By - July 05, 2023
Artificial intelligence (AI) is becoming more prominent across industries, and the laboratory is no exception. AI is not a new concept to laboratories, but as it becomes more integrated into our lives, concerns mount that AI is going to put people out of jobs.
While AI does offer a lot of benefits for the laboratory, and its future capacities are, as of yet, hard to predict, it’s unlikely to replace laboratory professionals or pathologists. For one thing, AI is not yet sophisticated enough to know when it is wrong and will require human oversight for most of its functions for the foreseeable future. AI has made strides in recent years, but it simply can’t replace the human experience.
AI is an umbrella term that refers to “a set of approaches for making machines or computers behave in an intelligentlike fashion,” according to Toby Cornish, MD, PhD, Associate Professor and Vice Chair for Informatics in the Department of Pathology at the University of Colorado School of Medicine.
Beneath that umbrella you find machine learning and deep learning models, all of which are ways for “machines to learn from input knowledge without explicitly being programmed to do so,” says David McClintock, MD, Chair of the Division of Computational Pathology and Artificial Intelligence and a senior associate consultant in the Department of Laboratory Medicine and Pathology at The Mayo Clinic.
People have expressed concern over the recent release of OpenAI’s software ChatGPT, because it appears to be able to communicate like a human. However, ChatGPT is just a large language model, which essentially recognizes patterns within text and generates what the user is asking for based on the data it has seen, Dr. McClintock explains. “It’s not actually thinking; it applies its algorithm(s) to your request and best presents to you an output that matches your request, based on its extensive and diverse set of training data,” he says.
One of ChatGPT’s problems is that it can populate very convincing sounding answers that also happen to be completely made up, according to Ulysses Balis, MD, FASCP, Fellow AIMB, and the James French Professor of Pathology Informatics and Associate Chief Medical Information Officer at the University of Michigan.
That’s because this sort of AI is great at “interpolation”— “where you have a constellation of results that are similar to what was present in the training set.” It is not so good at “extrapolation,” which is finding patterns or making meaning out of variables that it has not been trained on, what Dr. Balis calls “edge conditions”—those conditions that are so rare they aren’t in the training set of data.
Whether you realize it or not, AI is already in the laboratory, , but it’s mostly used in the core lab and other instrumentation, Dr. McClintock says. “You have instrument manufacturers who are including AI algorithms that analyze the data as it comes off the line and give you results.”
Currently, AI is most effective at narrow tasks, Dr. McClintock explains, which is “using a specific tool for a specific purpose. Most AI tools typically do a single task.”
Dr. Cornish tries to allay the concern about AI taking laboratory and pathology jobs. “There is no new paradigm where you can implement AI and suddenly the AI takes responsibility for all the decisions it’s making. We still need a human there,” he says. “At the end of the day, even a lab that was relying entirely on AI, the responsibility would have to fall on someone.”
Dr. McClintock agrees, saying, “It’s going to augment your job.” He explains a scenario regarding kidney transplant biopsy scoring. Pathology looks for BANFF criteria, which include factors such as fibrosis and inflammation. Pathologists have to make a guess to determine total inflammation. “They’re not counting every lymphocyte or cell. They’re giving you a kind of gestalt,” he says. AI could potentially quantify the data to come up with a consistent, standardized number that accurately quantifies the inflammation and better classifies when there is concern for organ rejection.
“Ideally that will translate to better patient care, better patient outcomes, more streamlined [treatments] because we don’t have the variation of three different pathologists interpreting slides three different ways,” Dr. McClintock says.
AI could also potentially help fill gaps with the laboratory professional shortage, Dr. McClintock says. “We are constantly fighting to bring people in. We can’t seem to get enough. So, there are a lot of opportunities to look at automating with AI.” Particularly those tasks that feel menial or that people don’t enjoy doing. “Just like how Amazon prompts people with ‘Looks like you might want to buy this item,’ in pathology you could have it prompt the pathologist ‘Do you want this data? Would you like to order this lab test?’” Or it could prefill reports based off measurements taken off a slide that were quantified with an algorithm, he explains.
Though ChatGPT is nowhere near ready to engage in thorough research or diagnosis, what it is useful for, Dr. Balis says, is “production work where you’re dealing with common things.” For example, for a routine case of results, “ChatGPT could give you a nice narrative text, which is grammatically correct and has all the data elements in the right order. Huge time saver!”
Another area where AI can improve pathologists’ and laboratory professionals’ efficiency is by saving the pathologist the trouble of manually looking for micrometastases in cancer slides, by going through galleries of images and marking the ones it thinks are cancer. Then a pathologist would simply have to go through the pre-marked slides and determine whether it is, or isn’t, cancer. “That’s an example where it transforms workflow from screening, which is stressful and time-consuming, to directed review, which is fast and adds a higher level of performance,” Dr. Balis says.
With all AI, Dr. Balis says, “The analogy is, you want to have an adult in the room. And that’s where pathology and laboratory medicine expertise really shine. For recognizing, through expertise and experience, what is unusual or unique and then bringing that to bear on the problem the additional investigation needed to arrive at that new diagnosis or insight.”
AI tools don’t have the ability to do that, he says. “They’re good at recognizing patterns that have already been seen, but in terms of carrying out the scientific method of hypothesis generation and further investigation to come to a real answer, we’re not there yet.”
Dr. Balis’ recommendation is that these tools be used to enhance productivity, but with expert oversight. “You have to know their limitations. And one of their limitations is if you give AI a set of starting conditions that have never been characterized before, it will give you an answer that’s very believable but completely wrong.”
Dr. Cornish adds, “People should be interested in AI because it has the potential make their jobs easier, but right now AI isn’t really capable of replacing people, especially very technical people.” He notes that, “AI is a long way off from displacing people that physically interact with patients or even with instruments. We’re not going to be seeing AI phlebotomists, AI lab techs, med techs, cytotechs, etc.”
While AI may feel like new territory for laboratories right now, as these tools become more sophisticated they should be able to enhance and support laboratory professionals and pathologists to make their jobs easier; they won’t take them away.