Machine learning (ML) in intraoperative neuromonitoring (IONM): proof of concept.

PURPOSE: Intraoperative neuromonitoring (IONM) improves safety during pediatric spinal deformity surgery by providing real-time neurophysiological assessment, enabling the earlier detection of neural compromise and the potential prevention of permanent injury. However, current IONM interpretation is subject to variability and relies on human expertise. This study evaluates a machine learning (ML) algorithm designed to identify subtle changes in motor evoked potentials (MEPs) that may precede neurological injury. METHODS: IONM data from 84 pediatric spine surgeries at a single high-volume academic center were retrospectively analyzed. Fourteen patients experienced intraoperative MEP signal loss, with six developing postoperative deficits. An ML model was trained on baseline muscle specific MEPs from each patient to identify signal changes during surgery. The model continuously analyzed real-time MEP data, generating a similarity score relative to baseline. A "red flag" alert was triggered if signal deviation exceeded 10% per minute. Model performance was assessed using sensitivity, specificity, predictive values, and accuracy, and compared against standard clinical alerts. RESULTS: The model achieved 78.6% sensitivity, 67.1% specificity, 94.0% negative predictive value, and 69.0% overall accuracy (AUC = 0.78). ML-generated alerts preceded clinical IONM alerts by an average of 23.3 min. Among patients with postoperative deficits, 5 of 6 were correctly flagged. CONCLUSION: This ML-based tool demonstrated promising early detection of intraoperative neuromonitoring changes, often preceding clinical alerts. Its high sensitivity and negative predictive value suggest potential for real-time support during pediatric spine surgery. Further validation in larger, prospective cohorts is underway. LEVEL OF EVIDENCE: III.