Clinical Validation of a Handheld Deep Learning Tool for Identification of Glaucoma Medications

Abstract

Abstract Purpose: To validate a convolutional neural network (CNN)-based smartphone application for the identification of glaucoma eye drop medications in patients with normal and impaired vision. Methods: Sixty-eight patients with visual acuity (VA) of 20/70 or worse in at least one eye who presented to an academic glaucoma clinic from January 2021 through August 2022 were included. Non-English-speaking patients were excluded. Enrolled subjects participated in an activity in which they identified a predetermined and preordered set of six topical glaucoma medications, first without the CNN and then with the CNN for a total of six sequential measurements per subject. Responses to a standardized survey were collected during and after the activity. Primary quantitative outcomes were medication identification accuracy and time. Primary qualitative outcomes were subjective ratings of ease of smartphone application use. Results: Topical glaucoma medication identification accuracy (OR = 12.005, P < 0.001) and time (OR = 0.007, P < 0.001) both independently improved with CNN use. CNN use significantly improved medication accuracy in patients with glaucoma (OR = 4.771, P = 0.036) or VA ≤ 20/70 in at least one eye (OR = 4.463, P = 0.013) and medication identification time in patients with glaucoma (OR = 0.065, P = 0.017). CNN use had a significant positive association with subject-reported ease of medication identification (X2(1) = 66.117, P < 0.001). Conclusion: Our CNN-based smartphone application is efficacious at improving glaucoma eye drop identification accuracy and time. This tool can be used in the outpatient setting to avert preventable vision loss by improving medication adherence in patients with glaucoma.

Keywords

• convolutional neural network
• deep learning
• glaucoma
• medication tools