Related Experiment Video
Updated: Jul 2, 2025
Intense Pulsed Light for the Treatment of Dry Eye Owing to Meibomian Gland Dysfunction
Published on: April 1, 2019
Comparison of deep learning-assisted blinking analysis system and Lipiview interferometer in dry eye patients: a
Yueping Ren1, Han Wen1, Furong Bai1
1National Clinical Research Center for Ocular Diseases, Eye Hospital, Wenzhou Medical University, Wenzhou, 325027, Zhejiang, China.
Background:
Abnormal blinking pattern is associated with ocular surface diseases. However, blink is difficult to analyze due to the rapid movement of eyelids. Deep learning machine (DLM) has been proposed as an optional tool for blinking analysis, but its clinical practicability still needs to be proven. Therefore, the study aims to compare the DLM-assisted Keratograph 5M (K5M) as a novel method with the currently available Lipiview in the clinic and assess whether blinking parameters can be applied in the diagnosis of dry eye disease (DED).
Methods:
Thirty-five DED participants and 35 normal subjects were recruited in this cross-sectional study. DED questionnaire and ocular surface signs were evaluated. Blinking parameters including number of blinks, number of incomplete blinking (IB), and IB rate were collected from the blinking videos recorded by the K5M and Lipiview. Blinking parameters were individually collected from the DLM analyzed K5M videos and Lipiview generated results. The agreement and consistency of blinking parameters were compared between the two devices. The association of blinking parameters to DED symptoms and signs were evaluated via heatmap.
Results:
In total, 140 eyes of 70 participants were included in this study. Lipiview presented a higher number of IB and IB rate than those from DLM-assisted K5M (P ≤ 0.006). DLM-assisted K5M captured significant differences in number of blinks, number of IB and IB rate between DED and normal subjects (P ≤ 0.035). In all three parameters, DLM-assisted K5M also showed a better consistency in repeated measurements than Lipiview with higher intraclass correlation coefficients (number of blinks: 0.841 versus 0.665; number of IB: 0.750 versus 0.564; IB rate: 0.633 versus 0.589). More correlations between blinking parameters and DED symptoms and signs were found by DLM-assisted K5M. Moreover, the receiver operating characteristic analysis showed the number of IB from K5M exhibiting the highest area under curve of 0.773.
Conclusions:
DLM-assisted K5M is a useful tool to analyze blinking videos and detect abnormal blinking patterns, especially in distinguishing DED patients from normal subjects. Large sample investigations are therefore warranted to assess its clinical utility before implementation.
