurrently, Epic is embroiled in litigation with Health Gorilla and Particle Health regarding patient record access. Amidst this turmoil, Epic has launched a new AI charting tool, an ambient scribe designed to auto-generate patient notes from appointments. This development poses a direct challenge to standalone AI scribe startups, forcing them to reassess their value proposition against Epic's integrated solution. This confluence of legal battles and technological innovation highlights the dynamic and competitive nature of the EHR and AI healthcare sectors. #HealthTech #EHR #AIinHealthcare #Antitrust #HealthcareInnovation

9

1 Comment

𝗛𝗟𝗧𝗛𝟮𝟱: 𝗔𝗜 𝗜𝗻𝗻𝗼𝘃𝗮𝘁𝗶𝗼𝗻𝘀 𝗶𝗻 𝗠𝗲𝗻𝘁𝗮𝗹 𝗮𝗻𝗱 𝗣𝗵𝘆𝘀𝗶𝗰𝗮𝗹 𝗛𝗲𝗮𝗹𝘁𝗵 At HLTH 2025, Spring Health, Oura, and Hinge Health debuted major AI advancements:  • 𝗦𝗽𝗿𝗶𝗻𝗴 𝗛𝗲𝗮𝗹𝘁𝗵 launched VERA-MH, an open-source benchmark to assess the ethics and safety of mental health chatbots.   • 𝗢𝘂𝗿𝗮 announced a large-scale study to detect early signs of hypertension through its wearable ring, expanding its cardiovascular health focus.   • 𝗛𝗶𝗻𝗴𝗲 𝗛𝗲𝗮𝗹𝘁𝗵 introduced AI-powered movement analysis and a 24/7 assistant, Robin, to support musculoskeletal care. Together, these innovations reflect the industry’s shift toward safer, more personalized, and preventive AI-driven healthcare.

1