Predicting neurological outcomes following spinal surgery: A machine learning approach using intraoperative neuromonitoring data

Abstract

Background context: Intraoperative neuromonitoring (IONM) has proven effective in reducing postoperative neurological complications. However, current understanding of IONM is limited and its precise meaning in relation to neurological outcomes remains unclear. Machine learning (ML) is a promising solution to analyze the excessive amount of IONM data quickly, objectively and in real-time.
Purpose: The goal is to develop a ML algorithm that can effectively predict neurological outcomes after spinal surgery using IONM data that include both motor evoked potentials (MEPs) and somatosensory evoked potentials (SSEPs), and analyze its key predicting features. To more effectively determine the specific independent contribution of both separate modalities, a separate ML model will be created for both MEP and SSEP in addition to a combined MEP-SSEP model.
Study setting: Retrospective study.
Patient sample: A total of 67 patients were analyzed.
Outcome measures: The neurological status three months postoperatively compared to the preoperative status, categorized into three classes: 'Neurological stable deficits', ‘Neurologically intact’ and 'Neurological improvement'.
Methods: 260 features were obtained from patients who underwent spinal surgery monitored by IONM. During nested cross-validation, the data was split into five folds, for both the inner and the outer loop. The four ML classifiers developed were support vector machine, K-nearest neighbors, random forest and extreme gradient boosting, and tested along the three modalities MEP, SSEP, and MEP-SSEP combination.
Results: Extreme gradient boosting outperformed the other classifiers on all performance metrics. The combined MEP-SSEP model exhibited the highest scores for sensitivity: 70.4%, specificity: 88.3% and accuracy: 87.1%, while the MEP model exhibited the highest performance for precision: 75.6%. Highest predicting scores per individual class were also obtained by this XGBoost classifier on the combined MEP-SSEP model. Key predicting features were the presence or absence of preoperative neurological deficits and last measured signal latency compared to baseline, with a contribution of 29% and 13.5% in the best performing model, respectively.
Conclusion: A reliable prediction of neurological outcomes three months postoperatively can be made combining MEP and SSEP IONM features, provided that the patient's preoperative status is accurately documented and included in the prediction. Though either MEP or SSEP features alone offer predictive value, MEP features show superior predictive values compared to SSEP features when both modalities are accessible, with latency emerging as a prominent predictive IONM feature.