Michael Schwartz

There appears to be a "Superstar Effect" in healthcare and biomedical AI research right now. A small percentage of publications are truly novel methods contributions that introduce new capabilities or behaviors; whereas a majority of contemporary research publications are focused on implementation science to assess the real-world utility and performance of new methods. This dynamic seems different from prior eras, when the translational infrastructure for implementing novel AI methods was largely unavailable. While methods research often captures more attention and accolades, both types of research are critical – novel methods without implementation science lacks translatability; implementation science without novel methods lacks novelty. If our goal is to invent and deploy transformative AI methods in real-world healthcare and life sciences contexts, we ought to continue encouraging and rewarding stellar implementation science work alongside breakthrough methods development.

90

3 Comments

"AI has got to go!" "Trust Nurses Not AI" from Nurses protest AI at Kaiser Permanente in Becker's Healthcare shows how much we have lost the plot. And its *all* the fault of the vendors in this Industry. IBM started with the mythical presentation of Watson for medicine. Then you have vendor after vendor making claims that studies do not show; videos of capabilities their products do not have; and muddling the water of AI is and can do so much that hospital technologists will say "thats not real AI" b/c your reality does not match their fantasy. We desperately need a Peterson Center on Healthcare for AI, but it would have to be 1000X the size to deal with the amount of BS in the market. In the mean time, what can help. 1. To Hospitals/Health Systems - Force vendors into trials and real world evaluation. Allow them to shape the environment to enable speed for your benefit and theirs - but make them show it work with real patients. No more 137 point RFP evaluations before you evaluate one solution. This is best practice b/c each AI requirement is worth 100 other ones (if it works) and you need to cut beyond the marketing. 2. To the Industry/Influencers - We need a SOC2 / HITRUST for AI. It does not have to validate models work, just that its real data, real training, force transparency, etc. This will not solve the problem but will eliminiate a significant portion. This is far better than government stepping in. 3. To Vendors - every video should be of your real product working (for computer vision AI), every demo should include an anonymized real data set (for predictive models) and every generative AI company should have a transparent set of test cases that can show the capabilities and boundaries of the tech. Also, would love to learn of educators who are helping teach healthcare professionals about AI evaluation in practical ways. As to the topic, our Director of Nursing at LookDeep Health (Tiffany Wyatt, RN BSN ) and many others have made the point that done well, AI can enable vs abstract the relationship between patient and nurse. One of the most valuable and scarcest resource in the hospital is clinical attention of nurses and doctor - AI can be the ultimate sorting hat that helps that expertise always be in the right place. Patients, nurses and even the financial drivers of the hospital can be aligned because of AI - but we need to show nursing and providers more generally thats what we are actually aiming for. #healthcare #AI #hype https://lnkd.in/ge8mW6WT

26

11 Comments

An important study nicely summarized by Spencer Dorn. The reality that genAI can have paradoxical effects that make a problem worse rather than better (providers taking more time on notes rather than less) emphasizes the need for studies like this. Kudos to UCSD for their work to evaluate this. A focus on summarization may be a better place for genAI to focus and kudos to technologists like Ruben Amarasingham, MD who have been working in that space. Patients have countless data points stored in tables (e.g., labs), structured content (e.g., medications, allergies, problem list), and free text (e.g., radiology reports and notes). The generative AI that can summarize this information and present it to the provider could make clinical visits more efficient and may help reduce cognitive load. Reductions in provider cognitive load could support a reduction in the use of heuristics to manage the overwhelming amount of data and the risk of cognitive biases which negatively impact care.

11

3 Comments

Very interesting, insightful study. Flare Capital Partners AI in healthcare is hype. Most AI stories for healthcare are tech advancements. Not real clinical advancements or fitting into their workflows, bearing few. MedTech Innovators face these paradox. 1.Adoption requires Evidence. But evidence comes only from adoption. Clinical data can not be synthetic data for good AI. 2.Regulatory, validation requires money. But investors require regulatory, validation. 3.MedTech need faster adoption, but healthcare is slower among every profession, a cut-throat industry. They won't use it, if it is more than 2-3 mouse clicks. https://lnkd.in/gUh8sDHe

