Predicting Postoperative Motor Function After Brain Tumor Resection With Motor Evoked Potential Monitoring Using Decision Tree Analysis

Abstract

Background Motor evoked potential (MEP) monitoring is a commonly employed method in neurosurgery to prevent postoperative motor dysfunction. However, it has low prediction accuracy for postoperative paralysis. This study aimed to develop a decision tree (DT) model for predicting postoperative motor function using MEP monitoring data. Methodology In this retrospective cohort study, we used datasets, comprising 14 variables including MEP amplitudes, obtained from 125 patients who underwent brain tumor resection with intraoperative MEP monitoring at our hospital. Prediction models were developed using DT and receiver operating characteristic (ROC) curve analyses. Model performance was assessed for accuracy, sensitivity, specificity, kappa (κ) coefficient, and area under the ROC curve (AUC) for internal and external validation. For the external validation of the classification model, we retrospectively collected data from an additional 28 patients who underwent brain tumor surgery with MEP monitoring. Results The amplitude of the last measured MEP and amplitude ratio were independent predictors of outcomes. The DT model achieved an accuracy of 0.921, sensitivity of 0.917, specificity of 0.923, and AUC of 0.931 using the internal test. In comparison, the ROC curve based on the amplitude of the last measured MEP achieved a sensitivity of 0.875, specificity of 0.906, and AUC of 0.941. External validation was performed and the DT model was superior to prediction by cutoff values from ROC curves in terms of accuracy, sensitivity, specificity, and κ coefficient. Conclusions Our study suggested the usefulness of DT modeling for predicting postoperative paralysis. However, this study has several limitations, such as the retrospective design and small sample size of the validation dataset. Nonetheless, the DT modeling presented in this study might be applicable to surgeries using MEP monitoring and is expected to contribute to devising treatment strategies by predicting postoperative motor function in various patients.

Keywords: brain tumor; decision tree; machine learning; motor evoked potentials; paralysis.