There is no doubt that interest in artificial intelligence (AI) and machine learning (ML) is booming in healthcare. Of course, at the end of the day, AI is simply a form of advanced analytics that uses artificially created algorithms to calculate beyond the speed or ability of any human being, so in that sense the journey of healthcare to AI is just a subset of its larger journey to advanced analytics, a journey that has taken many years to prepare.
What’s interesting now, at the very end of 2022, is that the thinking around artificial intelligence is changing rapidly, even as the use cases to which AI is applied are multiplying very rapidly. I remember something that happened about six years ago now, when the top executives of a well-known provider organization publicly announced that they were going to capture millions of diagnostic images and millions of points of data, in the field of diagnostic imaging, and essentially creating a huge lake of data and images, from which they planned to derive a huge number of data points that would help radiologists diagnose in a number of key clinical areas (such as breast imaging, cardiothoracic imaging, etc.). Long story short, this massive experiment ended in dismal failure, as the images and data points simply couldn’t be mined in the way that had been imagined. But even then, clinicians and IT managers in individual patient care organizations were beginning to create algorithms that they could test in diagnostic situations, as well as others related to clinician workflow, and so on.
As Brian Patterson, MD, practicing emergency physician and assistant professor in the Department of Emergency Medicine at the University of Wisconsin School of Medicine, and director of medical informatics for predictive analytics at UW Health, Madison , Wisconsin, told me this spring, “As for the Gartner Hype Cycle, a few years ago there was so much hype around AI. In the meantime, I have to say that I don’t think that the door is closed to generalizable patterns. We might be able to derive patterns across a large number of use cases. But in many cases, it’s very fair to say that just because your DSE [electronic health record] model works in one place, it will not necessarily work in another. So it varies depending on the patient populations you’re building the models on; and second, in terms of how data is entered into the EHR [electronic health record]. For example, in one facility, only fluid boluses will be recorded, and in another facility, each medication ordered would come with an order for the fluid that pushed it. And just based on that difference, a patient may appear to have received more or less fluid. And this is one of the great threats to generalizability. And what I totally agree with is that models need to be developed by specific teams in specific organizations, in specific work environments; in other words, where models are given to physicians as part of their workflow, how they are used, these must be individualized. And a few recent publications have pointed this out.
Meanwhile, Dr. Patterson and his colleagues at UW Health have indeed made progress in several areas. “We have at least three or four in use, and several more in the works,” he told me. “We have an algorithm for the likelihood of falls after ER visits. Others being implemented are algorithms on the risk of developing sepsis; implemented a clinical deterioration algorithm, albeit one that we implement from a vendor, albeit developed by UW faculty,” among other efforts. An absolutely essential learning, Patterson said, is that “No matter how well you think a model will work from outside of your system, you have to validate it against your data. The other lesson is that many models can provide insights, but if we want to change patient care, we need to find actionable insights and present it to clinicians at the right time. We need to have good, accurate predictions and then feed them into well-designed, well-implemented decision support in the workflow. »
This reality was underscored by Suchi Saria, Ph.D., John C. Malone Associate Professor and Director of the Machine Learning and Healthcare Lab at Johns Hopkins University in Baltimore, on March 14 at HIMSS22 in Orlando, in her closing speech. in the Machine Learning & AI for Healthcare Forum. Saria told her audience that “89% of providers have adopted some sort of sepsis tool.” But as I reported from his speech, “[C]An in-depth review of sepsis tool implementations found that when his Bayesian team looked closely at the success levels of sepsis alert algorithms, they found that actual rates of improvement in intervention were turned out to be much more modest than they first appeared. glance. In fact, she says, “I saw an incorrect review.” People measured sepsis for mortality, then deployed the tool, then used the billing code data and assessed. But it looks like you’ve improved mortality, but there’s a dilution effect. In other words, it turns out that clinicians and clinical informatics managers must necessarily test and recalibrate all algorithms developed elsewhere, in their own organizations, since, as Patterson told me, clinicians document their own organization’s electronic health records on an individual basis.
