Authors :
Presenting Author: Petr Nejedly, MSc – The International Clinical Research Center of St. Anne's University Hospital in Brno
Vojtech Travnicek, Researcher – Institute of Scientific Instruments, The Czech Academy of Sciences, Brno, Czech Republic; Kristyna Pijackova, Researcher – Institute of Scientific Instruments, The Czech Academy of Sciences, Brno, Czech Republic; Valentina Hrtonova, Researcher – Institute of Scientific Instruments, The Czech Academy of Sciences, Brno, Czech Republic; Jan Cimbalnik, Researcher – International Clinical Research Center, St. Anne's University Hospital, Brno, Czech Republic; Martin Pail, MD – International Clinical Research Center, St. Anne's University Hospital, Brno, Czech Republic; Pavel Jurak, Researcher – Institute of Scientific Instruments, The Czech Academy of Sciences, Brno, Czech Republic; Milan Brazdil, MD – 1 St Department of Neurology, Faculty of Medicine, Masaryk University, Brno, Czech Republic; Petr Klimes, Researcher – Institute of Scientific Instruments, The Czech Academy of Sciences, Brno, Czech Republic
Rationale:
Accurate localization of the seizure onset zone (SOZ) is a fundamental step in epilepsy surgery. Intracranial electroencephalogram (iEEG) biomarkers such as interictal epileptiform discharges, high-frequency oscillations, and relative entropy have been developed to automate the SOZ localization process. In this study, we compare the performance of commonly used features with a deep-learning model for SOZ localization. The study was evaluated on a cohort of 39 patients (pseudo-prospective hold-out test set) who underwent epilepsy surgery at St. Anne's University Hospital, Brno, the Czech Republic. The results and implications of this comparative analysis are discussed, highlighting the potential of deep-learning techniques in improving the accuracy and efficiency of SOZ localization for epilepsy surgery.Methods:
Here, we introduce a deep-learning approach based on U-net architecture for the processing of iEEG data. The primary objective was to train a neural network capable of accurately detecting interictal epileptiform discharges (IED) within the seizure onset zone (SOZ) regions. The motivation to train the neural network for spike detection is based on assumptions that SOZ has the highest spike occurrence compared with other non-SOZ electrodes. A training set comprising data from 50 patients was utilized for model development. To enhance the classifier's robustness, the training signals were deliberately corrupted with noise and artifacts, simulating common artifacts encountered in real-world scenarios. This methodology aimed to ensure that the resulting classifier remained insensitive to such artifacts, thereby improving its reliability and applicability in practical settings.Results:
To evaluate the results, we deployed detectors on 30-minute recordings while patients were awake and resting (39 patients in the holdout test set), according to standard protocol at St. Anne's University Hospital. We show that the proposed neural network was able to localize SOZ with the highest area under the precision-recall curve (AUPRC), while compared with state-of-the-art methods such as IED Janca-detector, HFO-CS detector, and Relative entropy (frequency band 80-250 Hz). The results (Table 1) are sorted according to AUPRC for Good outcome patients. The AUPRC was chosen as the comparison metric since it is less sensitive to True Negative detections in heavily imbalanced datasets (e.g., the proposed dataset has approximately 4% of iEEG channels marked as SOZ).Conclusions:
In this pseudo-prospective study, we evaluated the performance of a deep-learning method based on U-net architecture in comparison to state-of-the-art techniques for detecting interictal epileptiform discharges (IEDs) and high-frequency oscillations (HFOs). Our findings demonstrate that the deep-learning approach outperformed other methods in localizing seizure onset zones from interictal segments in patients with drug-resistant epilepsy. The results highlight the potential of deep-learning methods, particularly in the context of IED detection for seizure onset zone localization, and provide valuable insights for further advancements.Funding:
Supported by the Czech Science Foundation, project n. 21-44843L.