Introduction: Monitoring electromyograms (EMGs) during skull base surgeries is crucial to prevent cranial nerve injuries, which are common complications of skull base surgery. In order to enhance the safety and efficacy of surgical procedures, using machine learning (ML) algorithms to classify EMG signals can improve the recognition of muscle activation patterns.
Methods: This research utilized a public dataset (DOI: 10.17632/7hyptcbkkd.2) to monitor the EMG obtained from five cranial nerves of 11 patients during cerebellopontine angle tumour surgery. Specifically, the EMG data were collected using the Neuromaster G1 MEE-2000 (Nihon Kohden, Inc., Tokyo, Japan) from the V, VII, XI, X, and XII cranial nerves. An ML model was developed using MATLAB 2023b, based on an ensemble of decision trees, to classify EMG signals into 'Injury', 'Artifact', and 'Healthy' categories. The features used include the amplitude of the rectified value, root mean square value, median frequency, total power, and mean normalized frequency. The data were split using holdout with 80% for training and 20% for testing. Synthetic minority oversampling was applied to the training data to balance the classes; 800 maximum splits per tree were configured with limits of 5 observations per leaf and 10 per parent node. The model was trained through 250 learning cycles with pruning enabled to improve generalization. Subsequently, the model was validated using 5-fold cross-validation, ensuring a robust evaluation of its performance.
Results and discussion: The model achieved on the test set an overall accuracy of 81.12%, with 32.49% precision and 81.01% recall for Injury, 70.00% precision and 75.68% recall for Artifact, and 97.54% precision and 82.12% recall for Healthy, with F1-scores of 46.38%, 72.73%, and 89.17%, respectively.
Conclusion: This study demonstrated the potential of ML in EMG for intraoperative monitoring of cranial nerves, suggesting future optimizations and the integration of advanced algorithms to further improve diagnostic accuracy and clinical utility.
