Abstracts

Rationale: Hippocampal paroxysmal discharges (HPDs) are a well-known marker of abnormal brain tissue hyperexcitability observed rodent models of epilepsy (Riban et al., 2002). HPDs are characterized by a succession of rhythmic high-amplitude spikes, over a duration that last a few tens of seconds, on average. As opposed to seizure activity, HPDs are not correlated with any behavioral changes. Since the number and length of HPDs increases with tissue excitability, HPDs could be used to quantify the impact of neuromodulatory treatments. This objective motivates the development of automatic detection and classification specific to these events. In this work, we developed two different methods of HPDs detection and quantified their efficiency (sensitivity/specificity).Methods: We used EEG data recorded at 2048 Hz using monopolar electrodes placed in the hippocampus of four epileptic mice (kainate model), between 4 and 6 weeks after kainic acid injection. One challenge in the automatic detection of HPDs is that their frequency content is variable between animals. Therefore, in order to develop robust automated detection techniques, we designed two methods: 1) a method summing the squared gamma power and delta spectral power in the recorded EEG (filter-based); and 2) a method using threshold criteria both on the EEG signal amplitude and its derivative (derivative-based). Receiver operating characteristic (ROC) curves were computed for both techniques in order to evaluate their performance in HPDs detection. Detection was tested using a procedure mimicking ""online"" conditions, where EEG data is read “sample-by-sample” from a datafile already stored on hard disk.Results: The two detectors (filter- and derivative-based) developed (Matlab, The Mathworks, USA) were able to detect reliably HPDs (see an example on Figure 1), and were relatively comparable in detection sensitivity and specificity. The detectors correctly identified on average 90% of HPDs for the derivative-based detector, and 87% of HPDs for the filter-based detector for 10% of false positives, as can be seen on the ROC curves presented in Figure 2. Both methods were able to detect the onset of HPDs in typically less than 100 ms. Regarding the computing time, the derivative-based method was significantly faster than the filter-based method (87 vs. 488 seconds to analyze a one-hour datafile sampled at 2048 Hz).Conclusions: We have developed two automated methods for HPDs detection aiming to reliably detect HPDs despite the variability in the frequency content of HPDs in epileptic mice. These methods could be used in the analysis pipeline of EEG data acquired pre- and post-treatment to evaluate the impact of various forms of therapy on pathological hyperexcitability of brain tissue. Another possible use of these methods would be their implementation in responsive (closed-loop) stimulation protocols triggered at the onset of HPDs.