Efficient unsupervised algorithms for the detection of seizures in continuous EEG recordings from rats after brain injury

Abstract

Long-term EEG monitoring in chronically epileptic animals produces very large EEG data files which require efficient algorithms to differentiate interictal spikes and seizures from normal brain activity, noise, and, artifact. We compared four methods for seizure detection based on (1) EEG power as computed using amplitude squared (the power method), (2) the sum of the distances between consecutive data points (the coastline method), (3) automated spike frequency and duration detection (the spike frequency method), and (4) data range autocorrelation combined with spike frequency (the autocorrelation method). These methods were used to analyze a randomly selected test set of 13 days of continuous EEG data in which 75 seizures were imbedded. The EEG recordings were from eight different rats representing two different models of chronic epilepsy (five kainate-treated and three hypoxic-ischemic). The EEG power method had a positive predictive value (PPV, or true positives divided by the sum of true positives and false positives) of 18% and a sensitivity (true positives divided by the sum of true positives and false negatives) of 95%, the coastline method had a PPV of 78% and sensitivity of 99.59, the spike frequency method had a PPV of 78% and a sensitivity of 95%, and the autocorrelation method yielded a PPV of 96% and a sensitivity of 100%. It is possible to detect seizures automatically in a prolonged EEG recording using computationally efficient unsupervised algorithms. Both the quality of the EEG and the analysis method employed affect PPV and sensitivity.

Introduction

A large fraction of epilepsy patients have poorly controlled epilepsy. Several recent conferences have addressed the need for better therapies for this patient population (Stab et al., 2002, Stab et al., 2003). These conferences recommended the development of more predictive epilepsy models to assess the clinical efficacy of drugs to suppress seizures and epileptogenesis. In the rat, such chronic epilepsy models include kainic acid (Ben-Ari et al., 1979, Nadler et al., 1978), pilocarpine (Turski et al., 1983), self-sustained status epilepticus (Lothman et al., 1993), and hypoxia-ischemia (Williams et al., 2004).

To determine latency to first seizure after injury, seizure frequency, and mean seizure duration, these models require long-term seizure monitoring. Previous attempts at seizure monitoring have employed simple behavioral observations (Hellier et al., 1998). These attempts suffered from the difficulties implicit in providing 24-h monitoring and from the inability to detect subtle or sub-clinical (non-convulsive) seizures. Continuous EEG monitoring can circumvent these problems. Digitization of the EEG data permits the use of computerized algorithms that automate the seizure detection process; this is a substantial improvement over visual analysis of thousands of hours of video or EEG data. However, long-term digital EEG monitoring generates massive data files, so that efficiency becomes an important factor in the design of the seizure detection algorithms.

Computer algorithms for spike and seizure detection in human EEG recordings have been developed (Dumpelmann and Elger, 1999, Flanagan et al., 2003, Gotman, 1985, Gotman, 1990, Gotman, 1999). The detection techniques developed and marketed for humans are designed to evaluate data from large electrode arrays acquired over relatively short-term recordings (i.e. hours). The algorithms are tuned to state, spike and seizure parameters found in human patients, and require significant computational resources. They are therefore not optimized for analysis of long-term (i.e., months-long) recordings of continuous EEG data from smaller electrode arrays acquired in animal models of epileptogenesis and chronic epilepsy. In addition, available analysis packages for such data are not in the public domain.

In this paper, we describe the signal characteristics of seizures and spikes in two animal models of chronic epilepsy, and then discuss several computationally efficient algorithms used to identify seizures and spikes in continuously acquired radiotelemetric EEG data in these epileptic rats. We also report the sensitivity (true positives divided by the sum of true positives and false negatives, or the likelihood of detecting a true spike) and specificity (true negatives divided by the sum of true negatives and false positives, or the likelihood of identifying an instance where there is no event) of each seizure-detection algorithm.