Development of Machine Learning Models to Detect and Predict Signal Changes in Somatosensory Evoked Potentials Following Lumbar Decompression

Introduction: The fidelity of intraoperative neuromonitoring (IONM) during spinal surgeries is pivotal for patient outcomes. While somatosensory evoked potential (SSEP) monitoring is instrumental, discerning signal variations during lumbar decompression surgeries remains challenging. This study introduces a machine learning (ML)-based approach to augment the precision of SSEP signal analysis, facilitating objective detection of signal fluctuations and predictive modeling of waveform presentations post-decompression.

Methods: We collected 4,104 SSEP trials from 85 patients undergoing lumbar decompression and utilized MATLAB R2023a for feature generation, selection, and prediction modeling. The metrics for signal processing included amplitude and latency of the evoked response, time-frequency measures, phase-space curve assessment, and principal components, yielding 132 features per trial (four nerves for each extremity, 3 SSEP channels per nerve, 11 features per channel). We employed the minimum redundancy maximum relevance (MRMR) technique to isolate the top 30 predictive features. Subsequently, a range of ML classifiers were trained to discriminate between pre- and post-decompression SSEP waveforms, with selection based on optimal performance.

Results: MRMR identified the top 30 features used for modeling, with the most significant predictive feature being the latency of the evoked response in the left posterior tibial nerve in channel CPz-Fpz. The ensemble model, utilizing bagged trees preset and decision tree learners, demonstrated a superior accuracy of 84.6% in SSEP signal classification. Additional performance assessments displayed a high-speed prediction capacity (20,000 observations per second) and rapid training time (27.99 seconds).

Conclusion : Achieving 84.6% accuracy underscores the concept of using ML models for robust assessments of complex SSEP waveforms that are typically arduous for clinicians to interpret. Such complexity can lead to erroneous intraoperative decision-making or missed detection of emerging neurological deficits. Ongoing expansion of the dataset, including diverse surgical variables and model refinement, is anticipated to enhance IONM effectiveness further and usher in a new era of ML-assisted spine surgery.