Diagnostic accuracy of MRI-derived imaging biomarkers and machine learning-based classification in distal and myofibrillar myopathies.

Abstract

Objectives: To characterise and validate disease-specific muscle imaging patterns in distal and myofibrillar myopathies, quantify their diagnostic performance, and develop an imaging-based classification framework incorporating machine learning.

Materials and methods: Muscle MRI and CT scans from a multicentre national cohort of patients with genetically defined distal and myofibrillar myopathies were retrospectively analysed. The pattern of muscle involvement for each disease entity was described, and its sensitivity and specificity calculated. Supervised machine learning algorithms were trained using muscle scores and imaging-derived features, and their diagnostic performance was compared using balanced accuracy.

Results: A total of 165 patients (mean age 48 ± 16 years; 87 men), including 138 molecularly confirmed cases spanning 19 disease genes, were analysed. Overall, 6350 muscles were systematically evaluated. Key findings included early biceps femoris short head involvement in GNE myopathy, "fat pockets" in myotilinopathy, and early flexor hallucis longus involvement in desminopathy. All imaging patterns showed high sensitivity and specificity (from 75 to 100%). Among the tested machine learning models, XGBoost achieved the highest performance, correctly distinguishing among 11 genetic disease entities with a balanced accuracy of 69%.

Conclusion: Distal and myofibrillar myopathies demonstrate reproducible disease-specific imaging signatures with high diagnostic accuracy. Integration of these imaging biomarkers into a machine learning framework enables systematic prioritisation of likely genetic diagnoses and may facilitate diagnostic work-up in these heterogeneous inherited muscle disorders.

Key points: Question How accurately do muscle MRI patterns distinguish distal and myofibrillar myopathies, and how do imaging features perform in an updated MRI-based diagnostic algorithm? Findings Disease-specific muscle MRI patterns showed high diagnostic accuracy, while XGBoost achieved the highest balanced accuracy among machine learning models for imaging-based disease classification. Clinical relevance Refined muscle MRI patterns provide diagnostically useful imaging biomarkers in distal and myofibrillar myopathies. Their integration into machine learning-based classification models may facilitate precision diagnostics and guide targeted genetic testing.