Application of a deep learning system to detect papilledema on nonmydriatic ocular fundus photographs in an emergency department

PURPOSE: The Fundus photography vs. Ophthalmoscopy Trial Outcomes in the Emergency Department (FOTO-ED) studies showed that ED providers poorly recognized funduscopic findings in ED patients. We tested a modified version of the Brain and Optic Nerve Study Artificial Intelligence (BONSAI) deep learning system on nonmydriatic fundus photographs from the FOTO-ED studies to determine if the DLS could have improved the detection of papilledema had it been available to ED providers as a real-time diagnostic aid.

DESIGN: Retrospective secondary analysis of a cohort of patients included in the FOTO-ED studies.

METHODS: The testing dataset included 1608 photographs obtained in 828 patients from the FOTO-ED studies. Photographs were reclassified according to the optic disc classification system used by the deep learning system ["normal optic discs"; "papilledema"; "other optic disc abnormalities"]. The system's performance was evaluated by calculating the AUC, sensitivity and specificity using a one-vs-rest strategy, with reference to expert neuro-ophthalmologists.

RESULTS: The BONSAI-deep learning system successfully distinguished normal from abnormal optic discs [(AUC 0.92 (95%CI, 0.90-0.93); sensitivity 75.6% (73.7%-77.5%) and specificity 89.6% (86.3%-92.8%)], and papilledema from normal and others [(AUC 0.97 (0.95-0.99); sensitivity 84.0% (75.0%-92.6%) and specificity 98.9% (98.5%-99.4%)]. Six patients with missed papilledema in one eye were correctly identified by the deep learning system as having papilledema in the other eye.

CONCLUSIONS: The BONSAI deep learning system was able to reliably identify papilledema and normal optic discs on non-mydriatic photographs obtained in the FOTO-ED studies. Our deep learning system has excellent potential as a diagnostic aid in EDs and non-ophthalmology clinics equipped with nonmydriatic fundus cameras.