Abstracts

Rationale: Neonatal hypoxia-ischemia (NHI) is serious conditioned that often results in seizures and significant additional morbidity. It is encountered in between 0.1 and 0.7% of all births in this country. Abnormal electrical activity following NHI may be related to the likelihood and also frequency of subsequent seizures. Specifically, we would like to investigate if interictal spikes serve as a biomarker for chronic seizure activity. To do this, it is necessary to accurately, or at least consistently, quantify these discharges. Further, because of the large quantity of data that must be processed, the use of automated detection is highly desirable. In this abstract we review several automated methods of spike detection and compare these with the results of expert analysis. Methods: The EEG data was generated using electrode implanted P10 Sprague-Dawley rats that had undergone hypoxia using a method developed by Jensen (7% oxygen for 8 minutes, 6% for 4 minutes, 5% for 2 minutes and 4% for 1 minute). The SVERAT computer code, written using Visual Basic routines, was used both in the automated analysis, and also to facilitate the expert analysis. For the automated analysis, algorithms using slope, total travel, and deterministic finite automata (DFA) were used. For the expert analysis, the code scrolled through the data and the experts would use mouse clicks to indicate the presence of spikes. Experts included a neurologist, a basic science researcher and a technician, all of whom had significant experience in the field. Because none of the evaluations could be considered the "gold standard," neither sensitivity nor specificity could be defined. Instead, we use compute percent agreement as the number of common spikes divided by the number of total spikes. Comparisons were performed using a T-test or an ANOVA with Bonferroni correction. Results: Expert review resulted in an average percent agreement of 61%. Average agreement among the automated methods was not significantly better (P=0.08) at 77%. Agreement with automated methods was better as expertise increased (75% for the neurologist, 68% for the basic science researcher and 59% for the technician, p<0.01). The average percent agreement was better, but not significantly so, for the DFA method than either the slope method or the travel method. Conclusions: Expert identification of spikes in this model, as has been noted in the human case, remains inconsistent. It is also inconsistent with the results of automated spike identification. Automated spike detection does seem to improve slightly as the complexity of the algorithm increases. Results of more experienced investigators seem to more closely match those of the automated results.