A Combined Deep Learning Approach to Screen Patients for Neuromuscular Pathology

Neuromuscular diseases (NMD), comprising over 600 different conditions, severely impact nerve and/or muscle function and lead to significant morbidity. Ultrasound is a non-invasive tool that is gaining acceptance for diagnosing NMD. In clinical practice, muscle ultrasound can be evaluated quantitatively or visually using an ordinal four-point grading score (Heckmatt score). Its current application is limited by time investment in manual analysis and lack of result transferability to other centers. Here, we present a single-center multi-modal deep learning framework using intermediate data fusion that improved the speed and diagnostic performance of muscle ultrasound. Our approach used neural networks enriched with patient-specific data of body mass index and age to predict neuromuscular pathology with an area under the precision-recall curve of 0.87 on a test set of 320 patients (220 with NMD and 100 in whom the diagnosis was refuted). SHAP analysis showed that adding BMI and age did not affect the model's performance. By leveraging Heckmatt scores from ultrasound images of six key muscles, our model efficiently and effectively identified the presence or absence of a neuromuscular disease. This approach may enhance the clinical utility of ultrasound by facilitating a more efficient diagnostic process for neuromuscular pathology, helping to guide the subsequent workup toward a definitive diagnosis.