A Wearable Approach to RAPD Detection: Using a Virtual Reality Headset for Pupillometry

Abstract

Purpose : To describe and evaluate a portable diagnostic tool to determine relative afferent pupillary defects (RAPD) using parameters derived from a virtual swinging light test.

Methods : We used Heru (Heru Inc., FL) software on a virtual reality headset device (Pico N3P Eye, ByteDance, China) to collect pupillary response data to light. The data was collected on 73 subjects (36 females, 37 males; aged 20-89). The normal cohort had 40 subjects without any pathology impacting their pupillary responses. The RAPD cohort consisted of 20 subjects with RAPD in the left eye (OS) and 13 with RAPD in the right eye (OD). The sequence started by normalizing both pupil sizes by showing a dim screen to both eyes. Once normalized, a virtual swinging light test commenced, in which, one eye was lit with a maximum intensity display, followed by a period of rest. This was repeated for the fellow eye. The cycle occurred at least two times to minimize outlying responses. During the test, pupil size data was recorded. From the data, we derived 28 pupillary response parameters (14 for each eye): direct/consensual latency, end of constriction, constriction/dilation amplitudes, and constriction/dilation velocities; and the mean low and high. We used all 28 parameters to determine the existence and cardinality of RAPD using a histogram gradient boosting classifier. Outcomes of the classifier predictions: RAPD OS, No RAPD, or RAPD OD, were compared to the ground truth diagnoses to determine the classification efficiency.

Results : Data analysis of the parameters showed a statistically significant difference between the No RAPD, RAPD OS, and RAPD OD groups (p<0.0001, MANOVA test, 95% confidence level) indicating that the existence and cardinality of RAPD influences the parameters. The outcomes of the histogram gradient classifier were found to be significant, as shown in the confusion matrix in Figure 1. Based on this matrix, we calculated the model’s accuracy, sensitivity, and specificity. They were 0.97, 0.94, and 1.00, respectively.

Conclusions : Our novel diagnostic model utilized a portable solution that may allow clinicians to screen for RAPD. The statistically significant difference between the parameters shows that this test can detect the existence and cardinality of RAPD with relatively high accuracy. Future studies would focus on improving the model to be able to quantify the severity of the RAPD in units of neutral-density filter.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.