A Retrospective Study for Development of an Artificial Intelligence-Based Quantitative Facial Palsy Assessment Tool

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Objectives: This study aims to evaluate the clinical utility of an artificial intelligence (AI)-based mobile application for the automated, quantitative analysis of facial palsy. It also aims to establish optimal cutoff values that distinguish normal from impaired performance, thereby quantifying the severity and recovery of facial palsy.Methods: The electronic medical record data of 100 patients (199 photos) were used retrospectively. It covered 10 facial expressions, which were analyzed by a mobile application (Smilelab) based on the MediaPipe algorithm. 478 3D landmarks were extracted to measure mobility, symmetry, and mouth symmetry. Three clinicians independently evaluated the photos using the Yanagihara grading system. Statistical analyses, including the Kruskal-Wallis test and Receiver Operating Characteristic (ROC) curve analysis based on Youden’s index, were performed.Results: Significant associations were found between the clinicians’ evaluations and the application’s measurements. The mobility and symmetry indicators showed significant differences in forehead wrinkling, light eye closure, nose wrinkling, whistling, and grinning (p < 0.05). In mouth symmetry, correlation was found in oral movements (p < 0.001). ROC analysis showed discrimination; lower lip depression center deviation achieved an outstanding AUC of 0.893 (95% CI: 0.848–0.938; cutoff: 0.026).Conclusions: The AI-based mobile application successfully quantified facial palsy. The calculated mobility, symmetry, and mouth symmetry serve as useful quantitative markers. Furthermore, the optimal cutoff values can offer an objective, data-driven basis for assessing movement-specific impairment and for customizing rehabilitation, including acupoint selection.