An automated, machine learning-based detection algorithm for spike-wave discharges (SWDs) in a mouse model of absence epilepsy

Manual detection of spike-wave discharges (SWDs) from EEG records is time intensive, costly and subject to inconsistencies and biases. Additionally, manual scoring often omits information on SWD confidence/intensity which may be important for the investigation of mechanistic-based research questions. While there are some automated and semi-automated methods for the detection of SWDs in humans and rats, there has been minimal development of these methods for SWDs in mice. Here we develop a support vector machine (SVM)-based algorithm for the automated detection of SWDs in the gamma2R43Q mouse model of absence epilepsy. The algorithm first identifies putative SWD events using frequency- and amplitude-based peak detection. Four humans experienced at identifying SWDs scored a set of 2500 putative events identified by the algorithm. Then, using predictors calculated from the Morlet wavelet transform of each event and the labels from the human scored events, we trained a SVM to classify (SWD/nonSWD) and assign confidence scores to each event identified from 60 24-hour EEG records (11 animals). We provide a detailed assessment of intra- and inter-rater scoring that demonstrate the advantages in reliability of automated scoring over human scoring. The result of this algorithm is a scoring of SWDs along a confidence continuum that is highly correlated with human confidence and that allows us to more effectively characterize ambiguous events in order to gain a better understanding of ground truth in SWD detection (i.e., the most reliable features agreed upon by human scorers that also correlate with objective physiological measures). Finally, we demonstrate that events along our scoring continuum are temporally and proportionately correlated, on the scale of seconds, with abrupt changes in spectral power bands relevant to normal behavioral states including sleep. Our results demonstrate the value of viewing detection of SWDs as a continuous, rather than discrete, classification problem.