Subscribe to RSS
Patient-specific multivariate waveform detector for epileptic seizure detection
Intracranial EEG recordings of epilepsy patients consist of numerous channels, which can be highly correlated due to their spatial proximity. We propose a vector-autoregressive (VAR)-based multivariate waveform detector (MWD), which is able to exploit inter-channel correlations for the detection of patient-specific waveforms. The technical novelty is to design the MWD patient-individually and to process several channels simultaneously. Here, the MWD is used for epileptic seizure detection.
We analyzed typical waveforms appearing in seizures of 2 patients suffering from pharmacoresistant focal epilepsy. They underwent implantation of subdural electrodes for presurgical long-term EEG monitoring. We selected 6 channels from a frontal grid electrode in patient A and 3 left temporal channels from strip electrodes of patient B. These channels showed the individual EEG patterns of interest. For analysis, EEG at a sampling rate of 256Hz and a common average reference was used. From the first 3 seizures of each patient, we estimated the matrix correlation function of the interesting signal portion (i.e. the seizure onset waveform, ictal high-frequency (HF) content, or ictal spikes). By computing a VAR model from the correlation estimate, we obtained the MWD.
The signals which were processed by the MWD were averaged and the result was squared, lowpass filtered and the ROC statistics was computed. To demonstrate the benefit of the multivariate design, we also present a degenerated detector without taking the inter-channel correlations into account.
For patient A we designed two different MWDs, one matched to the initial slow wave, and one matched to the ictal HF content. For patient B one MWD was designed matched to ictal spike patterns. The full MWD outperforms the degenerated WD in terms of ROC statistics, at a specificity of 99%, the sensitivity of the MWD was 10% above the degenerated WD. This shows that exploiting the spatial structure increases performance of waveform detection.
We presented a multivariate waveform detector, which is able to exploit both temporal and spatial correlations of neighboring epileptic EEG channels with characteristic patient-specific EEG waveforms. The detector was designed using few given seizure patterns from the respective patient. This waveform detector is not limited to detect epileptic seizures, it can also be designed to detect other graphoelements, such as interictal epileptiform potentials.