Medical Student Department of Neurosurgery, Louisiana State University Health Shreveport Shreveport, LA, US
Introduction: The difficulty of accurately interpreting intraoperative neuromonitoring (IONM) during spinal surgeries prevents the widespread standardization of its uses. While somatosensory evoked potential (SSEP) and motor evoked potential (MEP) monitoring are instrumental, discerning signal variations due to pharmacology (anesthetic influences and other intraoperative medications) and surgical manipulations during lumbar decompression surgeries remain challenging. This study introduces a machine learning (ML)-based approach to combine these various intraoperative surgical factors in a predictive model. Allowing further credibility to be assigned when signal variations are present without an inciting influence suggesting alarm criteria.
Methods: A dataset comprising SSEP and MEP signals from 150 lumbar procedures was created with associated end-tidal concentrations of various volatile anesthetics (isoflurane, sevoflurane, and desflurane), administered intraoperative medications, and effects of decompression from patients IONM data. Time-synchronized SSEP and MEP trials were analyzed using Matlab R2023a. Features generated via signal processing tools were evaluated and further extracted using minimum redundancy maximum relevance (MRMR) or Chi-squared statistic for assigning different levels of importance for signal variability. Subsequent ML models were then generated and compared based on their ability to accurately discern causes of signal variability.
Results: Using classification modeling with pre-/post-decompression data, 84.6% predictive accuracy was achieved. In the anesthetic influences subset, using regression modeling, a root mean square error (RMSE) of only 0.01262 was achieved.
Conclusion : The superior performance of ML IONM prediction models with only volatile anesthetic influences and surgical manipulations due to decompression is highly suggestive of the ability to quantify seemingly minuscule fluctuations in SSEP and MEP waveforms. Once the impact of intraoperative medications is analyzed, these different aspects of IONM factors affecting signal waveforms can be compiled into a comprehensive signal model for alarm threshold detection. These insights are instrumental in developing a comprehensive prediction model that deciphers the multifaceted influences on IONM signals, enhancing intraoperative decision-making.
.jpg)
