Classification of neuromuscular diseases with facial recognition based on deep learning

Objective: Neuromuscular disorders (NMDs), especially when affecting facial muscles, present challenging and specialized clinical presentations. The aim of this study is to develop a deep learning (DL)-based model to distinguish patients diagnosed with Myasthenia Gravis (MG) and Hemifacial Spasm (HFS) from healthy controls and from each other, and to evaluate its potential in clinical decision support processes by comparing the model's performance with the diagnostic success of physicians with different levels of experience. Materials and Methods: This prospective cross-sectional study included a total of 50 participants: 20 MG patients, 15 HFS patients, and 15 healthy volunteers. Photographs of participants taken in six different poses (normal posture, smiling, squinting, and looking in three different directions) under standardized conditions were used as the dataset. A novel deep learning model utilizing the ResNet-18 backbone for feature extraction, with multi-view and relation-aware capabilities, was developed. The model's success was compared to the predictions of a group of physicians consisting of two general practitioners, two neurology residents, and one specialist neurologist using accuracy, F1-score, and Cohen's Kappa coefficient metrics. Findings: No statistically significant difference was found between the age and gender distributions of the groups (p > 0.05). In diagnostic performance analyses, the specialist neurologist achieved the highest success rate with 90.9% accuracy and a Cohen's Kappa value of 0.863. The developed deep learning model (DLM) demonstrated statistically significant success with 72.7% accuracy and a Kappa value of 0.566. The model's success was found to be higher than that of assistant physicians (63.6%) and general practitioners (45.5%–63.6%). The model's sensitivity was particularly high for MG diagnosis; it was observed that the relationship-aware architecture successfully encoded temporal changes between different facial expressions. Conclusion: This study demonstrates that deep learning-based facial recognition systems are a promising tool for non-invasive screening and differential diagnosis of neuromuscular diseases. The fact that the developed model performed better than non-specialist physicians indicates that the system could be a valuable clinical decision support tool, particularly in primary care settings and regions with limited access to specialist physicians. In the future, explainable artificial intelligence models supported by larger and more heterogeneous datasets will contribute significantly to the digital transformation of neuromuscular examination processes.