gms | German Medical Science

72. Jahrestagung der Deutschen Gesellschaft für Neurochirurgie (DGNC)
Joint Meeting mit der Polnischen Gesellschaft für Neurochirurgie

Deutsche Gesellschaft für Neurochirurgie (DGNC) e. V.

06.06. - 09.06.2021

Data-driven surgical decision-making – prediction of resectability in patients with glioblastoma using machine learning

Datengesteuerte chirurgische Entscheidungsfindung – Vorhersage der Resektabilität bei Glioblastompatienten mittels Machine Learning

Meeting Abstract

  • presenting/speaker Julius Maximilian Kernbach - Universitätsklinikum Aachen, Klinik für Neurochirurgie, Aachen, Deutschland; Universitätsklinikum Aachen, Neurosurgical Artificial Intelligence Laboratory Aachen (NAILA), Aachen, Deutschland
  • Georg Neuloh - Universitätsklinikum Aachen, Klinik für Neurochirurgie, Aachen, Deutschland
  • Lasse Dührsen - University Medical Center Hamburg-Eppendorf, Department of Neurosurgery, Hamburg, Deutschland
  • Franz Lennard Ricklefs - University Medical Center Hamburg-Eppendorf, Department of Neurosurgery, Hamburg, Deutschland
  • Stefan Grau - University Hospital Cologne, University of Cologne, Department of Neurosurgery, Köln, Deutschland
  • Hans Clusmann - Universitätsklinikum Aachen, Klinik für Neurochirurgie, Aachen, Deutschland
  • Daniel Delev - Universitätsklinikum Aachen, Klinik für Neurochirurgie, Aachen, Deutschland; Universitätsklinikum Aachen, Neurosurgical Artificial Intelligence Laboratory Aachen (NAILA), Aachen, Deutschland

Deutsche Gesellschaft für Neurochirurgie. 72. Jahrestagung der Deutschen Gesellschaft für Neurochirurgie (DGNC), Joint Meeting mit der Polnischen Gesellschaft für Neurochirurgie. sine loco [digital], 06.-09.06.2021. Düsseldorf: German Medical Science GMS Publishing House; 2021. DocV127

doi: 10.3205/21dgnc122, urn:nbn:de:0183-21dgnc1229

Published: June 4, 2021

© 2021 Kernbach et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution 4.0 License. See license information at http://creativecommons.org/licenses/by/4.0/.


Outline

Text

Objective: Surgical decision-making in patients with glioblastoma (GBM) remains challenging as it aims at achieving maximal resection without causing neurological deterioration. Computational approaches leveraging novel artificial intelligence and machine-learning methods may enhance the decision-making process in a “data-driven” fashion – realizing a step towards personalized medicine.

Methods: We conducted a machine learning-based examination of multicentric data from four sites, including 480 patients with GBM (mean [SD] age 62 [12] years; 191 female) in total, of which 47% were found to have complete excision of the contrast-enhanced tumour area. Resection was classified as gross-total versus partial. Data were randomly split into a development set (80%) and a validation set (20%). The model selection included AdaBoost, GradientBoost, Logistic Regression, and Random Forest. L2-penalized Logistic Regression was selected as most competitive and hence evaluated on the validation set in 1000 bootstrap iterations.

Results: The overall accuracy of the learned predictive model was determined by comparing the predicted resectability with the actual rate of removal, which resulted in an area-under-the-curve (AUC, Figure 1 [Fig. 1]) value of 0.70 (95% CI 0.61- 0.79) on the validation set, confirming the generalizability of the approach. Precision and recall of 0.68 (95% CI 0.60-0.77) and 0.67 (0.58-0.75) were observed, respectively. Based on the coefficient weights of the predictive machine learning approach (Figure 2 [Fig. 2]), the use of intraoperative monitoring and the application of 5-ALA, positively impacted the rate of gross-total resection, while unfavorable locations, as well as overall tumor volume, had a negative impact.

Conclusion: The proposed machine learning-based framework allows for reliable prediction of gross-total resectability in patients with GBM and is hoped to complement and ease surgical decision-making, highlighting the benefits of intraoperative fluorescence and monitoring, as well as unfavorable topographical tumour patterns.