Google Research and DeepMind recently introduced MedPaLM, an open-sourced large language model specifically designed to answer medical queries. The launch acts as a reminder of Google's leading position in the AI race, amidst the rising popularity of ChatGPT, an AI chatbot created by OpenAI.
Why it's notable:
Compared to ChatGPT, which is designed to generate human-like text, MedPaLM is specifically tailored to the medical field, with the ability to analyze medical language and abbreviations. Beyond understanding language inputs, MedPaLM can also summarize medical records, and identify relevant information within those records.
MedPaLM is benchmarked on MultiMedQA, an open-source benchmark that combines a free-response dataset of 3,375 medical questions frequently sought online, with six open-question answering datasets covering professional medical exams, research, and consumer queries. The model is designed to evaluate the quality of human responses along several axes, including factuality, precision, potential harm, and bias, therefore ensuring unbiased answers are produced.
MedPaLM has been shown to yield results that are on par with a medical expert's judgment. Doctors have determined that 92.6% of Med-PaLM responses were on par with clinician-generated answers.
Industry Implications:
Google joins a number of big tech players operating in the AI-driven healthcare space, most notably Microsoft who are working with OpenAI to employ GPT-3 to improve healthcare teams’ efficiency. Meta AI also released Galactica, which was designed to generate literature reviews on any subject. However this endeavor was criticized for producing unreliable results. Google’s launch of MedPaLM comes as ChatGPT’s dataset is evolving at rapid speed. However Google appears well placed to compete. For instance, MedPaLM can be scaled up to 540 billion parameters, which indicates higher performance with increasing scale, whereas Chat-GPT reportedly has about 175 billion parameters .
MedPaLM has the potential to revolutionize the way medical information is processed and used. Clinicians, researchers and consumers will be able to access medical information much quicker and more accurately than before, which could greatly increase accuracy (HCPs determined that 92.6% of Med-PaLM responses were on par with clinician-generated answers (92.9%)) and accessibility of information in the medical field. Compared with OpenAI’s ChatGPT, MedPaLM’s focus on the medical field appears to give it an edge that could make MedPaLM an invaluable resource for the industry in the future.
