M42 Unveils its Next-Gen Clinical LLM at Abu Dhabi Global Healthcare Week
M42, a global tech-enabled health powerhouse, has launched the next generation of its open-access generative AI Clinical Large Language Model (LLM), Med42, during Abu Dhabi Global Healthcare Week. Building upon the success of the initial launch, Med42 represents the next evolution in AI-driven health innovation, poised to advance healthcare AI and transform patient care, while helping reduce administrative burden for clinicians.
M42 unveiled Med42 in October 2023, setting a new standard for accuracy, efficiency, and accessibility in healthcare AI. With a score of 72 percent on the United States Medical Licensing Examination (USMLE) Sample Exam questions, the initial release outperformed OpenAI’s GPT3.5 and Google’s MedPaLM models. The model counts over 8,200 downloads to date from independent developers who have tested and validated the model’s performance. It continues to attract attention and acclaim worldwide.
This second iteration is a fully optimized 70-billion parameter finetuned model based on Llama-3, Meta’s latest version of open-access LLMs. The new model’s performance is on par with proprietary (closed) models such as OpenAI’s GPT-4 and Google’s Med-Gemini, with a zero-shot accuracy score of 85.1% and a maximum accuracy score of 87.3% using specialized prompting on the USMLE. This highlights Med42’s efficiency and accessibility compared to larger, closed models and solidifies its position as a frontrunner in open-access clinical LLM technology.
Ashish Koshy, Group Chief Operating Officer of M42, says, “We are excited to release the second version of Med42, which represents a significant step forward in our mission to responsibly harness the power of AI for the benefit of patients and healthcare professionals. This release brings us closer to our vision of a composable multimodal AI platform that integrates expert models across health data modalities, including text, genomics, health records, and imaging. By enabling these models to collaborate and generate insights from diverse data sources, we aim to tackle complex healthcare challenges that require a multimodal approach. Our sustained efforts in developing the Med42 platform underscore our commitment to making this vision a reality.
Ashish further emphasized the importance of considering both leading and lagging indicators when assessing Med42’s performance: “Benchmarks like the USMLE serve as valuable leading indicators of a model’s capabilities. However, the true measure of success lies in its real-world performance across specific applications. To that end, our interdisciplinary teams of AI experts and physician-scientists are actively conducting research studies to comprehensively evaluate and refine Med42’s performance in clinical settings.” Med42 showcases the potential of automation in healthcare by streamlining documentation processes, alleviating administrative burdens, and supporting healthcare professionals. The AI model facilitates the retrieval of relevant insights, searches through patient histories, and summarizes records. Moreover, Med42 is being evaluated for its ability to craft tailored treatment strategies by analyzing patients’ medical histories and determining optimal courses of action. The model also shows promise in serving as a clinical decision support tool for physicians, assisting pharmacists with accurate dosage decisions, and helping researchers efficiently review treatment literature.
Med42 is accessible through the HuggingFace platform, enabling extensive testing, contribution, review, and assessment by the global scientific and developer community. By making Med42 widely available, M42 aims to catalyze global innovation in AI, empowering experts worldwide to refine and expand the model’s applications across diverse medical domains, ultimately driving advancements in patient care and medical research.
Last Updated on 7 months by News Desk 2