1

Great presentation by Dan Buckland, MD for those looking to design Health AI/ML models. "...if we have been poorly treating a demographic or population, that will show up in the clinical data. We might not have race, gender or zip code as an identifier in the original clinical data." shared Dr Buckland. Good insights to support the need of Asher Informatics PBC AI/ML independent equity performance audits, AI-Assess. Through a funded federal research contract, we've designed a tool in which we can provide independent real-world data subgroup population testing on clinical and signal AI/ML models and products. Right now, the MVP is designed to support #oncologyAI FDA 510K submissions and enable the FDA access to the testing of an algorithms performance and therefore better evaluate an AI/ML submission. We've partnered with Gradient Health to provide data that includes race, gender, age and we're utilizing Rhino Health's federated platform to be able to pull testing data from multiple sources. If you're an oncology AI developer, we're currently conducting feedback interviews and would love to connect.

9

1 Comment

This isn't a post about #AI, but about data foundation and #interoperability in #healthcare. The demonstration of AI through #LLM is not sexiest part of this presentation. There's no silver bullet, we should know that at this point. Instead, I draw your attention to the rich discussion around the applicability of #OpenEHR and #HL7 #FHIR standards, and the importance of understanding why each of these #standards was developed and what for. Data definitions, data schemas, and how we apply standards to structure and represent information, both in storage and transportation, are key. Any further comments from me would be spoilers. I hope you enjoy the presentation and share your point of view, folks. Have a great week!

26

2 Comments

"Healthcare innovation is accelerating at an unprecedented pace, and cardiology is on the brink of transformative change using new forms of remote monitoring and artificial intelligence (AI). Thomas M. Maddox, MD, SM, professor of medicine and vice president of digital products and innovation at BJC HealthCare, Washington University School of Medicine, gave a presentation on new technology advances at ACC.24, the annual meeting of the American College of Cardiology (ACC). He then spoke with Cardiovascular Business, providing insights into cardiology's future. "Digital health is really starting to mature and become a key part of how we're going to practice cardiology going forward. There are some newer ways that we are collecting data about patient's health, and particularly when patients are not with us in a clinic or a hospital. So we are monitoring them at home and making sure that we're not missing early signs of disease worsening or disease progression," Maddox explained. " American Heart Association MedTech Innovator LSI Heart Rhythm Society #epeeps

8

2 Comments

In a recent STAT interview, FDA Commissioner Robert Califf highlighted a crucial challenge in AI for healthcare systems: "Developers can’t fully guarantee the algorithm’s performance... an AI model’s performance evolves over time and performs differently in different hospitals’ patient populations, unlike a drug. And that model drift is 'happening in health systems.' So, for a health system to say, ‘This is just a manufacturer [problem],’ then the whole thing will break down and won’t work." https://lnkd.in/deHW2_PX While the debate over who bears responsibility continues, it's becoming increasingly clear that the liability for ensuring patient safety will ultimately rest with health systems. This aligns with one of the critical domains of the CLINICAL AI SOCIOTECHNICAL FRAMEWORK, developed during my PhD research: "Monitoring and Support." This domain addresses explicitly the complexities of building, purchasing, and implementing AI systems in clinical environments and includes: User Experience and Feedback: Continuously evaluating and incorporating user feedback to identify areas for improvement. Data Drift and Shift Management: Monitoring and managing changes in data patterns over time to maintain AI model accuracy and relevance. Updates and Review: Regular updates and periodic reviews are implemented to meet emerging needs and technological advancements. Continuous Education and Support: Providing ongoing training and support ensures users can effectively utilize AI systems. Audit and Surveillance: Conduct regular audits to assess AI systems' performance, compliance, and security post-implementation. To assist with these assessments and planning, we developed a checklist available for free here: https://lnkd.in/dHmbAt8s Also, here is a pre-print on the checklist: https://lnkd.in/gKnGxcFH

