A new family of artificial intelligence (AI) models from Google was introduced to the medical industry last month. Although the public cannot currently access these AI models, or Med-Gemini, the tech giant has published a pre-print version of a research paper detailing its capabilities and workings.
According to the business, AI models perform better in benchmark tests than GPT-4 models. Its long-context capabilities, which enable it to process and comprehend research publications and health data, are one of this AI model’s unique features. The study has been published on arXiv, an open-access online repository for academic publications, and is presently in the pre-print stage.
“I’m very excited about the possibilities of these models to help clinicians deliver better care, as well as to help patients better understand their medical conditions. AI for healthcare is going to be one of the most impactful application domains for AI, in my opinion,” Jeff Dean, Chief Scientist, Google DeepMind and Google Research, said on X.
The Med-Gemini AI models are based on the Gemini 1.0 and 1.5 LLMs. Med-Gemini-L 1.0, Med-Gemini-M 1.0, Med-Gemini-M 1.5, and Med-Gemini-S 1.0 are the four versions in total. Every multimodal model has the ability to generate text, image, and video outputs.
When displaying results for complex clinical reasoning tasks, the models are combined with an online search that has been improved through self-training to make the models “more factually accurate, dependable, and nuanced.”
The business goes on to say that the AI model has been improved for faster long-context processing. Even when the questions are not precisely queried or when the chatbot needs to examine a lot of medical records, better long-context processing would enable it to provide more focused and accurate responses.
Google statistics show that in the GeneTuring dataset, Med-Gemini AI models performed better than OpenAI’s GPT-4 models on text-based reasoning tasks. On the MedQA (USMLE), Med-Gemini-L 1.0 likewise achieved 91.1% accuracy, surpassing the performance of its predecessor, Med-PaLM 2, by 4.5%. Notably, neither beta testing nor the general public can access the AI model. The corporation is probably going to make some more tweaks before selling the model to the general public.