Abu Dhabi-based M42 Health, has released a promising new open access clinical large language model called Med42. The unveiling of this 70 billion parameter model marks a major milestone in expanding open access to powerful AI capabilities aimed at transforming healthcare.
Finetuned from Meta's Llama-2 - 70B model, Med42 demonstrates significantly improved performance over previous open source medical AI models. When evaluated across medical question answering datasets, the model achieved up to 72% accuracy in a zero-shot evaluation on the United States Medical Licensing Examination (USMLE), and outperforms OpenAI's ChatGPT 3.5. This underscores Med42's potential to aid clinical decision-making by providing doctors with quick access to synthesized medical knowledge.
The M42 Health AI team developed Med42 using its extensive curated dataset of medical literature and patient records. Fine-tuning was performed on the Condor Galaxy 1 supercomputer through a collaboration between M42, Cerebras, and Core42 (a subsidiary of M42). Researchers from Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) also evaluated the model's capabilities.
Hasan Jasem Al Nowais, the Group CEO of M42, stressed the revolutionary potential of Med42 in democratizing access to healthcare knowledge. Med42 is not just an embodiment of M42’s commitment to spearheading healthcare innovation, but also an indicator of the influential role that the United Arab Emirates aspires to play in the global AI landscape.
Med42’s versatility is one of its standout features. Envisioned as an AI assistant, it promises to revolutionize clinical decision-making processes. Its applications range from devising patient-specific treatment plans based on their medical history to aiding researchers in sifting through vast medical literatures efficiently.
To foster collaboration and continuous development, the code and weights of Med42 have been made available on Hugging Face, inviting a broad spectrum of scientific evaluation and feedback. Taking cues from the licensing terms of Meta’s Llama 2 model, Med42 can be accessed freely for research and non-commercial use, albeit with necessary controls, keeping in mind the challenges and responsibilities of AI implementation in healthcare.
While initial results are promising, the researchers caution Med42 requires extensive real-world validation before clinical use. Possible risks include incorrect or harmful outputs, and perpetuating biases in the training data. M42 stresses responsible testing will be critical as Med42 progresses from benchmarks to potential real patient benefits.
Med42 demonstrates the impressive progress of medical AI, while highlighting the continued need for ethics and safety in developing these technologies. Its open access availability will empower researchers globally to move the field forward responsibly. If rigorously validated, models like Med42 could one day aid doctors in improved decision-making and enhance healthcare access worldwide. For now, its release represents a major milestone in healthcare AI - one that will require ongoing transparency and collaboration to reach its full potential.