75

9 Comments

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

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

This paper discusses the integration of theory-informed health IT (HIT) evaluation frameworks into guidelines for evaluating AI in healthcare to enhance technology development, implementation, and scaling. 1️⃣ AI in healthcare promises improvements in treatment quality, safety, and efficiency, but its effectiveness often remains unproven, leading to overoptimism and challenges in practical application. 2️⃣ Evidence from other sectors, like finance and retail, has limited applicability due to the distinct social, economic, technical, and legal challenges specific to healthcare environments. 3️⃣ Theory-informed frameworks can ensure AI technologies are validated, implemented, and adopted safely, reducing risks such as inappropriate care or clinician burnout. 4️⃣ Adoption of AI in healthcare is limited compared to sectors like finance and retail, with most AI applications still in early stages or focused on narrow areas like diagnostic imaging. 5️⃣ Current AI evaluation practices often lack theoretical grounding, focusing more on demonstrating technology functionality rather than on evidence-based outcomes, which impedes effective implementation. 6️⃣ Frameworks with technology, user, and organizational focuses can guide AI development and integration by addressing design usability, user context, and organizational strategies, respectively. 7️⃣ Existing HIT evaluation frameworks can be adapted for AI applications, helping stakeholders navigate technology design, implementation, and optimization to mitigate risks and maximize benefits. ✍🏻 Kathrin Cresswell, Nicolette De Keizer, Farah Magrabi, Robin Williams, Michael Rigby, Mirela Prgomet, Polina Kukhareva, PhD, MPH, FAMIA, Zoie Shui-Yee Wong Z, Philip Scott, Catherine Craven, Andrew Georgiou, Stephanie Medlock, Jytte Brender, Elske Ammenwerth. Evaluating Artificial Intelligence in Clinical Settings—Let Us Not Reinvent the Wheel. J Med Internet Res 2024. DOI: 10.2196/46407 📄 Planning to submit your research to a peer-reviewed, open-access journal? Use the discount code AMJB200 at checkout to save $200 USD on the article processing charge at any JMIR Publications journal. Start your submission here: https://lnkd.in/deuAEN2r. It's important to me that you know this isn't a sponsored post. I'm simply committed to keeping you informed about opportunities to disseminate your research effectively.

205

9 Comments

Get to know Dr Aengus Tran, co-founder and CEO of Harrison.ai, a Microsoft for Startups company, specializing in #AI for #medicalimaging. In this interview from Axios, hear what he has to say about challenges in healthcare; get to know more about his athletic pursuits; and read about his favorite tech device. Bethany Cordes Sinead Hewson Jacques Gilbert Dimitry Tran Randee Stapleton #founder #interview #healthcare #diagnostics #AI #AIinHealthcare https://lnkd.in/eNuYv7Vn

53

1 Comment

🔄 The Case for an Enterprise GenAI Platform: Promoting Reusability to Drive Innovation 🚀   Building on my recent post about the necessity of an Enterprise GenAI Platform (https://lnkd.in/eeQBtV9p), let's delve deeper into its key benefits, starting with reusability. In my view, reusability is a crucial factor that can dramatically accelerate innovation within your organization. Why? Just consider the resource and budget fragmentation occurring in your company as each LoB (line of business) attempts to create their own GenAI POC. The majority of these use cases employ a RAG (retrieval augmented generation) design pattern, essentially applying NLP (natural language processing) to their corpus of unstructured data. Although we’re using the same fundamental architecture, there are numerous distinct codebases now being developed & maintained by different teams, leading to significant redundancy in efforts in large enterprises!   Reusability isn't just a technology component—it's a catalyst for efficiency in the GenAI ecosystem and a game-changer for innovation. By harnessing pre-built design patterns and common components, organizations can: 1/ Accelerate Development: Drastically reduce time-to-market for new GenAI solutions, 2/ Ensure Consistency: Maintain uniform quality and standards across projects, and 3/ Optimize Resources & budget Dilution: Focus on novel challenges instead of reinventing the wheel.   A well-architected Enterprise GenAI Platform should encompass these pivotal features (though this isn't an exhaustive list) to amplify reusability: 1/ Modular Architecture: Enables easy integration and reuse of components across various applications. 2/ Component Registry: A centralized repository for discovering and utilizing reusable assets like pre-trained models, APIs, data pipelines, and even GenAI applications. 3/ Standardized APIs: Facilitates seamless connections between GenAI models and internal systems. 4/ Version Control: Ensures consistency and allows teams to track changes in reusable assets. 5/ Pre-built Templates: Accelerates development by providing templates and best practices for common use cases (think document summarization, formatting documents, summarizing transcripts to generate minutes of the meeting, etc.) By focusing on reusability, organizations can significantly accelerate the development of GenAI use cases. While the exact impact would vary by org size and complexity, reusability through an Enterprise GenAI Platform should lead to substantial efficiency gains.   💬 How are you currently managing resource allocation & fragmentation across different lines of business when developing GenAI solutions?   📢 Subscribe to my newsletter for more insights: https://lnkd.in/g3bdneR7   #aws #ai #EnterpriseAI #GenAI #Innovation #AIReusability #MLOps #AIPlatform #DigitalTransformation #AIStrategy #TechLeadership #FutureOfAI #GenerativeAI #cio #ciso #cto #cdo

