✨ Introducing Med-Gemini, our new family of AI research models specialized for medicine! ✨ Med-Gemini models are tuned from Gemini, building on its advanced reasoning, multimodal, and long-context capabilities to unlock new possibilities for medicine. Our research includes: - State-of-the-art performance across multiple benchmarks, outperforming other models. This includes an impressive 91.1% accuracy on the MedQA benchmark. - Ability to generate reports for 2D images as well as, for the first time, generate reports for complex 3D datasets like 3D brain CTs. - Enhanced ability to analyze medical images, videos, genomics, health records, and more. - Advanced reasoning and efficient use of web search for accurate medical answers. - Long-context abilities to summarize health records and analyze research papers. - The first large medical model to predict health outcomes from genomic data converted to polygenic risk scores. Further research is needed, but we’re excited by the potential of this research to support healthcare organizations, clinicians, and patient workflows. Learn more in our blog: https://goo.gle/4bkmyEa
Intersting share Google Health! Interested to know what data source this was trained on. As we embrace these AI-driven innovations, how do we ensure the ethical use and integration of such powerful tools in healthcare? Balancing technological advancement with patient safety and data privacy should be crucial.
Google Health - Is this available to the public? Could someone use this to make a "Personalized Super Symptom Checker" based on their personal EHR and the AI help of Google Bard/Gemini (which was trained in medical concerns) - and maybe to offer this as a chatbot (Patient Listener) sorting through the patient's story, which often the doctor does not have time to listen to? If interested, I made this comment based on the concepts in my book - "Future Healthcare Today: How Technology is Revolutionizing Holistic Wellness” - https://books2read.com/u/3nBMDo Greg Corrado Joëlle Barral
This will undoubtedly fuel better consideration of a multitude of AI applications in healthcare...tackling a key obstacle to adoption. For more about challenges to healthcare's adoption, I wrote on this earlier today https://www.linkedin.com/pulse/great-healthcare-ai-paradox-stakes-too-high-scott-vudvc/?trackingId=yUXCw7PqSsuOAayvkczm8g%3D%3D
Pre-print manuscript of this work (for scientific readers). https://arxiv.org/pdf/2405.03162
In this blog post (https://research.google/blog/advancing-medical-ai-with-med-gemini), Google Health wrote that, "We then benchmark Med-Gemini models on 14 tasks spanning text, multimodal and long-context applications. In the popular MedQA US Medical Licensing Exam (USMLE)-style question benchmark, Med-Gemini achieves a state-of-the-art performance of 91.1% accuracy, surpassing our prior best of Med-PaLM 2 by 4.6% (shown below)." Is the MedQA US Medical Licensing Exam (USMLE)-style question benchmark reflective of step 1, step 2, or step 3 of the USMLE exam? Or all the steps of the USMLE exam? Thanks!
👀🚀🚀🚀 congratulations, who can I talk to about API? Google Health
A potential game changer!
Google Health, again leading the way for breakthrough innovation and social Impact. This is going to life-changing… literally. We are excited to work with our academic healthcare customers to accelerate development and joint research.
Predicting health outcomes from genetic data with polygenic risk scores is a big step forward. What challenges did you encounter while developing this feature?
Physician | Data Scientist | Engineer
3wGoogle Health please provide clarification on a discrepancy between a figure in the arXiv paper “Advancing Multimodal Medical Capabilities of Gemini” and the images in this post. Figure 9 in the arXiv paper reads that the model detected a "basilar ganglia" infarction that was missed by a Radiologist. It had previously been pointed out that the "basilar ganglia" is not a real anatomic structure. This means that the Med-Gemini output was incorrect despite being highlighted as correctly finding something missed by a Radiologist. There is hypodensity in the left internal capsule and seemingly the claustrum / extreme capsule (not part of the basal ganglia) on the image that is consistent with an infarction. I see an updated image in this LinkedIn post with the same figure saying "basal ganglia" infarction. I would like to confirm that this is unedited output from Med-Gemini and not an undisclosed edit to the figure in the original manuscript that does not correspond to verbatim Med-Gemini output. If it is unedited Med-Gemini output then I would like to know if any changes were made to Med-Gemini in the interim or if this was simply passing the same test data into the model and getting a different (correct) model output this time.