To evaluate the value of artificial intelligence (AI) for recommendation of pupil dilation test using medical interview and basic ophthalmologic examinations.
Retrospective, cross-sectional study.
Medical records of 56,811 patients who visited our outpatient clinic for the first time between 2017 and 2020 were included in the training dataset. Patients who visited the clinic in 2021 were included in the test dataset. Among these, 3,885 asymptomatic patients, including eye check-up patients, were initially included in test dataset I. Subsequently, 14,199 symptomatic patients who visited the clinic in 2021 were included in test dataset II.
All patients underwent a medical interview and basic ophthalmologic examinations such as uncorrected distance visual acuity, corrected distance visual acuity, non-contact tonometry, auto-keratometry, slit-lamp examination, dilated pupil test, and fundus examination. A clinically significant lesion in the lens, vitreous, and fundus was defined by subspecialists, and the need for a pupil dilation test was determined when the participants had one or more clinically significant lesions in any eye. Input variables of AI consisted of a medical interview and basic ophthalmologic examinations, and the AI was evaluated with predictive performance for the need of a pupil dilation test.
Accuracy, sensitivity, specificity, and positive predictive value.
Clinically significant lesions were present in 26.5 and 59.1% of patients in test datasets I and II, respectively. In test dataset I, the model performances were as follows: accuracy, 0.908 (95% confidence interval (CI): 0.880–0.936); sensitivity, 0.757 (95% CI: 0.713–0.801); specificity, 0.962 (95% CI: 0.947–0.977); positive predictive value, 0.878 (95% CI: 0.834–0.922); and F1 score, 0.813. In test dataset II, the model had an accuracy of 0.949 (95% CI: 0.934–0.964), a sensitivity of 0.942 (95% CI: 0.928–956), a specificity of 0.960 (95% CI: 0.927–0.993), a positive predictive value of 0.971 (95% CI: 0.957–0.985), and a F1 score of 0.956.
The AI model performing a medical interview and basic ophthalmologic examinations to determine the need for a pupil dilation test had good sensitivity and specificity for symptomatic patients, although there was a limitation in identifying asymptomatic patients.