Standard electrode arrays for recording EEG are placed on the surface of the brain. Detection of High Frequency Oscillations Using Support Vector Machines: A Case Study Margret Una Kjartansdottir³, Haimonti Dutta¹, Catherine A. Schevon², Ansaf Salleb-Aouissi¹, David Waltz¹ and Ronald Emerson² ¹Center for Computational Learning Systems, ²Columbia University Medical School (CUMC), ³Department of Computer Science, Columbia University, New York, NY. Epilepsy is a chronic neurological disorder. Patients suffer from recurrent, unprovoked seizures. 65% of patients can be treated with antiepileptic drugs. 8-10% need resective surgery to eliminate seizures. During surgery patients are subjected to long-term electroencephalogram (EEG) monitoring. Two features - Power and Mean Curve Length (MCL) were extracted from the positive and negative samples and normalized. Parameters for Linear, Polynomial and Radial Basis Function (RBF) kernels, C, d and gamma were chosen by grid search. The performance of the SVM is measured in terms of Area Under the ROC Curve (AUC). The Linear kernel with parameter C = 500 provides the best AUC= V. N. Vapnik The nature of statistical learning theory Springer-Verlag New York, Inc., C. A. Schevon, A. J. Trevelyan, et al. (2009b). Spatial characterization of interictal high frequency oscillations in epileptic neocortex. Brain 132(11): This work is supported by funding from Epilepsy Research Foundation and National Science Foundation IIS The authors would like to thank Hatim Diab and Sam Lee for their help during different phases of the project. Support Vector Machines (SVM) were used to distinguish between HFOs and a randomly chosen snippet of EEG. Positive samples (HFOs) were non-overlapping windows of size 3 seconds including one or more epochs of size 6 ms labeled as an HFO by the detection algorithm. The negative samples (non-HFOs) were formed by random sampling. The final data set consisted of 1686 training instances and 722 test instances. Signal is down sampled to 2000 Hz. Filtered with 4 th Order Butterworth high pass filter Artifacts > 100 microvolt and <-100 microvolt removed. Original signal Count of HFO’s by channel layout. One electrode is supplemented with a specialized Micro Electrode Array (MEA). The ROC curve obtained from the best performing SVM with Linear Kernel. Difference in average MCL of HFO’s and non-HFO’s. Difference in average Power of HFO’s and non-HFO’s. Introduction Data Collection Filter and Artifact Removal High Frequency Oscillation detection algorithm Acknowledgement References Conclusions Features Machine Learning: SVM MEA measures 4 mm x 4 mm with 96 channels recording signals at 30 kHz After filtering and artifact removal = Micro Electrode Array = Seizure Onset HFOs may be potential biomarkers for epileptogenic activity. HFOs are brief bursts in the high gamma band ( Hz). HFOs last for milliseconds. High Frequency Oscillations (HFOs) Criteria for detections: 1.An epoch of 6 ms must have root mean square (RMS) amplitude of at least 5 standard deviations above the mean for that channel. 2. At least 4 positive or negative peaks whose amplitude is greater than 3 standard deviations above the mean should be present. Automatic HFO detection algorithm 90 second sample recording in 96 channels. HFO detection algorithm was run on 12 hours of MEA data from one patient. ParametersAUC Linear Kernel C = C = C = C = C = C = Polynomial Kernel d = RBF Kernel gamma = gamma = The results obtained from building the HFO classifier. False Positive Rate True Positive Rate Decision boundary obtained to separate the HFOs vs. Non-HFOs. Power MCL