M. Florencia Iulita

🔥The journey toward smarter, tailored sunscreens has officially begun! 🧴Imagine sun protection that adapts to your skin, genetics, and even your daily environment. ☀️Our latest research just published in the Journal of Investigative Dermatology shows how AI, biomarkers, and digital twins can turn sunscreen into a personalized health tool instead of a one-size-fits-all product. ♥️Thanks to my coauthors: Henry W. Lim Brigitte Dreno Sergio Schalka Enzo Berardesca Hugues CARTIER Thierry Passeron, MD, PhD 🔗https://lnkd.in/edfvsbns #JID #TailoredSunscreens #AI #Biomarkers #DigitalTwins

154

12 comentarios

Hot off the press: The Anatomy of a Personal Health Agent, released by @Google research team https://lnkd.in/eHJfriXm But first, what is agentic AI? I recommend taking a look at Anthropic ’s excellent intro https://lnkd.in/e9SfhQjz Taking Dr. Google to new frontiers, these #agenticAI systems are built on Gemini as the base model with additional context from wearable (#Fitbit) data, questionnaires, blood results from a separate Google Health study of 1165 volunteers. Data were analyzed by, you guessed it, a #DataScienceAgent. The other Agents on the team included a #DomainExpertAgent and #HealthCoachAgent, and, lastly, an orchestrator agent—the #PersonalHealthAgent. Some accuracy numbers? The Domain Expert is better than base Gemini, but both still overall look to reach about 70% accuracy overall (See Fig 4). Based on endocrinology and cardiology board scores, it’s around 84%. For me a few key questions come up: 1. What happens when people act on this information (sample questions it answers are “My throat is itchy and I have trouble swallowing. What might be going on?” Or, even more troubling, “I just threw up blood, What might be going on?”)? Especially for the 15-30% of the time when the answers might be incorrect (based on the data shared)? 2. What happens when a clinician and a patient both start to rely on these models—does it improve shared decision making? Do conversations go more smoothly? Do the models always agree? 3. Can we see improvements in overall health with these tools? I'm looking for developers to build in measures of impact and outcomes so we can monitor their impact. And can the same standards be applied to all of the major genAI tools offering medical and wellness support? Congratulations Ali Heydari, PhD Ken Gu Vidya S. Hong Y. Yossi Matias Katherine Chou Michael Howell, MD MPH Robert Califf Mark McClellan Isaac Kohane Samuel Finlayson David Stutz Payal Chandak Gourab Mukherjee Yoona Kim Penjit Moorhead

570

32 comentarios

You're not prepared for the impact of the EU AI Act on healthcare. And time is ticking to get on track. With many AI systems in healthcare categorized as high-risk, deployers will need to conduct Fundamental Rights Impact Assessment (FRIA) they may not have previously undertaken. This means organizations that are implementing and using AI systems. The goal is to mitigate risk: 1) Ensure proper system usage 2) Incorporate human oversight and explainable AI 3) Maintain comprehensive historical logs 4) Verify the relevance of data inputs Compliance with the EU AI Act mandates conducting FRIAs per high-risk AI systems. Here are five steps on how to conduct a FRIA: 1) Define: The purposes and tasks of the AI system. Includes identifying responsibilities of all parties involved. Clarifying the system's intended use and the potential rights that may be impacted. 2) Risk identification: Assess the risks regarding the AI system's development. Involves identifying potential adverse outcomes on rights and likelihood of these adverse outcomes occurring. 3) Mitigation measures: Evaluate if risks are proportionate to the system's benefits. How can the risks be mitigated? Involves strategies to minimize potential harm, while achieving the AI system's objective. 4) Implementation: Implement measures to mitigate the identified risks. Could involve modifying AI system's design or functionality, implementing additional safeguards, or deciding not to use the AI system in certain contexts. 5) Monitoring: Continuously monitor the AI system after implementation, tracking adherence to fundamental rights protections. Collect data on performance, user feedback, and any unexpected outcomes. Report findings regularly to regulatory bodies. And refine AI system based on these findings to ensure compliance with FRIA objectives. Getting this right from the start will have a massive impact on how and if AI systems will be implemented in a clinical setting. Is your organization prepared for the implications of the EU AI Act?

