Abstracts

Rationale: Electroencephalography (EEG) is the gold standard for the diagnosis of epilepsy and detection of seizure activity. The amount of time associated with setting up an EEG recording is directly associated with the number of contacts utilized in the recording. Reducing the number of electrodes required for analysis not only shortens set up times but also reduces caregiver burden for signal quality checks and increases patient comfort during monitoring. There is an increasing number of automated, clinical EEG analysis tools available to aid in the detection of seizures and diagnosis of epilepsy. The performance of any such algorithm is directly tied to the input data used to detect potential seizures. Optimizing the montage used in an automated seizure detection algorithm can improve sensitivity and specificity while reducing computational burden. We investigated the extent to which seizure detection performance is maintained with systematically reduced montages. We also investigated whether the channels included in high sensitivity reduced montages corresponded to the annotated region of seizure onset.

Methods:

EEG recordings from 60 adult patients (8.79 ± 1.3 hours) were reviewed by 3 epileptologists for seizures. Seizure labels were determined through agreement between at least 2 of the 3 reviewers. This yielded 14 patients with a combined 42 seizure events (3 ± 0.43 seizures per patient). Reviewer annotations included a determination of lateralization at seizure onset. All montages were recorded using the 10-20 system. Seizures were automatically detected using an algorithm based on nonlinear systems dynamics. Detection performance was determined by the sensitivity and specificity of detections relative to the established gold standard. In an initial analysis, combinations of all channel pairs that preserved midline symmetry (e.g. C3-P3 and C4-P4) were evaluated. The 4 channel pairs (8 individual channels) which consistently yielded maximal performance across montages were then evaluated without requiring symmetry.

Results: Using symmetric channel pairs, performance converges with increased montage size. This indicates that some channels contain information that is more relevant to detecting seizures than others (Figure 1 A and B). With unconstrained montages, performance remains high across all channel counts (Figure 1C and D). 6 patients had seizures with agreed upon focality at onset (5R, 1L) with most exhibiting secondary generalization. For the unconstrained montages, the highest average sensitivity was obtained when including right hemisphere channels across montage sizes, consistent with the predominant onset hemisphere across the dataset (Figure 1C).

Conclusions:

Our results show that a subset of available electrodes contain the information necessary to detect seizures. Additionally, the optimal subset of electrodes appears to localize to the region of the brain where seizures begin. However, the area of localization is likely dataset specific. Further exploration is needed to utilize the information contained in specific channels to not only optimize automated seizure detection but localize to a general region of seizure onset without a priori knowledge.

Funding: None.