Evaluating pediatric peripheral neuromuscular disorders using deep neural networks on electrodiagnostic data.

Objective: This study aims to assess application of deep neural networks to predict neuromuscular disorders in patients based on electrodiagnostic data and compare with clinical assessment. Methods: Patients evaluated with electrodiagnostic tests in intensive care over a 10-year period were included in this study. The data set contained both electrodiagnostic and clinical information. Based on the final diagnosis, patients were classified into six groups: non-primary neuromuscular disorders, neuropathy, motor-neuronopathy, myopathy, neuromuscular-junction disorders, and critical-illness neuromyopathy. The neural network was trained on the data. Results: The data set was small, allowing training of the neural network on a standard laptop. The validation results were promising, with an accuracy of 0.92, an ROC-AUC of 0.99, and a precision recall AUC of 0.97. The confusion and positive predictive value matrix demonstrated high performance, with diagonal values exceeding 0.82. Conclusion: This study demonstrates the efficacy of neural networks in predicting neuromuscular disorders using electrodiagnostic tests. The performance of the model was comparable to clinical assessment. These findings suggest that with more extensive datasets, neural networks can provide reliable estimates of neuromuscular diagnoses. Significance: Incorporating neural networks into diagnostic workflows could enhance decision-making, especially in scenarios requiring reassessment or complementary investigations.