Abstracts

Rationale: Better methods to localize epileptogenic tissue for targeted resective surgery or neuromodulation in drug resistant epilepsies are required. While interictal spikes and high-frequency oscillations (ripples) are well known electrographic biomarkers for the epileptogenic zone (EZ), these features suffer from low spatial specificity and pathological specificity, respectively. Combined spike ripple (SR) events are a promising biomarker for the EZ that overcomes the limitations of spikes or ripples alone. Current approaches to accurately detect SRs have been developed for non-invasive data and require expert review, which is both prohibitively time consuming and subjective. We sought to develop a fully automated SR detector validated for use in prolonged human intracranial recordings.

Methods: We considered two existing SR detectors: a semi-automated feature-based detector that classifies noninvasive EEG time series data, and a fully automated convolutional neural network (CNN) detector that classifies noninvasive EEG spectrogram images. Here, we combined these two methods with the goal of developing a more precise dual detector suitable for multichannel, prolonged, and invasive EEG recording. To do so, we curated a training dataset of 18 subjects from 4 centers with Engel 1 surgical outcome (5 channels inside the EZ and 5 channels outside the EZ; 10 minutes of recording each). From these data we extracted: (a) 1700 hand-marked SR events (1800 images) from inside the resected volume; (b) 950 false positive detections (1800 images) of the semi-automated detector outside the resected volume; and (c) 1600 randomly selected 1 second samples (1600 images) without any true or false positive detections. We trained the CNN detector with spectrogram images of these events. We then tested performance of the feature-based detector, the CNN detector, and the dual detector against the hand-markings using leave-one-out cross-validation by patient. The dual detector combines the feature-based detector and the CNN detector by only considering events detected by both detectors as SRs. The probability thresholds that led to the optimal performance of the CNN (maximum F1 score) and dual detector (maximum precision) were used.

Results: The CNN detector has a more balanced performance (F1=0.46) compared to the feature-based detector (F1=0.40) or the dual detector (F1=0.43). However, the dual detector has the best precision (0.67) compared to the feature-based detector (0.32) or the CNN detector (0.56). The feature-based detector requires 3 to 15 min to analyze 10 min of 10 channels of data, whereas the CNN detector requires approximately 5 min. The dual detector’s runtime is a summation of the runtimes of the feature-based and CNN detectors.

Conclusions: We introduce a fully-automated CNN detector to detect spike ripples in intracranial EEG data. When precision is a priority, such as in prolonged, multichannel recordings, we introduce a dual detector that incorporates further features for spike ripple detection.

Funding: R01NS119483, R01NS110669