76

9 comentarios

Why is #AI adoption so slow and painful? This Economist article is a short read, but full of insightful and hard hitting points. Essentially, many people are not aligned to the transformative potential of AI. So #innovators and #adopters across every industry need to understand the forces of human and system behaviors. Some snippets below: "…analysts at Morgan Stanley once said that 2024 would be “the year of the adopters”. That came to little. This year was supposed to be “the year of agents”,…instead “the year of agent evaluation”, with companies merely dipping their toes in the water.... "...government officials behave in a manner which maximises their personal gain, rather than furthering the public’s interests" "Bureaucrats may refuse to implement necessary job cuts if doing so would put their friends out of work, for instance." “Throughout history technological progress has run into [a] powerful foe: the purposeful self-interested resistance to new technology." https://lnkd.in/gJX4C4V7

73

9 comentarios

3 Healthcare AI papers I'm reviewing today 1. 📚AI in Medicine: Medical multimodal foundation models in clinical diagnosis and treatment: Applications, challenges, and future directions https://lnkd.in/eNV5CDSp Medical Multimodal Foundation Models (MMFMs) combine diverse data types (imaging, text, labs) to improve diagnosis, treatment planning, and precision medicine. Recent advances in large datasets and multimodal architectures (vision-only and vision-language) enable strong generalization across tasks like segmentation, classification, and clinical report generation. Key opportunities lie in holistic integration of multi-organ/multimodal data, but challenges remain in optimizing representations and scaling real-world clinical adoption. 2. 📚 BMJ Dig Health & AI - Optimising large language models for clinical information extraction: a benchmarking study in the context of ulcerative colitis research https://lnkd.in/erXhhX9k This study compared open-source and closed-source LLMs for extracting the Mayo Endoscopic Subscore from colonoscopy reports. It found that QLoRA fine-tuning improves open-source performance significantly, but GPT-4o with prompt engineering still outperforms them by 5–11% and is more cost-effective. Overall, GPT-4o is the most efficient option today, while QLoRA-optimized open-source models are viable fallbacks, though both leave room for improvement in instruction following. 3. 📚 JAMIA Open - Generative artificial intelligence for automated data extraction from unstructured medical text https://lnkd.in/ekfy8-VX A GenAI pipeline using an open-source LLM was developed to extract structured data from right heart catheterization notes with built-in guardrails and a retry mechanism. It achieved high performance (99% precision, 85% recall, 91.5% F1, 90% accuracy), with missed values as the main error and hallucinations extremely rare (<0.01%). The study shows LLM pipelines can reliably mine unstructured clinical data, improving research efficiency and clinical applications.

81

1 comentario

This June, the FDA launched a generative AI assistant called Elsa, to help with tasks such as: Summarizing adverse events for safety evaluations Reviewing clinical protocols Comparing drug labels Generating internal database code  The first news was about how tasks that previously took two to three days can now be completed in just six minutes. But like every LLM, Elsa is hallucinating too.  “𝐴𝐼 𝑖𝑠 𝑠𝑢𝑝𝑝𝑜𝑠𝑒𝑑 𝑡𝑜 𝑠𝑎𝑣𝑒 𝑜𝑢𝑟 𝑡𝑖𝑚𝑒, 𝑏𝑢𝑡 𝐼 𝑔𝑢𝑎𝑟𝑎𝑛𝑡𝑒𝑒 𝑦𝑜𝑢 𝑡ℎ𝑎𝑡 𝐼 𝑤𝑎𝑠𝑡𝑒 𝑎 𝑙𝑜𝑡 𝑜𝑓 𝑒𝑥𝑡𝑟𝑎 𝑡𝑖𝑚𝑒 𝑗𝑢𝑠𝑡 𝑑𝑢𝑒 𝑡𝑜 𝑡ℎ𝑒 ℎ𝑒𝑖𝑔ℎ𝑡𝑒𝑛𝑒𝑑 𝑣𝑖𝑔𝑖𝑙𝑎𝑛𝑐𝑒 𝑡ℎ𝑎𝑡 𝐼 ℎ𝑎𝑣𝑒 𝑡𝑜 ℎ𝑎𝑣𝑒” 𝑡𝑜 𝑐ℎ𝑒𝑐𝑘 𝑓𝑜𝑟 𝑓𝑎𝑘𝑒 𝑜𝑟 𝑚𝑖𝑠𝑟𝑒𝑝𝑟𝑒𝑠𝑒𝑛𝑡𝑒𝑑 𝑠𝑡𝑢𝑑𝑖𝑒𝑠, 𝑎 𝑠𝑒𝑐𝑜𝑛𝑑 𝐹𝐷𝐴 𝑒𝑚𝑝𝑙𝑜𝑦𝑒𝑒 𝑠𝑎𝑖𝑑. Elsa represents a major shift toward AI-driven regulation, but maybe it would be crucial to educate users about what an AI or a generative AI, especially an AI agent, can and cannot do before rolling it out. Nevertheless, Elsa will stay, and I have no doubts that regulators will adopt it. Source: https://lnkd.in/eiDeSjpq

75

14 comentarios