Accelerating Telemedicine with Large Language Models
On Tues ., Sept . 5 , the GLMS Foundation Senior Physicians Speaker Series welcomed Michael Yared to speak on “ Accelerating Telemedicine with Large Language Models .” Michael is the CEO and Co-Founder of Echobind , a software development agency that creates web and mobile applications . He helps businesses navigate risk involved in conceiving , developing and launching software . He is a Louisville native and the son of Dr . Sam Yared , Senior Physicians Committee Chair .
A large language model ( LLM ) is essentially a computer program trained on a vast amount of text to understand and generate human-like language . We give an LLM text input , and it results in text output from the machine . A commonly known example of this is OpenAI ’ s Chat GPT . For many LLM ’ s , their training set is the entirety of the internet with some versions reaching over 1 trillion parameters . The theory is that the more you train it on and the more options you give it , the better it works .
When working with computers , they use numbers , not letters . Imagine a list of every single word in the English language ( roughly half a million ) and every word is given a corresponding number . Sentence input is then converted to numbers and the computer can compress it and see patterns . This can be used in the case of predictive text . For example , if you type “ good ” it may know that it is more likely the next word will be “ morning ” than the next word being “ time ” based on all of the data points it has studied . As you add words , it gets easier for the model to know what word is coming next .
The invention that made this a lot more usable and possible was the transformer . Imagine reading a complex medical case with symptoms , history and test results . You wouldn ’ t just read it from start to end , you ’ d focus on the relevant parts to diagnose the patient , so you ’ d jump back and forth between earlier symptoms to test results , etc ., which is how the transformer architecture works . It ’ s reading data by focusing on the most relevant parts and giving those extra attention to make it more accurate .
For example , if you told the model you were born in Louisville , Kentucky , and asked it to fill in the blank : “ The best football player in the world is ” the model would tell you it ’ s Lamar Jackson . But if you changed the location to South America , the model would tell you it ’ s Lionel Messi .
In application with telemedicine , LLMs can be useful in accessibility , such as sending personalized messages for appointment
by KATHRYN VANCE
and medication reminders , as well as answering questions about medical schedules , dosages and potential interactions . LLMs can also be useful in convenience and cost-savings . For example employees of a company needing answers about their health plan used to pick up the phone or open a chat box to speak with a real agent when they had questions – now , they can do the same thing , but with a chatbot .
In 2021 , Michael ’ s team partnered with Teladoc and Amazon to create the world ’ s first voice assistant telemedicine application . Just by asking Alexa to speak to a doctor , you could get a call back from a physician within 10 minutes . Now imagine if Alexa had an LLM installed . There are several benefits – you wouldn ’ t need an internet connection , the voice can be synthesized to the local population and most importantly , your wait would drop to 0 minutes . While he recognizes there is a whole slew of legal issues that would need to be navigated before a device acting like a physician could make it to market , today you can already use LLMs to diagnose issues , most just with a disclaimer to seek advice from a medical professional for serious issues .
Looking forward , there are many advances to expect in this fast-paced industry . Hospitals will have different models and may even take a more aggressive stance on certain treatments and be at the forefront of recommending a certain procedure . There will be cultural and linguistic mediation , translating medical advice to culturally appropriate , easily understandable language . We will also see physician training scenarios . LLMs are good at answering questions , but also asking them . If you were to give it a scenario , a model can propose complex medical scenarios from which students can attempt to learn .
As health care and technology change at every corner , it ’ s important to celebrate the advancements while we remember to appreciate the old-fashioned techniques as well ( and yes … my computer did just predict the end of that sentence as I was typing it ).
Kathryn Vance is the Communication Specialist at the Greater Louisville Medical Society .
November 2023 23