What Will AI Do To The Doctors?
What happens when AI tools begin to integrate in clinical practice? Does clinical practice & patient care improve? Do patients benefit? What happens to the doctors themselves, & their clinical skills?
As AI tools and technologies begin to integrate into medical practice, these are poignant, but unexplored questions. A new paper published in the Lancet Gastroenterology provides some preliminary answers.
As the paper reports, researchers conducted a retrospective analysis of 1443 colonoscopies done at four Polish Endoscopy centers. In 2021, an AI tool to help doctors localise adenomas during endoscopy was introduced as part of the ACCEPT (Artificial Intelligence in Colonoscopy for Cancer Prevention Trial) trial. The goal of this analysis was to compare the quality of endoscopies 3 months before and 3 months after the AI tool was introduced.
The analysis found that once doctors got used to relying on AI assisted tools to detect polyps, endoscopists got about 20% worse at detecting adenomas without the assistance of AI than they were before the AI tool was introduced. This drop is substantive from a clinical point of view. What is even more startling, is that the drop was noted within 3 months.
In the ACCEPT trial, researchers had set out to do a randomized controlled study of adenoma detection rates with and without the assistance of an AI tool to localise a possible adenoma. As part of the trial, patients were alternately enrolled for an endoscopy with an AI tool or without an AI tool depending on the day the endoscopy was done.
The ACCEPT trial was designed to compare adenoma detection rates with an AI tool versus without an AI tool. When researchers noticed what appeared to be an over reliance among endoscopists on the AI tool to detect adenomas, they set out to retrospectively compare adenoma detection rates before the AI tool was introduced to detection rates after the tools were introduced. On days the AI tools were not being used after the trial began (the control arm of the trial), researchers found a 20% decrease in the adenoma detection rate compared to the rates before the AI tools were introduced, as we noted above.
Although this analysis makes use of a natural experiment of sorts, there are still reasons to have reservations about the study design and its ability to signify causality. That is a discussion for another time. The question to ask for this discussion is, what if there is a grain of truth in the findings of this study. These results do have face validity, as illustrated by our tendency to overly rely on everyday technological tools like mobile phones.
AI tools could indeed get better than doctors at predefined clinical tasks, not just because they process and synthesise information better, but because doctors tend to get sloppier along the way. Corroboration of these findings will require multiple other more carefully designed studies–preferably prospective cohort or randomized controlled trials–studying several kinds of AI tools deployed in several clinical scenarios. If these findings do prove to be true, we will have to think deeper and harder about the consequences of the integration of AI tools in clinical practice and medical education.
It's been almost a decade since Richard and Daniel Susskind declared that AI will soon replace doctors. That reality has not come to pass yet, although AI technologies are making startling progress in information synthesis and medical decision making. The technical potential for AI to replace doctors may be within the realm of possibility—although, as people are beginning to discover, replacing doctors is not a mere technical challenge. But help may be coming from an unassuming corner—doctors themselves!
Kiran Raj Pandey is a physician and a health services & systems researcher. Learn more about him and his work at kiranrajpandey.com.
 | A guest post by
|