Automating Epilepsy Surgery Planning with Graph Neural Networks: Retrospective Study
Abstract number :
3.287
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2024
Submission ID :
434
Source :
www.aesnet.org
Presentation date :
12/9/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Petr Nejedly, Msc – St. Anne's university hospital Brno, Czech Republic
Valentina Hrtonova, MSc. – Institute of Scientific Instruments, The Czech Academy of Sciences, Brno, Czech Republic
Martin Pail, MD, PhD – Institute of Scientific Instruments of the CAS, Brno, Czech Republic
Irena Dolezalova, doc – 1 St Department of Neurology, Faculty of Medicine, Masaryk University, Brno, Czech Republic
Jan Cimbalnik, PhD – International Clinical Research Center, St. Anne’s University Hospital, Brno, Czech Republic
Vojtech Travnicek, Ing – International Clinical Research Center, St. Anne's University Hospital, Brno, Czech Republic
Pavel Jurak, PhD – Institute of Scientific Instruments of the CAS, Brno, Czech Republic
Petr Klimes, PhD – Institute of Scientific Instruments, The Czech Academy of Sciences, Brno, Czech Republic
Milan Brázdil, MD, PhD – 1st Department of Neurology, Faculty of Medicine, Masaryk University and St. Anne´s University Hospital, Brno, Czech Republic – member of ERN EpiCARE
Rationale: Epilepsy surgery planning is intricate, demanding precise localization of epileptogenic zones to improve surgical results. Traditional approaches to analyzing intracranial EEG typically overlook spatial features. This study explores the innovative integration of Graph Neural Networks (GNNs) with spatial coordinates of intracranial EEG electrodes along with EEG biomarkers and MRI imaging features, presenting promising advancements in surgical planning. The efficacy of GNNs is compared with traditional Neural Networks (NNs) to evaluate their ability in accurately predicting epileptogenic zones using iEEG data from patients scheduled for epilepsy surgery.
Methods: In this retrospective study, data from 75 patients who underwent epilepsy surgery at St. Anne's University Hospital, Brno, Czech Republic were analyzed. Both GNNs and traditional NNs were employed, utilizing the same set of features and model architecture. The distinction was that GNNs additionally considered the spatial distance between electrodes. Performance metrics were evaluated through a leave-one-out cross-validation testing scheme.
Results: GNNs demonstrated superior performance with an Area Under the Receiver Operating Characteristic (AUROC) of 0.894 and an Area Under the Precision-Recall Curve (AUPRC) of 0.667, compared to NNs which showed an AUROC of 0.735 and an AUPRC of 0.411 in patients with good surgical outcomes (Engel I). In contrast, for patients with poor outcomes (Engel III+IV), the AUROC was 0.79 and AUPRC was 0.31, indicating lesser alignment with actual resected electrodes. Additionally, 70% of patients with good outcomes had the "model’s selected most epileptic electrode" resected, versus only 30% in patients with poor outcomes.
Conclusions: The results indicate that integrating spatial features into Graph Neural Networks (GNNs) substantially improves the accuracy of epilepsy surgery planning, offering promising prospects for its automation. The GNN model demonstrated superior performance in patients with favorable outcomes, showing a high degree of alignment with clinical decisions. However, the model's recommendations diverged from clinical decisions in patients with poor outcomes. To validate these findings further, additional research using multicentric datasets is necessary.
Funding: Supported by the Czech Science Foundation, project n. 21-44843L.
Neurophysiology