Development and Validation of a Deep Learning‐Based Facial Weakness Score for Objective Assessment in Facioscapulohumeral Muscular Dystrophy

INTRODUCTION/AIMS: Facioscapulohumeral muscular dystrophy (FSHD) is a muscle disease that leads, among other manifestations, to facial weakness. This weakness can severely impact communication and quality of life, yet it remains under-researched with limited objective clinical measures. Current manual scoring methods are subjective and exhibit suboptimal inter-observer agreement. This study aimed to develop and validate a deep learning-based facial weakness score (DLFWS) as an objective clinical outcome measure for facial weakness in FSHD. METHODS: One hundred and twenty-two genetically confirmed FSHD patients and 56 controls were recruited. Sixty-four patients had a 5-year follow-up visit. Video recordings of participants performing seven facial exercises, each repeated three times, were analyzed using a convolutional neural network. The deep learning-based facial weakness score (DLFWS) was trained by comparing start and end frames from these exercises with the manual facial weakness scores (MFWS) assigned by three experienced observers. Pearson correlation coefficients and intraclass correlation coefficients (ICC) were used to evaluate the DLFWS's performance and reliability. RESULTS: The DLFWS showed a strong correlation with the MFWS across exercises, with a mean Pearson correlation of 0.79 and a maximum of 0.85 for individual exercises. The DLFWS demonstrated excellent test-retest reliability, with an ICC of 0.90. Over a 5-year follow-up period, no significant progression of facial weakness was detected. DISCUSSION: The DLFWS provides a reliable and objective assessment of facial weakness in FSHD patients. This automated tool holds potential for widespread clinical and research applications, enabling standardized assessment of facial weakness in FSHD.