The Clinician or AI – Who Will Diagnose?
Do we effectively diagnose and manage keratoconus and corneal ulcers? Are we being compassionate?
Cheryl Guttman Krader
Published: Monday, May 1, 2023
Cheryl Guttman Krader Reports
Debate seeks common ground on the value of a combined approach to corneal disease diagnosis.
Will artificial intelligence-based systems replace the experienced clinician in diagnosing keratoconus and infectious keratitis within the next 10 years? Cornea specialists debated this question at the American Academy of Ophthalmology annual meeting.
To support her conclusion that technology-based, clinician- support software will replace today’s clinicians alone, Dr Maria A Woodward described the shortcomings of today’s clinicians in delivering quality care and introduced opportunities through AI.
“A clinician is someone who triages, diagnoses, and manages disease and listens to, communicates with, and cares for the sick. But do we effectively diagnose and manage keratoconus and corneal ulcers? Are we being compassionate?” she asked.
Citing a study her group conducted, Dr Woodward said almost 50% of clinicians lacked confidence in their ability to differentiate organism types when diagnosing microbial keratitis.
“In contrast, AI-based algorithms could distinguish between organism categories and will only get better over the next decade.”
Similarly, she questioned whether clinicians effectively diagnose and manage keratoconus and suggested advantages for AI-based strategies.
“Do we know how frequently to see patients, which are at risk for progression, should get cross-linking, and should go on to transplantation? Now there is a lot of good evidence that AI algorithms can help distinguish between keratoconus subtypes and help us predict which patients should go on to get cross-linking or a transplant.”
These are the situations where Dr Woodward proposed AI could enhance clinician ability to provide compassionate quality care.
“How much time do we spend talking to our patients and interacting with them as human beings? There is plenty of data showing how we could be delivering better care, reducing our medical error rate, and reducing our burnout [with] compassion,” she concluded. “By leveraging technology for improving how we triage, diagnose, and manage disease, we can connect more personally with our patients. Technology is not the enemy—it is up to us to drive technology-aided care.”
Not the sole solution
Dr Stephen D Klyce assured his colleagues AI would not replace them in diagnosing keratoconus and microbial keratitis. Those tasks, he said, require more than machine-based analysis of corneal images because the original AI-based keratoconus screening program provided only an interpretation of topographic images, not a diagnosis.
“AI can identify the corneal topography has the characteristics associated with clinical keratoconus but is insufficient by itself to make a diagnosis,” Dr Klyce asserted.
Newer systems—including scanning devices that give information on anterior and posterior corneal surface thickness, corneal biomechanics, and epithelial thinning patterns—provide additional information to help determine whether a cornea has keratoconus or if a patient is at risk for ectasia after laser refractive surgery. Even if clinicians had the physical space and financial resources to install all the devices used to acquire those measurements, the collective data still do not provide a clinical diagnosis.
“You need to know the patient’s history,” Dr Klyce said, explaining that a patient who appears to have forme fruste keratoconus based on AI-based interpretation of images could instead have contact lens-related corneal warpage.
Regarding AI-based diagnosis of microbial keratitis, Dr Klyce noted some models are effective in classifying images, though they are not completely accurate.
Citing a study showing that an AI-based algorithm for diagnosing infectious keratitis based on clinical image classification had an 80% accuracy rate compared to just 49% for ophthalmologists, Dr Klyce focused on the algorithm’s error rate.
“What about the 20% of patients given the wrong therapy because of the algorithm’s mischaracterisation of the keratitis? The consequences could be disastrous.”
Describing other limitations of AI-based systems, Dr Klyce said they are only as good as the training data, whereas clinicians have access to many data sources beyond those available and used to train AI. In addition, AI algorithms need constant revision as technology advances. Even with expert consensus, there will be bias because independent groups have different ideas about the characteristics considered for an AI algorithm.
“AI will not replace the clinician but will continue to provide access to new tools to aid in clinical diagnosis,” Dr Klyce said in closing. “Clinicians and their patients will benefit from advances in the diagnostic capabilities afforded by the enormous strides in digital healthcare. The future will be AI and human intelligence working together.”
Maria A Woodward MD, MSc is Service Chief, Cornea, External Disease, and Refractive Surgery, University of Michigan, Ann Arbor, US. email@example.com
Stephen D Klyce PhD, FARVO is an adjunct professor of ophthalmology at the Icahn School of Medicine at Mount Sinai, New York, US. firstname.lastname@example.org