Abstracts

Rationale: We have proposed an epileptic seizure prediction method using a machine learning anomaly detection technique based on heart rate variability (HRV) during long-term scalp electroencephalogram (EEG) recordings. Herein, we aimed to validate the feasibility of our seizure prediction method in a case of post-traumatic epilepsy using electrocorticography (ECoG) with both stereoelectroencephalography (SEEG) and subdural electrodes.

Methods: A 24-year-old man, who had a right frontal lobe contusion due to trauma 6 years ago, experienced focal awareness and focal to bilateral tonic-clonic seizures. We considered surgery for drug-resistant seizures. Based on his seizure semiology and distribution of epileptic discharges, we presumed that the epieptogenic zone (EZ) was near the lesion. First, we implanted SEEG electrode. To obtain more detailed EZ information, subdural electrode implantation was also performed.

Two or more epilepsy specialists analyzed both long-term ECoG recordings and identified the seizure onset. To verify our seizure prediction method, we applied the algorithm to preictal and interictal HRV data. It was based on an anomaly detection method using a multivariate statistical process control model that learned interictal HRV data under long-term scalp EEG recordings. Preictal and interictal periods were defined as 15 min prior to seizure onset and an interval of 50 min apart from seizure, respectively. Data with remarkable electrocardiography (ECG) noise were excluded.

Results: For the HRV data during ECoG recordings with the SEEG electrode, the algorithm predicted two of three seizures (sensitivity = 67%),and detected 0.34 false positives (FPs) per hour. All FPs (14) were associated with physical movements. Seven FPs occurred just before awakening, and two of them were detected immediately after sleep onset. For the HRV data during ECoG recording with the subdural electrode, four of five seizures were predicted (sensitivity = 90%), and no FP was reported.

Conclusions: These results suggest that our algorithm can predict seizures with preferable sensitivity and acceptable FP rates for HRV data during ECoG recordings. The performance of the algorithm was similar to that of the EEG recordings. ECoG can detect epileptic discharge more reliably than EEG; thus, the results support the feasibility of our algorithm. Hence, physical movement was suggested to be one of the causes of FPs. Moreover, FPs were likely to occur at the transition between sleep and wakefulness, although the sleep stage needs to be determined not with ECoG, but with scalp EEG. In case of seizure clusters, the interval between the last and current seizures was short. Thus, the former seizure can affect the preictal HRV of subsequent seizures, which is a possible cause of FPs._x000D_ We performed seizure prediction based on HRV in a patient with ECoG recordings and achieved a high rate of seizure prediction and low rate of FPs. The relationship between subclinical seizures in ECoG and FPs was also investigated. Further investigation of ECoG is necessary.

Funding: Please list any funding that was received in support of this abstract.: The authors declare no conflicts of interest associated with this abstract.