Background: Intraoperative monitoring of motor evoked potentials (MEPs) is a well-established technique to assess the functional integrity of motor pathways during neurosurgical procedures. While signal interpretation and warning criteria are mostly based on macroscopic, easily recognizable changes, MEPs are significantly more complex and carry much more information. Recent advancement in data analysis and machine learning technique application proved to be extremely efficient in exploring biological signals. With the aim to explore potential, clinically useful machine learning applications for the analysis of intraoperative MEPs, we conducted a preliminary study to evaluate the ability of different machine learning techniques in classifying intraoperative MEPs to the related muscle. Methods: We retrospectively extracted intraoperative MEPs data for a total of 18 patients who underwent craniotomy for resection of brain tumor. The MEPs collected underwent a screening process by an expert neurophysiologist to select a set of MEPs clearly representing the related muscles. We therefore trained six different types of machine learning algorithms on both raw signals and signals expressed as a set of extracted features and evaluated their accuracy in classifying the different MEPs within the same patient (intra-patient recognition task) and between different patients (inter-patient recognition task). Results: A total number of 25.000 MEPs signals were extracted and a total number of 3300 signals were selected. The performance for the intra-patient muscle classification task reached 99.4% accuracy using random forest technique and a mean of 98% between the six models tested. The performance for the inter-patient muscle classification task ranged between 77% to 99% accuracy for using random forest technique depending on the number of muscles and models considered. Conclusions: This study proved that carefully customized machine learning techniques have the potential to further characterize intraoperative monitoring signals with the aim to refine current warning criteria and make surgeries safer and more effective.
Machine learning approaches for the automated classification of intraoperative motor evoked potentials. A pilot study
A. Boaro;A. Azzari;M. Bicego;F. Sala
2022-01-01
Abstract
Background: Intraoperative monitoring of motor evoked potentials (MEPs) is a well-established technique to assess the functional integrity of motor pathways during neurosurgical procedures. While signal interpretation and warning criteria are mostly based on macroscopic, easily recognizable changes, MEPs are significantly more complex and carry much more information. Recent advancement in data analysis and machine learning technique application proved to be extremely efficient in exploring biological signals. With the aim to explore potential, clinically useful machine learning applications for the analysis of intraoperative MEPs, we conducted a preliminary study to evaluate the ability of different machine learning techniques in classifying intraoperative MEPs to the related muscle. Methods: We retrospectively extracted intraoperative MEPs data for a total of 18 patients who underwent craniotomy for resection of brain tumor. The MEPs collected underwent a screening process by an expert neurophysiologist to select a set of MEPs clearly representing the related muscles. We therefore trained six different types of machine learning algorithms on both raw signals and signals expressed as a set of extracted features and evaluated their accuracy in classifying the different MEPs within the same patient (intra-patient recognition task) and between different patients (inter-patient recognition task). Results: A total number of 25.000 MEPs signals were extracted and a total number of 3300 signals were selected. The performance for the intra-patient muscle classification task reached 99.4% accuracy using random forest technique and a mean of 98% between the six models tested. The performance for the inter-patient muscle classification task ranged between 77% to 99% accuracy for using random forest technique depending on the number of muscles and models considered. Conclusions: This study proved that carefully customized machine learning techniques have the potential to further characterize intraoperative monitoring signals with the aim to refine current warning criteria and make surgeries safer and more effective.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.