When we think of AI’s potential in healthcare, images of robots performing surgery in the operating room, scanning images of moles for early cancer detection, and even researching new drugs to cure chronic diseases come to mind. However, natural language processing (NLP) is another aspect of AI that is just as impactful regarding its capabilities in healthcare.
One of the most impressive examples of this is Google’s Med-PaLM, a large language model (LLM) designed to provide high-quality answers to medical questions.
Existing Google LLMs like PaLM and Flan-PaLM provide the foundation for Med-PaLM. PaLM is a Google LLM with about 540 billion parameters and is based on Google’s Transformer architecture, which is responsible for applications like GPT-4. Flan-PaLM is a variation of PaLM that focuses on FAQs and dialogs. Med-PaLM is the result of adapting Flan-PaLM to the medical industry using instruction prompt tuning, an efficient low-cost method of guiding LLMs to perform specific tasks.
As a result of instruction prompt tuning, Med-PaLM shows great accuracy and high performance. It was the first AI system to achieve a passing score on the US Medical License Exam (USMLE), which focuses on multiple-choice questions with 67% accuracy in late 2022. However, Med-PaLM 2 is the newest iteration of the system, and it boasts an 85% accuracy on par with experts on the USMLE. Med-PaLM 2’s capabilities expand beyond simply multiple-choice questions, and Google claims that its long-form answers perform “encouragingly” based on both clinician and non-clinician evaluations made on 14 different criteria. However, Google states that Med-PaLM 2 does not currently meet their standards when it comes to product performance.
LLMs like Med-PaLM can be used to create chatbots that can answer personalized questions for people. If people have questions about symptoms they’re experiencing, it’s easier for them to resort to asking a chatbot as opposed to calling a clinic or their doctor to get an answer. In addition, it is also much quicker for people to ask a chatbot. Most importantly, studies comparing ChatGPT’s medical advice compared to that of a doctor showed that ChatGPT provided higher quality responses according to three licensed medical professionals. Taking into account that ChatGPT is not designed specifically for medical purposes, LLMs like Med-PaLM would likely perform extremely well in a study like this and be great as the basis of a medical chatbot. This frees up time for medical professionals that may have more important tasks to do than responding to common questions that a chatbot could answer.
Google doesn’t believe that Med-PaLM 2 is yet ready for implementation into real applications despite the fact that it is already at an expert level. This goes to show that when it is released it will be extremely impactful, especially considering its extremely fast rate of improvement of 18% USMLE accuracy in less than a year. Regardless of its availability, Med-PaLM 2 is an example of how NLP can positively change healthcare.
See more like this:
