Application of deep learning towards automated EMG wave classification in neuromonitoring for neck surgeries

Abstract Background Intraoperative neuromonitoring (IONM) - that is, recording of electromyography (EMG) signals - is routinely used during (para)thyroid surgery. Surgeons label selected signals to document nerve identity, body side and timepoint of stimulation (before/after resection), with a mislabelling rate of 20%. For the purpose of an automated error alert of mislabelled EMG signals, we developed a multi-task 1D convolutional neural network (CNN). Methods Raw IONM data were corrected using MIONQA software. Labelled EMG signals were extracted and metadata (surgery duration, timing; median EMG peak values of actual surgery) were added to each EMG wave. Between 150 and 280 extracted features were used to train, validate and test various CNNs. Results By repeated adjustments of the model architecture and the number of extracted features, an optimized 1D CNN was designed. After multiple runs with randomized training (13044 EMGs) and test data (3261 EMGs), the final optimized CNN achieved a mean accuracy of 95.7% ± 0.7 for correct identification of recurrent laryngeal, vagus and superior laryngeal nerves; 97.6% ± 0.7 for the correct prediction of the resected body side ; 97.6% ± 0.9 for correct identification of the stimulation timepoint (before vs. after resection); the ROC curve demonstrated an excellent AUC of 0.993. Conclusion The newly developed CNN enables accurate automated classification of EMG signals, facilitating the identification and correction of mislabelled IONM data. Optimised data quality is an essential prerequisite for artificial intelligence training, in future enabling neuromonitoring machines to alert the surgeon in the operating theatre of mislabelling.