Robbie Freeman

《As the A.I. pioneer Yoshua Bengio has recently emphasized, advancing generalized A.I. systems that can exhibit greater autonomy isn’t necessarily aligned with human interests. Humanity would be better served by labs devoting more resources to building specialized tools for science, medicine, technology and education.》

5

Google Cloud is expanding the use of AI agents in healthcare through new partnerships, building on a recent survey indicating 44% of healthcare executives use these tools. These collaborations include Hackensack Meridian Health for clinical note summarization and NICU assistance, IKS Health for automating prior authorizations, and #ColorHealth for developing an AI assistant for #breastcancerscreenings. For more details, https://lnkd.in/gwz9UfbN

5

1 Comment

Beyond the Hype: How AI and GenAI Are Transforming CDI AI is everywhere in healthcare, including within clinical documentation integrity programs. Every vendor seems to be offering some type of AI solution, promoting their use in improving physician documentation and financial performance. After seeing several vendor's AI documentation programs, I am not impressed and NOT convinced that these programs will drive better physician documentation while alleviating medical necessity and level of care downgrades, not to mention DRG downgrades and clinical validation denials. Reality- Only Physicians can communicate patient care after seeing the patient, taking the time to improve the quality and completeness of their documentation through guidance of physician advisors. AI may be used as a tool to identify opportunities for better documentation. #betterdocumentation, #AI, #realityofdocumentation, #onlyphysicianscandocument https://lnkd.in/ed_H5k3j

9

2 Comments

This article - not so much about which EHR vendor has the “better” AI strategy - points to the directions and differentiators being debated in the industry: Built-in vs Bolted-on AI: ​Built-in AI is integrated from the start, offering a more integrated experience with better data security. Bolted-on AI is added later, potentially causing fragmented workflows and data privacy risks. Semantic Database: A database that stores data but also understands the meaning/ context behind that data (e.g., knows that "MI" and "Myocardial Infarction" are the same thing). This helps AI understand different types of data (clinical, coverage rules, research...) Live, Evergreen Data: AI that works off live data, adjusts without needing re-training and uses the most current information in the system (e.g., new best practice guidelines, drug approvals, insurance rules). The contrast is an AI trained on static historical "snapshot" of data that becomes outdated over time. Knowledge Graph: A massive, intelligent map placed on top of the database to capture relationships between data points. It doesn't just know what "diabetes" is, but can make connections to HbA1c, insulin, and diet recommendations. This web of relationships provides the context for AI to make better inferences. AI Agents: Various agents that perform tasks like filling out prior authorizations, sending orders. e.g., instead of one general AI, Oracle uses many smaller agents: one for billing, one for lab orders, one for finding clinical trials...they all work together using the shared Knowledge Graph. Explainable AI (XAI): Clinicians need to know where AI decisions are coming from and need an AI that shows its work/ cites sources. If it recommends a treatment, it should be able to point the clinician to the clinical guideline, patient data point, or research paper it used to arrive at that decision. Partnerships with frontier AI (Reasoning Capabilities): Pairing AI agents with frontier AI’s reasoning models to understand instructions and draw inferences. Oracle says it's using LLMs for their NLP abilities to understand complex language and logic. Voice-first ambient technology: The system uses voice recognition to listen to patient-provider conversations in the background and suggests billing codes, orders, notes based on what it hears. Interoperability: Epic's strategy is centered around its Care Everywhere network that connects Epic customers with each other and with other EHRs through a point-to-point data exchange. They also leverage FHIR APIs to allow third-party applications to integrate with their system. Oracle Health is focusing on an open cloud-based platform, has applied to be a Qualified Health Information Network (QHIN) to participate in TEFCA. https://lnkd.in/gca-3bnZ

6

1 Comment

Interesting move. What becomes clear is this—connecting data is only half the problem. The real question is: Will this data actually flow into clinical workflows where decisions are made? Healthcare doesn’t lack data. It struggles with fragmentation, lack of interoperability, and context. If platforms like this can enable: • patient-owned, portable data • seamless integration with hospital systems • and real-time clinical usability —only then this becomes transformative. If not, we may just be building better datasets to train AI models, not better healthcare systems. Let’s see which direction this evolves.

