Using AI to Improve Ophthalmology Triage
By Jean Shaw
Selected by Emily Y. Chew, MD
Journal Highlights
Ophthalmology Science, March 2023
Download PDF
Tanya et al. evaluated a cloud-based triage system for on-call ophthalmology consults. They found that their clinical decision support system (CDSS) is accurate, standardizes information collection, and is superior to paper-based consults.
For this prospective comparative cohort study, the researchers used current practice guidelines and expert opinion to develop an artificial intelligence (AI)–based decision tree for referrals. Their CDSS collected specific information on patients’ ophthalmic symptoms, produced a provisional diagnosis, and sent an electronic referral to two community-based ophthalmology clinics in Canada.
Data were collected between November 2020 and December 2021. Ten categories of new-onset symptoms were developed, along with 10 categories covering previous ophthalmic diagnoses and seven levels of referral urgency. Primary outcome measures included the category, diagnosis, and level of urgency for the referring provider, the CDSS, and the on-call clinician. The CDSS provided multiple options to communicate directly with the on-call ophthalmologist if needed, and the clinician’s final diagnosis and assessment of urgency, based on a complete ophthalmic examination, served as the ground truth.
During the study period, 96 referrals that represented 59 unique diagnoses were processed. The most frequent referring categories were vitreoretinal pathology (37.5%), traumatic pathology (16.7%), and unspecified (14.6%).
Results showed that the CDSS performed as well as referring providers did in determining a disease category and better than they did in determining a diagnosis. With regard to assessing urgency, the CDSS correctly assigned 66 cases (68.8%), matching the ophthalmologist’s assessment. Among the other 30 cases, the system assigned 15 a lower level of urgency and eight a higher level of urgency than did the ophthalmologist. The remaining seven were of unspecified urgency and represented instances of “tree failure,” as information was unavailable for the initial data collection.
Future research directions include broadening the scope of clinical scenarios by increasing the number of decision trees, the authors wrote, with the goal of improving the system’s diagnostic accuracy.
The original article can be found here.