Abstracts

Rationale: Manual EEG spike quantification is frequently a subjective and time-consuming process. We have previously developed a double threshold wavelet algorithm for quantification of epileptic spikes, and now report improvements to this automated algorithm via addition of a machine learning component in order to further enhance sensitivity while substantially decreasing the rate of false positives. Methods: We applied our algorithm to EEG recordings from eleven pediatric patients (average age of 8 years) with documented ESES (Electrical Status Epilepticus in Sleep). For each patient, one segment of three to eight minutes duration was first analyzed using a double threshold wavelet method to highlight time points corresponding to potential spike events. For each detected spike, the calculated features were the amplitude, slope, and spike angle in every channel. These features were input to a machine learning algorithm to determine whether the time point was indeed a spike. Machine learning classifier training and testing for each set was performed using a 10-fold cross-validation technique. Results were compared against gold-standard visual review by two board certified clinical neurophysiologists. Results: Two different machine learning modalities were tested—Support vector machine (SVM) and Bayesian learning. When SVM was used, the overall sensitivity was 69.6% (+/- 15.3%) and the overall specificity was 95.7% (+/- 10.4%). The kappa statistic was 0.878 and 0.854 respectively for reviewers A and B. When Bayesian learning was utilized, the overall sensitivity was 77.6% (+/- 9.1%) and the overall specificity was 87.6% (+/- 8.3%), and the kappa statistic was 0.780 and 0.771 respectively for reviewers A and B. The detailed results are provided in Table 1, and a graphic summary is shown in Figure 1. Conclusions: Automated spike quantification by wavelet transform is feasible and can be improved by machine learning. Machine learning improved specificity from 64% to as much as 95%, while allowing for maintenance of higher sensitivities. Overall, there was good agreement with the gold-standard of EEG review by human visual inspection. In certain data sets, low sensitivity was noted, which was most likely due to presence of artifact, further highlighted by the fact that these sets displayed the highest amount of disagreement between the two human reviewers. We anticipate that upon further validation, the reported spike detection protocol method may be useful in quantification of epileptiform activity in research and in clinical practice.