7

Has the narrative fully moved on from "health tech will solve the unnecessary care problem?" In his excellent Healthcare AI Predictions for 2026 article, Graham Walker, MD calls it 'The Patient Trust Singularity': "the moment when patients routinely trust their AI system’s medical judgment more than their clinician’s." What's worrisome is that models are confident, often sycophantic, and rarely uncertain. If you've spent any time with an LLM, you know how persuadable they are. ChatGPT, Claude, and Gemini would not make a great debate team. Any doc will tell you that, once a patient is convinced they need a particular treatment, it's very difficult to convince them otherwise. Even if that treatment is not supported by the literature, low value, or even potentially harmful. Tuning and guardrails work to a point. Even the most well trained model can't overcome the garbage in, garbage out problem. In Ortho, we have Clinical Practice Guidelines meant to drive clinically appropriate, evidence-based decision-making. The problem is that recommendations are almost always vague because the evidence they're based on is weak. In other words, an exhaustive review of the literature uncovers that most studies are low quality, inconsistent, conflicting, and biased. That's not an indictment of Orthopedic research, it's a function of the complexity and cost of producing high quality studies. (Although, truth be told, medical publishing *does* have a quantity over quality problem). These gray areas, where weak evidence meets confident models tuned for resolution, are problematic. Doctors may feel obligated to order tests and treatments they don't think are indicated or may watch their patients go elsewhere if they hold their ground. Consumer-driven options are ready to fill that void, counterpositioning themselves as an alternative to the technophobic, paternalistic traditional system. The WISeR model says arthroscopy for knee arthritis should be avoided. A consumer LLM could be steered to a different conclusion. Liability and accountability are well known issues rightfully garnering a lot of attention. The long term clinical and financial impacts of AI are less certain. Will implementation be a net negative or net positive? Will patients seek out "clinical-grade" AI or rely on free, frictionless tools? (Dr. Google is already powered by AI). If AI increases patient conviction in low-value interventions, are we building a system that scales the very problem healthcare innovation has been trying to solve? (Link to Graham's article in the comments.)

38

20 Comments

We talk a lot about using EHR data for clinical AI – but most of that focus is on patient data only. A new npj Health Systems perspective argues we need a true dual lens: patient data and EHR use metadata (the clickstreams of how clinicians actually work). 🔗 Article: https://lnkd.in/gP-3e5T5 EHR audit logs (clicks, notes, orders, messages, barcode scans, time in chart, after-hours work) aren’t just IT exhaust. They’re a rich, underused signal about workflow, teamwork, and cognitive load. For those of us working in clinician wellbeing and systems redesign, this opens up big opportunities: • Use EHR metadata as core wellbeing outcomes when we evaluate new AI tools (scribes, inbox triage, early warning systems), not just traditional utilization metrics. • Track whether AI is truly reducing burden – fewer clicks, less after-hours work, smoother task flow – or simply reshuffling work to different people or times of day. • Identify “hot spots” of friction and overload at the service, team, or role level and pair them with targeted workflow and staffing interventions. At Michigan Medicine we are using this data to help us understand the impact of ambient documentation on our providers’ workload. Partnering with Dana Habers, Greta Branford, Michael Burns, Carolyn Winslow, Julie Ann Evans, Jeffrey Patterson, MHSA, MPH, Kevin Pearlman, Phillip Amara, and many others. The takeaway: if we want AI that supports better care and better work, we should treat EHR audit logs as first-class data for both model design aind clinician wellbeing, not an afterthought. Thank you Thomas Kannampallil and Julia Adler-Milstein for leading the way! David C. Miller, MD, MPH Mary Masson

21

1 Comment