Blink Reflex Neuromonitoring to Predict a Change in House Brackmann Score: An Initial Machine Learning Approach

Introduction: Cranial nerve preservation is paramount for neurosurgeons during skull base surgery, and intraoperative neurophysiological monitoring (IONM) poses a method for cranial nerve identification and functional tracking. The blink reflex has demonstrated use to monitor both central and peripheral facial nerve function. An electrophysiologic representation of the corneal reflex, the blink reflex uses an electrical stimulus to activate the supraorbital branch of the trigeminal nerve to elicit a response in the orbicularis oculi muscle involving both the trigeminal and facial nerves. The complete utility of the blink reflex as a predictive, intraoperative tool is unknown, and this study aims to understand warning signals for facial nerve injury that may be detected earlier than allowed by standard methods. We seek to do this by applying machine learning algorithms to IONM data.

Methods: Patients older than 18 years undergoing resection of a vestibular schwannoma at our institution were eligible for our study. Patients with a prior cerebellopontine angle surgery, previous cancer, previous facial nerve palsy, or for whom no blink reflex was found at baseline under anesthesia were excluded. In this initial retrospective review, 26 patients from 05/02/2022 until 06/16/2024 were included. A total of 10,282 ipsilateral blink reflexes were collected across all procedures. The data were standardized by centering the mean at zero and scaling it to variance of one, allowing for machine learning estimators and principal component analysis (PCA) to be conducted. Patients’ pre- and postoperative House-Brackmann scores were obtained a difference of greater than 1 was defined as a poor outcome, and a difference of 0 or 1 was defined as a favorable outcome.

Results: Initial results demonstrate patterns of clustering across differences in House-Brackmann score. PCA in both two and three dimensions display regions for potential investigation. These regions demonstrate high dimensional information in the data, allowing for potential trends to be identified in the blink response that could allow for uninterrupted intraoperative monitoring of central and peripheral facial nerve function. Clustering analysis shows distinct clusters that do not fully overlap with true class values, likely due to many poor or unreliable blink reflex traces included. This current study is limited by an unlabeled dataset, leading to potential noise in the data ([Figs. 1], [2]).

Conclusion: This study lays the foundation for further investigation using a larger cohort and robust IONM protocols. Traces are currently being labeled with key findings to allow for cleaner data and supervised machine learning paradigms, including with granularity high enough to enable clinical inference from a single trace. Eventually, we aim to undertake prospective testing of algorithms based on intraoperative changes in blink reflex traces, with the goal of influencing intraoperative strategies and improving postoperative facial nerve outcomes.

Fig. 1 Three-dimensional PCA. Purple: no HB difference, yellow: HB difference greater than or equal to 1.

Fig. 2 Two-dimensional PCA, color indicates HB difference.

Publikationsverlauf

Artikel online veröffentlicht:
07. Februar 2025