That said, there is no doubt that AI is becoming the topic of the day in patient care organizations nationwide, both for clinical diagnostic purposes, for clinician workflow purposes, for operational improvement purposes and for revenue cycle management purposes, among others. Indeed, as I noted in my coverage of RSNA22, the annual conference of the Radiological Society of North America, held in late November at the McCormick Place Convention Center in Chicago, AI was the absolute talk of the conference.
And I reported on the plenary address on Monday, November 28, delivered by renowned physician and author Siddhartha Mukherjee, MD, DPhil. As I wrote on November 28, “The 4,188 seats of the Arie Crown Theater were nearly full when Dr. Mukherjee, introduced by RSNA President Bruce G. Haffty, MD, took the stage. Mukherjee spent most of the hour giving his presentation, with the final quarter of an hour spent in a fireside chat between the two doctors. Mukherjee, assistant professor at Columbia University and practicing oncologist at Columbia University Medical Center, is an oncologist and hematologist who has spent decades involved in research in the diagnosis and treatment of cancer. Received the 2011 Pulitzer Prize for General Nonfiction for his 2010 book The emperor of all diseases: a biography of cancer; last month he published his latest book, cell song. He spoke on the topic “A Glimpse into the Future of Biomedical Transformation”.
As I wrote, “At the start of his lecture, Mukherjee said, ‘To begin with, I want to talk about deep learning; it means deploying learning algorithms that mimic human learning. How do we learn? He asked. “Can machines learn like us? Can machines learn medicine? He referred to the April 3, 2017 article he published in the new yorker, at the request of the editors of this magazine. He pointedly noted that the title the editors gave the article was “AI VERSUS MD” – but quickly added that “Interestingly, that’s a fake title”. I don’t think “versus” is the right word. Much of what I’m about to tell you is about “with”, not “versus”. And he then referred to the philosopher Gilbert Ryle, who, “long before the birth of modern AI, made a distinction between ‘knowing that’ as opposed to ‘knowing how.’ “It’s knowing a series of facts; knowing how is bringing those facts together to produce learning. Mukherjee told his audience that he believed there would be huge advances in healthcare in the use of technology. of AI and machine learning in the coming years, breakthroughs that will transform both patient care and medical research.Certainly, he and all other experts and observers note, the level of activity around AI now speeds up dramatically.
By this, as editor David Raths wrote on Dec. 22, “New C-Suite titles – such as Director of Patient Experience and Director of Digital – pop up in healthcare from time to time. Now, some industry insiders are saying that healthcare systems appoint AI managers will be the trend in 2023. I got this idea while reading some 2023 predictions from Punit Soni, CEO of a company called Suki, which builds AI voice interfaces. The company’s Suki assistant is used in more than 100 health systems and clinics in three dozen specialties. One of its predictions for 2023 is that “the director of ‘AI will become a position in healthcare systems differentiated from CIOs and data/digital executives’. Until I read this,” Raths wrote, “I had never heard of any health creating this title, but a search quick revealed several. For example, the CIO office of the Department of Health and Human Services has an AI director named Greg Singleton. HHS initially created an AI office, appointing its first AI director, Oki Mek, in March 2021.”
Yet, experts warn, health officials must also prepare for the inevitable emergence of cybercrime in AI in radiology. As I reported on Nov. 29, during a Nov. 28 RSNA session, “[A] three experts from different disciplines examined the prospects of cybercrime impacting the field of radiology and the development of artificial intelligence in radiology and healthcare, and issued serious warnings about the potential for harm to patients and patient care organizations, as the adoption of AI grows in healthcare, with criminals potentially turning AI algorithms against patient care itself.
Having said that, there is no doubt that AI has a huge future in the field of health, in different areas of activity. If there were any bright spots among the 2022 challenges, one was definitely in the area of AI and machine learning.