Unsupervised machine learning can delineate central sulcus by using the spatiotemporal characteristic of somatosensory evoked potentials

• Published in: Journal of neural engineering Vol. 18; no. 4; pp. 46038 - 46051
• Main Authors: Asman, Priscella, Prabhu, Sujit, Bastos, Dhiego, Tummala, Sudhakar, Bhavsar, Shreyas, McHugh, Thomas Michael, Ince, Nuri Firat
• Format: Journal Article
• Language: English
• Published: England IOP Publishing 29.04.2021
• Subjects: Evoked Potentials, Somatosensory; functional mapping; Hand; Humans; Intraoperative Neurophysiological Monitoring; machine learning; Motor Cortex; somatosensory evoked potentials; Unsupervised Machine Learning; machine learning; functional mapping; somatosensory evoked potentials
• ISSN: 1741-2560; 1741-2552
• DOI: 10.1088/1741-2552/abf68a

Somatosensory evoked potentials (SSEPs) recorded with electrocorticography (ECoG) for central sulcus (CS) identification is a widely accepted procedure in routine intraoperative neurophysiological monitoring. Clinical practices test the short-latency SSEPs for the phase reversal over strip electrodes. However, assessments based on waveform morphology are susceptible to variations in interpretations due to the hand area's localized nature and usually require multiple electrode placements or electrode relocation. We investigated the feasibility of unsupervised delineation of the CS by using the spatiotemporal patterns of the SSEP captured with the ECoG grid. . Intraoperatively, SSEPs were recorded from eight patients using ECoG grids placed over the sensorimotor cortex. Neurosurgeons blinded to the electrophysiology identified the sensory and motor gyri using neuronavigation based on sulcal anatomy. We quantified the most discriminatory time points in SSEPs temporal profile between the primary motor (M1) and somatosensory (S1) cortex using the Fisher discrimination criterion. We visualized the amplitude gradient of the SSEP over a 2D heat map to provide visual feedback for the delineation of the CS based on electrophysiology. Subsequently, we employed spectral clustering using the entire the SSEP waveform without selecting any time points and grouped ECoG channels in an unsupervised fashion. Consistently in all patients, two different time points provided almost equal discrimination between anterior and posterior channels, which vividly outlined the CS when we viewed the SSEP amplitude distribution as a spatial 2D heat map. The first discriminative time point was in proximity to the conventionally favored ∼20 ms peak (N20), and the second time point was slightly later than the markedly high ∼30 ms peak (P30). Still, the location of these time points varied noticeably across subjects. Unsupervised clustering approach separated the anterior and posterior channels with an accuracy of 96.3% based on the time derivative of the SSEP trace without the need for a subject-specific time point selection. In contrast, the raw trace resulted in an accuracy of 88.0%. . We show that the unsupervised clustering of the SSEP trace assessed with subdural electrode grids can delineate the CS automatically with high precision, and the constructed heat maps can localize the motor cortex. We anticipate that the spatiotemporal patterns of SSEP fused with machine learning can serve as a useful tool to assist in surgical planning.