Automatic grading of patients with a unilateral facial paralysis based on the Sunnybrook Facial Grading System - A deep learning study based on a convolutional neural network

• Published in: American journal of otolaryngology 2023-05, Vol.44 (3), p.103810-103810, Article 103810
• Main Authors: ten Harkel, Timen C., de Jong, Guido, Marres, Henri A.M., Ingels, Koen J.A.O., Speksnijder, Caroline M., Maal, Thomas J.J.
• Format: Article
• Language: English
• Subjects: Bell Palsy; Convolutional neural network; Deep Learning; Face; Facial paralysis; Facial Paralysis - diagnosis; Humans; Machine learning; Medical imaging; Reproducibility of Results; Sunnybrook facial grading system
• ISSN: 0196-0709; 1532-818X
• DOI: 10.1016/j.amjoto.2023.103810

In order to assess the severity and the progression of a unilateral peripheral facial palsy the Sunnybrook Facial Grading System (SFGS) is a well-established grading system due to its clinical relevance, sensitivity, and robust measuring method. However, training is required in order to achieve a high inter-rater reliability. This study investigated the automated grading of facial palsy patients based on the SFGS using a convolutional neural network. A total of 116 patients with a unilateral peripheral facial palsy and 9 healthy subjects were recorded performing the Sunnybrook poses. A separate model was trained for each of the 13 elements of the SFGS and then used to calculate the Sunnybrook subscores and composite score. The performance of the automated grading system was compared to three clinicians experienced in the grading of a facial palsy. The inter-rater reliability of the convolutional neural network was within the range of human observers, with an average intra-class correlation coefficient of 0.87 for the composite Sunnybrook score, 0.45 for the resting symmetry subscore, 0.89 for the symmetry of voluntary movement subscore, and 0.77 for the synkinesis subscore. This study showed the potential of the automated SFGS to be implemented in a clinical setting. The automated grading system adhered to the original SFGS, which makes the implementation and interpretation of the automated grading more straightforward. The automated system can be implemented in numerous settings such as online consults in an e-Health environment, since the model used 2D images captured from a video recording. •The automated grading was performed with a similar accuracy as human observers.•The automated grading can be implemented in a wide variety of clinical settings.•By using an existing grading system, interpretation of the automated score is easy.