Evaluation of Channel Performance in Seizure Prediction

Hoai An Nguyen, Zhinoos Razavi Hesabi, Levin Kuhlmann

Abstract: Since unprovoked seizures are identified as the biggest concern of epileptic patients, an effective seizure prediction device is definitely a game changer in epilepsy management. In order to apply seizure prediction in practical settings, seizure prediction algorithms are mandated to be optimized prior to installation into implantable devices which often come with low computational resources. To achieve this goal, past literature has demonstrated various machine learning approaches to reduce the number of channels or features needed in seizure prediction. However, the resultant channel or feature set generated by machine learning models can change over time, thus making the device configuration unfeasible once the device is implanted. Therefore, we evaluated the seizure prediction performance of each channel in a set of intracranial electrode-based electroencephalography (iEEG) recordings to uncover if electrode placement impacts on the predictive accuracy. The study was conducted on a dataset of three patients from the first-in-man seizure warning device trial and a state-of-the-art seizure prediction algorithm based on Extra Trees classification was applied. In three patients, the best channels achieved the area under the receiver-operating curve (AUC) scores of 0.77, 0.70, and 0.76, respectively. The result also indicated superior seizure prediction of some channels across all patients. Regarding AUC comparison between the best performing channel and all channels, in patient 1, using one channel as a single predictor surpassed the multiple channel model. While statistically significant differences were not obtained in patient 2, the one channel model offered practical benefits by preventing data overfitting and reducing computational complexity. In patient 3, the general model performed better than individual channels. These findings demonstrate the feasibility of selecting electrodes to build models in some patients and also emphasize the importance of patient-specific methods for seizure prediction.