BlinkNet-Q: A Deep Learning–Based Software for Real-Time Blink Rate Quantification and Incomplete-Blink Pattern Recognition

Résumé fourni par la source

Purpose: To develop and validate a dual deep learning model (DLM)-based software (BlinkNet-Q) for near real-time quantification of blink rate and incomplete blinks using standard red-green-blue video. Methods: BlinkNet-Q integrated two DLMs: DLM-1 classified open-eye versus complete-blink frames, and DLM-2 differentiated complete from incomplete blinks. This proof-of-concept technical validation study included separate model development and independent testing cohorts. The development cohort comprised 15 participants (30 eyes), yielding 2953 annotated frames for training and internal validation, with participant-level data splitting. The testing cohort included 39 participants (39 eyes) and was used for final evaluation of the BlinkNet-Q. Frame-wise analysis with computer vision-based eye localization enabled automated blink quantification. Performance was assessed using classification metrics, regression-based metrics, intraclass correlation coefficient (ICC), and Bland-Altman analysis. Results: DLM-1 achieved an area under the curve (AUC) of 0.989 (95% confidence interval [CI], 0.932-0.993), and DLM-2 reached an AUC of 0.884 (95% CI, 0.854-0.963). The BlinkNet-Q showed strong agreement with manual counting for blink rate (ICC = 0.973; R² = 0.948; mean absolute error [MAE] = 1.262 blinks/min) and incomplete-blink ratio (ICC = 0.937; R² = 0.876; MAE = 0.059), with small mean biases under standardized recording conditions. Conclusions: BlinkNet-Q demonstrated the technical feasibility of automated blink rate and incomplete-blink quantification under standardized recording conditions, supporting its potential as a software prototype for blink assessment. Further external validation in diverse clinical populations is needed before clinical implementation. Translational Relevance: BlinkNet-Q may support future development of accessible, camera-based blink assessment tools for ocular surface evaluation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.