Limin Yu, MD, MS

One key reason why healthcare technologies like AI, telemedicine, and software haven't structurally lowered costs or broadened access for the entire system after decades of experimentation (excluding pure financial engineering of payer contracts) is the limited reimagining of care delivery itself, with asynchronous telemedicine as a notable exception. In the same way that digitizing physical paper failed to exploit the capabilities of a new computing medium for word processing, hosting physician-patient conversations over video does not transform the actual delivery of care. In telemedicine, many of the inputs (e.g., clinician time) and therefore cost structures of in-person care persist, which is why we often see fee-for-service telemedicine companies struggle to achieve economies of scale (i.e., growing revenues non-linearly while growing costs linearly). Until we deconstruct the components of a medical appointment – from intake and triage, to chart review, history gathering, physical examination, diagnosis, treatment, and follow-up – we will continue to be underwhelmed by the disconnect between the transformative potential of these technologies and real-world patient/system impact. I’m optimistic that the emergence of reliable and performant medical AI agents will urge us to perform this deconstruction and redesign the fundamental building blocks of a “medical appointment.”

215

38 Comments

Oracle Health announced plans to launch a next-generation EHR platform in 2025, incorporating advanced cloud and AI capabilities. This new EHR will feature Oracle's clinical AI agent, voice-activated navigation, and Health Data Intelligence, which integrates patient data from over 300 sources to enhance clinical workflows and improve patient outcomes. The platform aims to replace traditional EHRs, which often dictate rigid workflows, with a more dynamic system that allows for personalized care plans. Oracle Health emphasizes that the new EHR is a complete reinvention rather than an upgrade of the previous Cerner system. Additionally, Oracle plans to become a Qualified Health Information Network to facilitate standardized data sharing across the healthcare ecosystem. Executives at Oracle also unveiled new cloud applications to improve patient management and streamline front-office operations, reinforcing the company's commitment to enhancing healthcare accessibility and data utility while addressing interoperability challenges in the industry.

1

As the discourse shifts from models to compound AI systems / agents, we need better AI benchmarks to evaluate multi-modal and multi-step task performance, especially in healthcare and life sciences. When we wrote the first paper demonstrating ChatGPT's performance on the USMLE, we chose the US Medical Licensing Exam as a benchmark for accessibility, speed, and ease. This benchmark was never intended to represent AI model performance on real-world clinical tasks. Today, I still see so many research teams and startups using benchmarks (like the USMLE) that are ill-fitted for assessing the true clinical or scientific performance and utility of the models they are developing for real-world contexts. Benchmark development may be seen as a "less sexy" area of research, but it is of paramount importance. Years after the rise of the transformer, we still lack adequate benchmarks for so many single-step tasks in biomedicine. With compound AI systems (i.e., architectures that integrate multiple AI models to perform complex tasks) emerging, we need new benchmarks for agentic behaviors. I'd argue that developing an agent with novel capabilities without at least proposing a companion benchmark (if an industry standard does not yet exist) may hinder the adoption of said agent, especially for high-stakes workflows. Designing more benchmarks that capture/simulate real-world clinical and scientific workflows will help us mitigate the major discrepancies observed between in silico and in vivo performance and better support safe + effective deployment of AI in biomedicine. There are already brilliant people focused here, and we need more. DMs are open if you're researching or working in this area of multi-step/multi-modal benchmarking in healthcare and life sciences! #healthcare #ai #artificialintelligence #generativeai

100

18 Comments

To follow up on Dede Ainbinder recent post: If we use the current EHR data to build a AI clinical decision support tool it is unlikely to capture our clinical decision making process. EHR's are inherently designed to maximize billing and minimize risk of litigation. A clinical decision support tool based on this data will likely do the same. There is a large sunken cost fallacy here where we continue to hang on to garbage data. Just because the data gathering and storage process was capital and time intensive doesnt mean the data is valuable. To truly capture the clinical decision process we need to capture data which sits outside billing and litigation. Its hard to imagine a future where this data isnt used for billing and cannot be called upon during litigation. #EHR #healthdata #clinicaldecision

