Diagnosing myopathies remains challenging due to their clinical heterogeneity and variability in muscle involvement patterns on muscle magnetic resonance imaging (MRI). Although artificial intelligence (AI) applied to muscle imaging shows promise, most models rely on a single modality and lack clinical integration and explainability. This study aims to evaluate whether integrating structured clinical data with MRI data improves the accuracy of AI in distinguishing between facioscapulohumeral dystrophy (FSHD1) and myositis, and to assess the explainability of AI decisions.

This retrospective study involved 40 patients with FSHD1 or myositis who underwent whole-body MRI and had their clinical data recorded in the French rare disease registry. MRI images were processed via a convolutional neural network. The clinical data were encoded into 36 variables. Monomodal (clinical) and hybrid multimodal (clinical and MRI) classifiers were trained using machine learning algorithms. Five-fold cross-validation was employed for each classifier.

The clinical classifier achieved an accuracy of 92.5%, heavily influenced by the "scapular winging" feature. When this feature was removed from the analysis, accuracy dropped to 80%. The multimodal classifier achieved an accuracy of 87.5%, outperforming the clinical classifier when "scapular winging" was excluded from the analysis. Explainability tools highlighted clinically relevant features (e.g. asymmetry, facial involvement and tibialis anterior atrophy), but also some areas of uncertain relevance.

Combining clinical and imaging data in a hybrid AI model could provide complementary information for the diagnosis of myopathies. Explainability tools provide insight into the features contributing to AI decisions, while allowing clinicians to critically review them.