Deep learning model for needle electromyography electrodiagnosis in comparison with physician assessment: A retrospective study

Research Objectives: To investigate whether the deep-learning approach can better classify EMG waveforms into neuropathy, myopathy, than normal physicians’ electrodiagnosis results.

Design: A one-dimensional convolutional neural network model was used as the deep-learning model. The deep-learning model and six physicians classified and electro-diagnosed the EMG waveforms as myopathic, neuropathic, or normal.

Setting: The nEMG signal data comprised those of 58 individuals who visited Seoul National University Hospital between June 2015 and July 2020. Each individual was classified as having peripheral neuropathy, myopathy, or as normal based on the final diagnosis.

Participants: There were 20 participants without any neuromuscular disease; 19 with neuropathy, including radiculopathy, motor axonal polyneuropathy, and motor neuron disease; and 19 with myopathy, including muscular dystrophy and inflammatory myopathy. The number of nEMG data points used for analysis were 124, 161, and 97 for participants with myopathy, neuropathy, and normal states, respectively.

Interventions: No intervention

Main Outcome Measures: The performance of the deep-learning model and the physicians were evaluated using the following metrics: accuracy, F1 score, area under the receiver operating characteristic (ROC) curve, positive predictive value (PPV; precision), sensitivity (recall), and specificity.

Results: The accuracy, sensitivity, specificity, positive predictive value and F1 score of the deep learning model were 0.720, 0.715, 0.858, 0.726, and 0.715, respectively, and the physicians’ mean scores were 0.537, 0.527, 0.770, 0.582, and 0.511, respectively. The performance of the deep-learning model for predicting myopathy, neuropathy, and normal state was also evaluated using the area under the receiver operating characteristic curve, and the results were 0.874 (95% confidence interval [CI] 0.858–0.889), 0.781 (95% CI 0.723–0.839), and 0.847 (95% CI 0.836–0.858), respectively.

Conclusions: This study demonstrated that deep-learning could contribute to interpreting the EMG of patients with neuromuscular disease on behalf of physicians and assist physicians’ decision-making regarding diagnosing patients with neuromuscular disease. Large prospective cohort studies with more diverse neuromuscular diseases will further improve the performance of deep-learning-based EMG interpretation in the future.

Author(s) Disclosures: The authors declare no competing financial interests.