Application of deep learning towards automated electromyographic wave classification for neuromonitoring in thyroid
Thomas J Musholt1, Petra B Musholt1, Tobias Kortus2
1Section of Endocrine Surgery, Department of General, Visceral and Transplantation Surgery, University Medical Center Mainz, Mainz, Germany.
Background:
Intraoperative neuromonitoring-that is, recording of electromyographic signals-is used routinely during (para)thyroid surgery. Surgeons label selected signals to document nerve identity, body side, and time point of stimulation (before or after resection), with a mislabelling rate of 20%. For the purpose of an automated error alert of mislabelled electromyographic signals, the authors developed a multitask one-dimensional convolutional neural network.
Methods:
Raw intraoperative neuromonitoring data were corrected using MIONQA software. Labelled electromyographic signals were extracted and metadata (duration of surgery, timing, median electromyographic peak values of actual surgery) were added to each electromyographic wave. Between 150 and 280 extracted features were used to train, validate, and test various convolutional neural networks.
Results:
Available raw data from a single centre including 1541 operations with continuous intraoperative nerve monitoring and 508 with intermittent intraoperative nerve monitoring between 2014 and 2024 were used. By repeated adjustments of the model architecture and the number of extracted features, an optimized one-dimensional convolutional neural network was designed. After multiple runs with randomized training (11 414 electromyograms) and test (4891) data, the final optimized convolutional neural network achieved a mean(standard deviation) accuracy of 95.72(0.76)% for correct identification of recurrent laryngeal, vagal, and superior laryngeal nerves; 97.68(0.72)% for correct prediction of the resected body side; and 97.61(0.89)% for correct identification of the stimulation time point (before versus after resection). The receiver operating characteristic curve for classification of the electromyographic peak signals had an excellent area under the curve of 0.993.
Conclusion:
The newly developed convolutional neural network enables accurate automated classification of electromyographic peak signals, facilitating the identification and correction of mislabelled intraoperative nerve monitoring data. Such optimized data quality is essential for artificial intelligence training, enabling neuromonitoring machines to alert the surgeon in the operating theatre of mislabelling. Future studies will aim to include a wider range of clinical scenarios and external data sets, in order to further optimize the existing labelling tool and allow clinical applications.
