Objective:

We assessed the feasibility, usefulness, and generalizability of the Eyer2 and its integrated AI tool for ophthalmology/neuro-ophthalmology/neurology inpatient and emergency department (ED) consultations.

Background:

Bedside ocular funduscopic examination is challenging, both for trainees and staffing attendings who must confirm remote findings. Eyer2 is a handheld non-mydriatic, true-color 45/55-degree field ocular fundus camera, with electronic-health-record compatibility, automated upload of images to the Eyer-Cloud, and artificial intelligence (AI) capabilities. Its reliability when used by non-eye-care providers at bedside is unclear.

Design/Methods:

We deployed the Eyer2 by trained students/medicine/neurology/ICU/residents/fellows/attendings in varied clinical settings including outpatient clinics, EDs, and inpatient units. Images were acquired in both eyes without pupillary dilation, and automatically uploaded to the Eyer-Cloud, with remote interpretation by ophthalmologists. Image quality was graded as good/adequate/poor. Images were assessed for pathology and AI interpretation accuracy.

Results:

Over 20 weeks, 482 photographs were obtained on 242 patients in the ED (19%), hospital floors (13%), ICU (19%), clinic (43%), elsewhere (5%). 230 images were graded as good (48%), 181 adequate (38%), 71 poor (15%). The most common reason for poor image quality was miosis (68%). 183 images were abnormal (38%), including papilledema (38); diabetic/hypertensive retinopathy (35). The AI tool was activated for 66% of images (AI tool does not activate if poor image quality) with sensitivity 76% and specificity 90% for detecting ocular fundus abnormalities.

Conclusions:

Most Eyer2 fundus photographs provided useful data for remote interpretation, allowing accurate and rapid identification of optic nerve/macula pathology, immediate backup for trainees, and documentation of findings. While image quality varied across user-type and location, the camera generated useful images across multiple different operators and clinical settings. Challenges included a steep learning curve in its use and pupillary miosis, especially in ICUs. The AI interpretation was helpful mostly in good quality photographs by alerting non-ophthalmology-trained providers of potential optic nerve or retinal pathology.