Artikel
Machine learning as applied to intraoperative monitoring of the visual evoked potentials during surgery near the anterior visual pathway
Einsatz von maschinellem Lernverfahren zur Analyse der intraoperativ visuell evozierten Potentiale während Schädelbasisoperationen
Suche in Medline nach
Autoren
Veröffentlicht: | 4. Juni 2021 |
---|
Gliederung
Text
Objective: Surgical interventions aimed at removing tumors involving the sellar and parasellar region are a common occurrence in the neurosurgical practice. Despite well established methods to monitor the function of the visual pathway, a clear understanding of the electrical potential changes during surgery and their objective postoperative consequences on the visual field are lacking.
Methods: Visual evoked potentials (VEP) were measured in 12 patients (24 eyes, median age=55, range=28-71, female=7, male=5) with sellar and parasellar tumors during microsurgical transcranial or endoscopic transnasal transsphenoidal approaches using LED pads (Inomed, Emerdingen, Germany) placed over closed eyelids and occipital needle electrodes placed subcutaneously according to the international EEG 10-20 system (O1, O2 as well as Oz and Fz). Changes in the visual field where quantified for each patient eye as percent change of the mean visual field defect (MD, arithmetic mean of sensitivity loss) using the formula (MDpreoperative - MDpostoperative)/MDpreoperative. For each set of visual evoked potentials recorded during surgery we applied a dimensionality reduction algorithm (principal components analysis) to identify the direction of largest variance in the set and subsequently separated the data into clusters using a simple unsupervised machine learning algorithm (k means).
Results: We obtained reproducible VEP waveforms in 75% of the cases (18/24 eyes) during surgery. Of the six eyes where VEP could not be recorded, two had a severe preoperative deficit whereas the other four could be ascribed to technical difficulties. Changes of visual evoked potentials are defined as patient recordings with > 1 cluster (66%, 10/18 eyes). A sanity check was performed in that the individual recordings assigned to clusters where inspected and the clustering rejected if noise clusters where found (10%, 2/18 eyes). We found that the machine learning algorithm reported changes in the recorded visual evoked potential more often than trained human observers (55% of cases as opposed to 20% of cases). Furthermore, detected changes of the VEP strongly associate to improved postoperative mean visual field defects (p=0.006, independent samples Mann-Withney test).
Conclusion: Unsupervised machine learning does better than human observers in identifying changes of the intraoperative visual evoked potentials that have a positive predictive value regarding clinical visual outcomes for surgeries near the anterior visual pathway.