Deep Learning-Driven Glaucoma Medication Bottle Recognition: A Multilingual Clinical Validation Study in Patients with Impaired Vision

Abstract

Objective: To clinically validate a convolutional neural network (CNN)-based Android smartphone app in the identification of topical glaucoma medications for patients with glaucoma and impaired vision.

Design: Nonrandomized prospective crossover study.

Participants: The study population included a total of 20 non-English-speaking (11 Spanish and 9 Vietnamese) and 21 English-speaking patients who presented to an academic glaucoma clinic from December 2023 through September 2024. Patients with poor vision were selected on the basis of visual acuity (VA) of 20/70 or worse in 1 eye as per the California Department of Motor Vehicles' driver's license screening standard.

Intervention: Enrolled subjects participated in a medication identification activity in which they identified a set of 6 topical glaucoma medications presented in a randomized order. Subjects first identified half of the medications without the CNN-based app. They then identified the remaining half of the medications with the app. Responses to a standardized ease-of-use survey were collected before and after using the app.

Main outcome measures: Primary quantitative outcomes from the medication identification activity were accuracy and time. Primary qualitative outcomes from the ease-of-use survey were subjective ratings of ease of smartphone app use.

Results: The CNN-based mobile app achieved a mean average precision of 98.8% and recall of 97.2%. Identification accuracy significantly improved from 27.6% without the app to 99.2% with the app across all participants, with no significant change in identification time. This observed improvement in accuracy was similar among non-English-speaking (71.6%) and English-speaking (71.4%) participants. The odds ratio (OR) for identification accuracy with the app was 319.353 (P < 0.001), with substantial improvement in both non-English-speaking (OR = 162.779, P < 0.001) and English-speaking (no applicable OR given 100% identification accuracy) participants. Survey data indicated that 81% of English speakers and 30% of non-English speakers found the app "very easy" to use, with the overall ease of use strongly associating with improved accuracy.

Conclusions: The CNN-based mobile app significantly improves medication identification accuracy in patients with glaucomatous vision loss without increasing the time to identification. This tool has the potential to enhance adherence in both English- and non-English-speaking populations and offers a practical adjunct to daily medication management for patients with glaucoma and low VA.

Financial disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Keywords: Artificial intelligence; Clinical validation; Convolutional neural network; Glaucoma medication; Medication compliance.

Figure 1

Demonstrated use of UCEye convolutional neural network-based application.

Figure 2

Mean average precision over epochs, i.e. a single iteration in the training process. mAP = mean average precision.

Figure 3

Box (A), class (B), and object (C) losses (over epochs).