Detection of the spontaneous blinking pattern of dry eye patients using the machine learning method

Abstract

Objective: To establish a method to record the spontaneous blink pattern with a machine learning model, and to clarify the spontaneous blink pattern in patients with dry eye. Methods: It was a cross-setional study.We selected 357 dry eye patients (102 males and 255 females), aged (46.2±13.3) years, who visited corneal specialist clinics of Beijing Tongren Eye Center in 2019, as the dry eye group. The control group enrolled 152 normal controls, including 32 males and 120 females, aged (48.1±13.9) years. All participants completed the Ocular Surface Disease Index questionnaire, blink video capture, lipid layer thickness measurement, tear break-up time measurement, corneal fluorescein staining, and Schirmer Ⅱ test. Based on the assembled model built using UNet image segmentation algorithm and ResNet image classification algorithm, single frames of the blink video were analyzed, and then the palpebral opening height of each frame was obtained in order to establish a spontaneous blink wave. Finally, the characteristics of spontaneous blinks in dry eye patients were analyzed based on different types of complete blinks (types A, B and C) and partial blinks (types Ⅰ, Ⅱ and Ⅲ). Independent sample t test and Wilcoxon rank-sum test were used to judge if there was significant difference between the dry eye group and the normal group. Results: The accuracy of the segmentation model and the classification model was 96.3% and 96.0%, respectively, and the consistency with the manual analysis was 97.9%. In dry eye patients, the number of blinks was 30 (18, 42)/min, which was higher than that in normal controls [20 (9, 46)/min] (U=18 132.50, P=0.002). The number of complete blinks in dry eye cases was significantly lower than that in normal controls [6 (3, 24)/min vs. 12 (3,33)/min; U=12 361.00, P=0.016], and the number of partial blinks was significantly higher than that in normal controls [15 (6, 27)/min vs. 3 (0, 10)/min; U=22 839.00, P<0.001]. In complete blinks, the proportion of type A blinks in dry eye patients was significantly higher than that in normal controls [53.7% (2 796/5 177) vs. 39.3% (633/1 698); χ²=101.83, P<0.001]; in partial blinks, the proportion of type Ⅱ blinks in dry eye patients was significantly higher than that in normal controls [36.0%(2 334/6 477) vs. 29.6%(126/426); χ²=6.99, P=0.007]. The average interblink interval of dry eye patients was 1.2 s, which was not significantly different from that of normal controls (1.1 s; U=15 230.00, P=0.093). The eyelid closed phase of dry eye patients was 0.8 s, which was significantly shorter than that of normal controls (1.3 s; U=16 291.50, P=0.006). There were no significant differences in eyelid closing phase, early opening phase and late opening phase between the two groups (all P>0.05). Conclusions: In dry eye patients, the number of partial blinks increased, the number of complete blinks decreased, and the duration of eyelid closed phase shortened significantly. The main blink patterns of dry eye patients included type Ⅱ partial blinks with a reduced closure amplitude and type A complete blinks with a shortened closure time.