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

Machine learning as a prognostic tool for aneurysmatic subarachnoid haemorrhages

Machine learning als Prognoseinstrument für aneurysmatische Subarachnoidalblutungen

Meeting Abstract

  • presenting/speaker Arthur Gubian - Heidelberg University Hospital, Department of Neurosurgery, Heidelberg, Deutschland
  • Johannes Walter - Heidelberg University Hospital, Department of Neurosurgery, Heidelberg, Deutschland
  • Renan Sánchez-Porras - Heidelberg University Hospital, Department of Neurosurgery, Heidelberg, Deutschland
  • Andreas W. Unterberg - Heidelberg University Hospital, Department of Neurosurgery, Heidelberg, Deutschland
  • Klaus Zweckberger - Heidelberg University Hospital, Department of Neurosurgery, Heidelberg, 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. DocV164

doi: 10.3205/21dgnc159, urn:nbn:de:0183-21dgnc1592

Published: June 4, 2021

© 2021 Gubian 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: Prognostication of functional outcome, drainage dependency and occurrence of delayed cerebral ischemia (DCI) has notoriously been a challenge in Neurosurgery. The development of machine learning algorithms in recent years has seen a revolution in the field of data analysis, with its application reaching medicine. We aimed to develop a tool to prognosticate (1) the long-term functional outcome, (2) the risk of drainage dependency at release/transfer and (3) the risk of occurrence of DCI.

Methods: A prospectively constituted database of patients treated for aneurysmal haemorrhage in our ICU between December 2015 and May 2020 was retrospectively consulted. The relevant variables and parameters for the used models were selected according to the existing literature and after a correlation analysis on the entire database itself. The analysis was performed on Jupyter lab using Python 3.0. Several machine learning models were then tested and, for each model, the accuracy, AUROC, specificity and sensitivity, as well as the NPV, PPV and the Matthew correlation coefficient, were calculated on a randomized validation dataset (consisting of 20% of the overall dataset). The chosen model was then improved using fine parameter tuning until optimal results were obtained. The endpoints were functional outcome at last follow up (dichotomized modified Rankin Scale), drainage dependency (need for a ventriculo-peritoneal shunt between the bleeding and last follow up, or need for an external ventricular drainage at release/transfer) and DCI (appearance of new infarct areas after exclusion of procedure-related infarction).

Results: Data was retrieved for 116 patients between 2015 and 2020.

For the functional outcome at last follow up, the best accuracy was found using the SVC algorithm. The Matthew correlation coefficient was 0.495, NPV0.94, PPP0.5, Sensitivity 0.67, specificity 0.89 and the accuracy was 85.7%.

For the drainage dependency, the best accuracy was found with the Linear Discriminant Algorithm. The Matthew correlation coefficient was 0.614, NPV0.92, PPP0.67, Sensitivity 0.8, specificity 0.85 and the accuracy was 83.3.7%.

For the occurrence of DCI, the best accuracy was found using the Naïves Bayes Gaussian algorithm. The Matthew correlation coefficient was 0.555, NPV 1.0, PPP 0.5, Sensitivity 1.0, specificity 0.62 and the accuracy was 72.2%.

Conclusion: Machine learning may offer itself as a valid prognostic tool in the management of aneurysmatic subarachnoid bleeding.