11

👉🏼 Generative artificial intelligence and ethical considerations in health care: a scoping review and ethics checklist 🤓 Yilin Ning 👇🏻 https://lnkd.in/emB6jqm3 🔍 Focus on data insights: - 📊 The scoping review identifies significant gaps in the ethical discussions surrounding GenAI applications, emphasizing a lack of operational frameworks. - 🔬 Analysis shows that existing ethical guidelines are often incomplete, particularly when addressing the nuances of different GenAI technologies and their data synthesis capabilities. - 📈 Data from various studies reveals that integrating ethical assessment checklists into peer review processes could enhance the accountability of GenAI research. 💡 Main outcomes and implications: - 📝 Development of a comprehensive checklist allows for improved evaluation of ethical considerations in GenAI research across various healthcare applications. - 🔗 The integration of the checklist into publication systems is expected to promote enhanced ethical disclosures and transparency in GenAI-driven healthcare solutions. - ⚖️ Future research is likely to benefit from this structured approach, thereby fostering a culture of ethical awareness in the rapidly evolving field of GenAI. 📚 Field significance: - 🌐 This work significantly contributes to the understanding of ethical implications in high-stakes fields like health care, where the use of GenAI is increasingly prevalent. - 🔍 The findings underscore the necessity for multidisciplinary collaboration between technologists, ethicists, and healthcare professionals to inform best practices. - 🚀 The recommendations may set a precedent for the ethical evaluation of emerging technologies beyond healthcare, influencing regulatory frameworks globally. 🗄️: [#generativeAI #ethics #healthcare #datainsights #sustainability #peerreview #transparency #accountability #checklist #research]

1

👉🏼 Generative artificial intelligence and ethical considerations in health care: a scoping review and ethics checklist 🤓 Yilin Ning 👇🏻 https://lnkd.in/epDvr6r5 🔍 Focus on data insights: - 📊 The current ethical frameworks for generative AI in health care are inadequate, highlighting a need for more comprehensive guidelines. - 🔬 There is a lack of operationalisable solutions derived from ethical discussions, leading to potential risks in high-stakes areas. - 🧩 Research often overlooks various forms of GenAI, such as image synthesis, and their distinct ethical implications. 💡 Main outcomes and implications: - ✅ Development of an ethics checklist can facilitate the integration of ethical considerations into GenAI research and applications. - 📅 Recommendations for incorporating this checklist into peer review processes could enhance the quality and ethical rigor of published research. - 🌍 The broader implications of ethical discussions on GenAI could lead to improved standards in health care technologies and products. 📚 Field significance: - 🚀 This work positions itself at the intersection of emerging technology and medical ethics, proving critical for future developments in AI applications in health care. - 🔍 By addressing gaps in ethical research, this study may encourage further investigations into the implications of GenAI usage in sensitive fields. - ⚖️ Provides a pathway for researchers and practitioners to better navigate the complexity of ethical considerations in AI technologies. 🗄️: [#generativeAI #healthcareethics #ChatGPT #ethicalguidelines #researchintegrity #AIapplications #medicaltechnology #dataethics]

1

A dynamic presentation by AIMI Co-director Nigam Shah and an engaging panel discussion featuring AIMI affiliated faculty Fatima Rodriguez, MD, MPH, and Stanford colleagues Alex Weihan Chu and Michael Pfeffer demonstrated the vital role of AI in health care, and emphasized the importance of using it safely, ethically, and cost-effectively. Drs. Shah and Rodriguez's fascinating insights -- shared as part of the "Artificial Intelligence: The Future of Medicine & Health Care Is Here" live event -- showed how Stanford is at the forefront of medical innovation by using AI responsibly for the benefit of the Tri-Valley community and beyond. From the 'Green Button Project' to enhancing patient prognostic models, Dr. Shah discussed how AI is playing a pivotal role in advancing clinical decision-making and patient care at Stanford. He also emphasized the need to ensure benefits of Large Language Models (LLMs) in medical implementations. In the panel discussion, Dr. Rodriguez shared her insights on an initiative that is underway to use an AI model to analyze existing imaging data to find patients at risk of heart disease, and then deliver subsequent—and often lifesaving—preventive care. Want to learn more about the potential of AI to advance health care? Watch Dr. Shah's full presentation: https://bit.ly/3XF6QOu Watch the full panel discussion with Dr. Rodriguez: https://bit.ly/4de3GqA The event included opening and closing remarks by Stanford Health Care Tri-Valley President and CEO Rick Shumway. #Datasets #LargeLanguageModels #AIinHealthcare #InnovationInMedicine #PatientCare

