Abstract
This paper presents a patient-specific approach for electroencephalography (EEG) channel selection and seizure prediction based on statistical probability distributions of the EEG signals. This approach has two main phases; training and testing phases. In the training phase, few hours of multi-channel nature for each patient representing normal, pre-ictal, and ictal activities are selected. These hours are segmented into non-overlapping 10-s segments and probability density functions (PDFs) are estimated for the signals, their derivatives, local means, local variances, and medians. These PDFs have multiple bins, which are studied separately as random variables across different segments of the same nature. Depending on the PDFs of these random variables for different signal activities and on predefined prediction and false-alarm probability thresholds, bins are selected from certain channel distributions for seizure prediction. In the testing phase, the selected bins only are used for classification of each signal segment activity into pre-ictal or normal states in the prediction process. In the final prediction step, an equal gain decision fusion process is performed leading to a discrete decision sequence representing the activities of all segments. This sequence is filtered with a moving average filter and compared to a patient-specific prediction threshold. Moreover, we have studied the effect of a lossy compression technique on the accuracy of the proposed algorithm using discrete sine transform (DST) compression. This system can be implemented for communication between headset and mobile to give alerts for patients.