Virtual Reality-Based Infrared Pupillometry (VIP) for Long-COVID.

Abstract

To evaluate the use of virtual reality-based infrared pupillometry (VIP) to detect individuals with long coronavirus disease (LCVD). Prospective, case-control cross-sectional study. Participants 20 to 60 years of age were recruited from a community eye screening program. Pupillary light responses (PLRs) were recorded in response to 3 intensities of light stimuli (L6, L7, and L8) using a virtual reality head-mount display (VRHMD). Nine PLR waveform features for each stimulus were extracted by 2 masked observers and were analyzed statistically. We also used trained, validated, and tested (6:3:1) methods on the entire PLR waveform by machine learning models for 2-class and 3-class classification into LCVD, post-COVID (PCVD), or control groups. Accuracies and areas under the receiver operating characteristic curve (AUCs) of individual or a combination of PLR features and machine learning models analyzing PLR features or whole pupillometric waveform. Pupillary light responses from a total of 185 participants, including 112 in the LCVD group, 44 in the PCVD group, and 29 in the age- and sex-matched control group were analyzed. Models examined the independent effects of age and sex. Constriction time (CT) after the brightest stimulus (L8) is associated significantly with LCVD status (false discovery rate [FDR] < 0.001, 2-way analysis of variance; FDR < 0.05, multinominal logistic regression). The overall accuracy and AUC of CT after L8 alone in differentiating the LCVD group from the control or PCVD group were 0.7808 and 0.8711, respectively, and 0.8654 and 0.8140, respectively. Using cross-validated backward stepwise variable selection, CT after L8, CT after L6, and constriction velocity (CV) after L6 were most useful to detect LCVD, whereas CV after L8 was most useful for distinguishing the PCVD group from other groups. The accuracy and AUC of selected features were 0.8000 and 0.9000 (control vs. LCVD groups) and 0.9062 and 0.9710 (PCVD vs. LCVD groups), respectively, better than when all 27 pupillometric features were combined. A long short-term memory model analyzing whole pupillometric waveform achieved the highest accuracy and AUC at 0.9375 and 1.000 in differentiating the LCVD from PCVD group and a lower accuracy of 0.7838 for 3-class classification (LCVD, PCVD, and control group). We report specific pupillometric signatures in differentiating LCVD from PCVD or control groups using a VRHMD. Combining statistical methods to identify specific pupillometric features and machine learning algorithms to analyze the whole pupillometric waveform further enhanced the performance of VIP as a nonintrusive, low-cost, portable, and objective method to detect LCVD. Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.