108

1 Comment

AI Bias in Healthcare can lead to serious patient harm if left unchecked. Fundamentally, #AI algorithms are trained from the data we provide it. If that data that is non-representative of the population that we are serving, There is a 𝘳𝘦𝘢𝘭 𝘤𝘩𝘢𝘯𝘤𝘦 that we end up benefiting one subgroup of patients, while ignoring or possibly harming another. 👇 𝐋𝐞𝐭 𝐦𝐞 𝐢𝐥𝐥𝐮𝐬𝐭𝐫𝐚𝐭𝐞. Picture a breast cancer screening algorithm that is able to analyse a women's risk using their mammography images. However, the algorithm is trained only on European women's mammography images as it is what's most readily available. But we know that breast density between Asian women and European women differ in real life. So, there is a very real possibility that the algorithm will be able to detect cancers very well in European women, but less so in Asian Women. 🤔 Now imagine if the algorithm then systematically: A. Overdiagnose Asian/African women with breast cancer (Which leads to more invasive investigations such as biopsies/surgery) B. Overlooks suspicious signs of cancer in Asian/African women (Which leads to the cancer not being detected early) 𝐓𝐡𝐢𝐬 𝐢𝐬 𝐡𝐨𝐰 𝐀𝐈 𝐛𝐢𝐚𝐬 𝐜𝐚𝐧 𝐥𝐞𝐚𝐝 𝐭𝐨 𝐩𝐚𝐭𝐢𝐞𝐧𝐭 𝐡𝐚𝐫𝐦. _______________________________ The scenario above is not as hypothetical as one might assume. In fact, it has been extensively demonstrated that #data poverty leads to poorer performing algorithms - particularly in #healthcare. There is a real imperative for AI developers to ensure that: 1. Their training datasets reflect the real world population make up as closely as possible, 2. There aren't any inherent biases baked into the datasets initially, 3. The algorithm is not inherently biased against any one population group. Because if our new data-driven technologies serve only one population, at the expense of others - 𝐓𝐡𝐞 𝐠𝐚𝐩 𝐢𝐧 𝐡𝐞𝐚𝐥𝐭𝐡 𝐞𝐪𝐮𝐢𝐭𝐲 𝐰𝐢𝐥𝐥 𝐨𝐧𝐥𝐲 𝐜𝐨𝐧𝐭𝐢𝐧𝐮𝐞 𝐭𝐨 𝐰𝐢𝐝𝐞𝐧. 👇 Would you agree/disagree? Konrad Dobschuetz | Dr. Elsa Zekeng | Nelson Jordan | Gerald Lip | Josh Au Yeung | Shubs Upadhyay | Xiao Liu, MBChB PhD

98

34 Comments

🌟 A Better Future: Greater Trust of Generative AI in Healthcare 🌟 Just released: My latest Forbes blog, where I share thoughts on some surprising statistics that were revealed in a recent survey from Deloitte. For all the upside that generative AI will offer to patients, consumers, and organizational stakeholders, we must respect the crucial cornerstone of trust. As AI and generative AI are pouring into healthcare biggest challenges and needs, how it lands matters. We need to respect that patients are at the center of care and related transactions. Setting the right foundation is extremely important and wholly necessary. And in my latest blog, I share 6 strategies to help healthcare organizations who will be utilizing generative AI. #Trust #GenerativeAI #ForbesBlog #FutureOfHealthcare

35

2 Comments

This article is an interesting take on the age old dichotomy of governance versus control, now as seen in the world of AI. What we have historically done in other domains is accept that direct control isn't going to work, and instead shifted to guidance, education, delegation, and some amount of post-fact monitoring. Though I agree that in the AI space deviation from control policies can be even harder to detect, I do believe the root challenge to be one we have seen before and expect that the same compensating mechanisms can be put to good use. https://lnkd.in/g7ferE8g

29

2 Comments