Abstracts

Rationale:
The epileptic seizure is a fast, dynamic, pathological process. Therefore, high temporal resolution modalities (EEG, MEG, and iEEG) are used as primary diagnostic tools for detecting the onset of a seizure and identifying the network associated with it. The data outcome from these modalities is massive, complex, and must be processed prior to surgery. Currently, a major obstacle in the field is that existing tools for detecting seizure onset rely on a manual examination of the entire collected datasets. Quite often, there are different types of waveforms associated with seizure onsets. As a result, clinicians must review an entire dataset to detect the different wave patterns corresponding to seizure onset. This expert-driven analysis is tedious, time-consuming, and subjective.
Method:
We designed a novel algorithm, Maximum Influence Node Detector (MIND), that automatically detects the seizure onset on EEG, MEG, and iEEG waveforms. The algorithm relies only on the patient’s brain activity that is free from epileptogenic waveform as input for seizure detection. This removes the need to scan the whole dataset manually. To identify the abnormal dynamics that characterize the epileptic seizures, MIND is based on concepts taken from information and graph theories, which are not modality-specific and do not depend on the shape of the brain-waveforms. We use mutual information to calculate the connectivity matrix, and eigencentrality, which measures the impact of a zone on the rest of the network, to find the zones that are involved in the seizure onset. We hypothesize that prior to seizure onset, (i) the zone that generates the seizure has the maximum influence (influmax) among all zones in the brain, (ii) the influence of this zone will be significantly stronger than at any other time without epileptic activity. MIND uses these statistically significant differences in influence to detect the seizure onsets with an accurate time resolution. We trained the MIND algorithm to classify data segments of 7 s-10s and detect if an epileptic seizure occurred within the data segment or not. 
Results:
Our preliminary results show that our novel algorithm, MIND, can detect the onset of epileptic seizures in all three modalities, EEG, MEG, and iEEG. To test the algorithm, we analyzed five cases, one with iEEG, three with EEG and MEG (simultaneously acquired), and one patient with only MEG (with seven seizures). Our experts detected the seizure onset by eye, where for each seizure, the first spike or sharp-wave preceding the propagation of the waveform was considered as the true seizure onset. The results show that the algorithm can detect seizure onset by using only epileptogenic free activity as a baseline.
Conclusion:
There are various automated and semi-automated detectors of seizure onset. None of them can successfully meet all of these essential features: 1) Objective and accurate detection spatially and temporally, 2) Non modal-dependent, 3) High sensitivity and specificity, 4) Non-patient-specific training or manual identification of seizure exemplars, and 5) Sensitive enough to complicated cases, as very brief seizures, very long seizures, fast propagated seizures, or a mix of interictal discharges with frequent ictal activities. MIND meets all of these requirements. The scientific premise of our work is that when epileptic seizures occur, the epileptogenic zone influences the rest of the brain far more than any other region, and the exact timing of such influence. MIND is non-modality-dependent, non-patient-specific, and precise. Unlike most existing algorithms, MIND also sheds light on the underlying mechanism of the seizure-phenomena by investigating the epileptic seizure network.
Funding:
:NIH: R21-NS101373-01A1 1R01DC016915-01 R01-NS069696