17

8 Comments

Big healthcare AI news yesterday as Tempus AI received the first-ever CPT code for an algorithm-only analysis of previously-sequenced, laboratory developed test transcriptomic data. TLDR: AI molecular phenotyping has a new business model. The PurISTSM test uses RNA profiling to identify molecular subtypes in patients with advanced pancreatic ductal adenocarcinoma, which may inform first-line therapy selection and management. While I'm skeptical that payer balance sheets are going to solve the "who pays for healthcare AI" problem, the reimbursement pipes for clinical AI are maturing. I am surprised to see more progress for usage-based vs. performance-based models given the many challenges with FFS economics. It will be interesting to see how payers assess the "medical necessity" and therefore reimbursability of clinical AI as these models come to market. https://t.co/eLe18SmihV

420

9 Comments

Here’s your recap of last week’s health IT news 🗞️ 👇 🤖 AI Products • Emergency Services, Inc. + Augmedix Go ED: The emergency medicine group adopted the new ED-specific AI ambient scribe solution.    • Color + OpenAI: The cancer care tech company partnered with the GenAI giant to build a AI clinical decision support copilot.    • AKASA Medical Coding: The AI RCM vendor known for its AI prior auth tool AKASA Authorization Advisor released a new GenAI medical coding solution.    • Humata Health: The AI prior auth for providers vendor closed a $25M funding round led by healthcare org-backed funds The Blue Venture Fund and LRVHealth.    • Wellsheet + Concord Hospital Health System: The AI-enabled Smart EHR UI solution announced a client-reported 16.3% reduction in length of stay and 40% decrease in time spent in the EHR via its partnership with Concord Health, as well as the release of its new LLM-generated handoff summary feature. 🍎 VBC • BSIM Healthcare Services + Innovaccer: The healthcare provider—embarking on its first VBC contract—adopted the pop health data platform for analytics. 🩼 Durable Medical Equipment • Parachute Health: The DME e-prescribing and ordering solution released an update that integrates prior authorization. ⚕️Clinical Staffing • Matchwell + Indiana Hospital Association: The clinical staffing service partnered with the hospital member org to create the first state-wide resource pool for temp staffing. 🧑‍💻 Virtual and At-Home Care • Deaconess Health System + KeyCare: The Indiana-based health system adopted the Epic-based virtual care provider for around-the-clock virtual urgent care.    • UC Davis Health + Current Health: The health system launched a new RPM program for high blood pressure in partnership with Best Buy Health’s care-at-home platform. 🧠 Food for Thought: We came across a couple of good resources this past week on the current state of healthcare data information exchange: • This article in Health Affairs from Deven McGraw and Tina Grande provides a comprehensive overview of some of the patient health data abuses we've seen come to light in recent months, as well as a number of policy recommendations to improve the health information exchange landscape for all participants. https://lnkd.in/e_GEQ-JJ  (h/t Alya Sulaiman)    • This episode of Health Gorilla's InteropTalk gives a helpful update on the current state of affairs, with a particular focus on the state of trust across the health data exchange community and how the definition of "Treatment" may evolve. https://lnkd.in/e96AJG-5 (h/t Brendan Keeler) --- Reminder: you can sign up for the Elion Briefing to get the latest healthcare tech news delivered to your inbox each week 📬 👉 https://lnkd.in/e2UJ3ckP

36

2 Comments

Learn about ethical and practical integration of AI, focusing on creating systems that improve patient care without compromising physician autonomy or patient privacy from speaker Jesse Ehrenfeld MD MPH at AIMed24. He advocates for responsible AI use that supports clinical decision-making, enhances patient safety, and alleviates administrative burdens on healthcare providers. #AIMed24

