Google announced the release of MedLM, two foundational models built upon Med-PaLM 2 and specifically designed to answer medical queries, generate insights from unstructured data, and summarize medical information.
The tech giant emphasized that it leveraged its experience with large language models in healthcare institutions and learned that effective artificial intelligence models are designed to address specific use cases.
Google has designed a large-scale MedLM model to handle challenging tasks, while a smaller version of MedLM is designed in a way that makes it adjustable and expandable across various tasks.
The company stated that summarizing conversations using one model may be preferable, while searching for medications using a different model may be more effective.
Google announced its partnership with Augmedix, a startup in clinical documentation, to utilize MedLM technology in converting data into medical notes. BenchSci, a drug research and development company, announced that they are using the technology to discover drugs and accelerate their development process.
Similar to the collaboration between Google and Accenture, Google is also working with Accenture to enhance the adoption of healthcare institutions and provide assistance to healthcare providers in using MedLM.
The company collaborated with Deloitte and healthcare service providers to assist care teams in obtaining information from sources such as benefit documents and provider certifications. This is aimed at helping call center agents in identifying suitable providers for members.
The MedLM model is now available to Google Cloud customers listed in the allowed list through the Vertex AI artificial intelligence development platform.
The company stated: “We will conclude the year by expanding the reach of MedLM to more healthcare institutions. We are excited about the progress and future possibilities, and our ongoing efforts to advance cutting-edge research in health and life sciences.”
In March, Google tested their large language model Med-PaLM 2 using questions similar to the US Medical Licensing Examination. The test was conducted at the level of advanced experts, and the model achieved an accuracy exceeding 85%.
The large language model also succeeded in MedMCQA, a multiple-choice dataset designed to address actual medical entrance exam questions.
Upon the first appearance of the large language model Med-PaLM, it showed an accuracy of 67.6% in answering MedQA questions similar to the US Medical Licensing Examination. The accuracy increased to 85% after two months, and currently, the accuracy has reached 92.6%.