35

Here’s your recap of last week’s health IT news 🗞️ 👇 🤖 AI Clinician Assistant • Avaamo launched an AI ambient scribe solution with support for over 50 specialties with pre-trained models. • Wound AI, an AI clinical decision support solution built by VANTIQ, Telemedicine Solutions, LLC, and NTT DATA, launched and is now available 📞 AI Contact Center • Premier Inc. awarded a group purchasing agreement to TeleVox Healthcare. 🏥 AI Facility Management • Medely has integrated with Oracle, allowing users to leverage Oracle’s Cloud HCM capabilities. 🌡️ AI Remote Patient Monitoring and Triage • The Alliance for Smart Healthcare Excellence—established by care.ai, now a part of Stryker, and developed with assistance from CHIME—was established to promote the Smart Hospital Maturity Model (SHMM). 📈 Data & Analytics • Epic expanded its set of APIs available to developers to support USCDI v3. • Health Catalyst will acquire Intraprise Health, a cybersecurity provider; it was also selected for analytics and interoperability by CyncHealth, Iowa’s health information exchange (HIE). • Health Commons Project acquired OneHealthPort • Socially Determined released a new application that helps organizations access data for minors. 🧾 Orders & Fulfillment • Surescripts is expanding its Medication History for Populations product to health plans, enabling care management teams to identify gaps in care and improve outcomes. 🧑‍💻 Virtual & At-Home Care • Mass General Brigham launched a partnership with the New England Center and Home for Veterans to provide hospital at home services for veterans experiencing homelessness. • MedStar Health expanded its partnership with DispatchHealth to support care transitions.    🍎 VBC • i2i Population Health is partnering with CureMD to surface actionable data to providers at the point of care. 🧠 Food For Thought • Brendan Keeler shared an interesting post differentiating AI solutions across augmentative vs. autonomous and then further breaking down each. One take we found interesting was, “Active interfaces are fun, but ultimately have some of the same problems as Siri - they require explicit engagement and cognitive overhead to use. This means their adoption and impact often fall short of expectations - users must actively remember to use them and overcome the friction of engagement.” https://lnkd.in/eR39qFpK --- Want the latest healthcare tech news delivered to your inbox each week? Sign up here 📬 👉 https://lnkd.in/e2UJ3ckP

34

AccendoWave - A Pain #Data Company A new California law went into effect Jan. 1 that prohibits payers from making coverage decisions based solely on artificial intelligence algorithms. Under the new law, any denial, delay or modification of care based on medical necessity must be reviewed and decided by a licensed physician or qualified healthcare provider with expertise in the specific clinical issues at hand, according to a Dec. 9 news release from California Sen. Josh Becker, the bill's sponsor. 

5

GPT versus Resident Physicians. Future of doctors The advancement from GPT-3.5 to GPT-4 marks a critical milestone in which LLMs achieved physician-level performance. These findings underscore the potential maturity of LLM technology.

4

This paper evaluates a cost-efficient strategy for using LLMs in health systems, addressing the underexplored economic and computational challenges of their utilization at scale. 1️⃣ The paper focuses on query concatenation—a method of grouping multiple clinical tasks and notes into a single input—to optimize performance, scalability, and cost-effectiveness without compromising accuracy. 2️⃣ Over 300,000 experiments were conducted with real-world clinical data, showing that performance deteriorates as the number of simultaneous tasks and text inputs increases. 3️⃣ High-capacity models like GPT-4-turbo-128k and Llama-3–70B showed strong resilience, maintaining accuracy and formatting even under heavy task loads. 4️⃣ Optimal task burden was identified as 50 simultaneous tasks, beyond which accuracy and formatting declined for most models. 5️⃣ The proposed concatenation method for combining multiple queries led to up to 17-fold cost savings at scale compared to traditional single-query methods. 6️⃣ External validation with public datasets confirmed the trends, supporting the utility of high-capacity LLMs in medical settings for cost-effective, large-scale tasks. ✍🏻 Eyal Klang, Donald Apakama, Ethan Abbott, Akhil Vaid, Joshua Lampert, MD, Ankit Sakhuja, Robert Freeman, Alexander Charney, David Reich, Monica Kraft, Girish Nadkarni, Ben Glicksberg. A strategy for cost-effective large language model use at health system-scale. npj Digital Medicine. 2024. DOI: 10.1038/s41746-024-01315-1

192

13 Comments