31. Clinical evaluation versus automatic detection of interictal epileptiform discharges – Who can we trust?

Abstract

Objective Interictal epileptiform discharges (IEDs) are electrographic hallmark of epilepsy. Information about the spatiotemporal distribution of IEDs in intracranial EEG is utilized to localize epileptogenic zone during the presurgical evaluation and plan the resection. Visual evaluation of long-term multi-channel intracranial recordings is extremely difficult and prone to bias. Clinicians usually assess only high-amplitude (high signal to noise ratio) discharge and low-amplitude IEDs can be overlooked or considered clinically insignificant. The goal of our study was to develop reliable automatic IED detectors to facilitate analysis of long-term recordings and increase the information yield of intracranial recordings. Methods Seven intracranial EEG recordings were randomly selected from our database. Samples of five minutes duration from fifteen high-rate IED channels (525 min in total) were presented to three experienced EEG specialists for spike labelling. The readers independently reviewed the data and classified IEDs into two groups: obvious and ambiguous. The inter-reader agreement was evaluated and IEDs labelled by at least two readers were considered as a gold standard (GS). We have developed, tested and optimized novel IED detector using GS datasets and compared its performance with published detectors. Our detecting approach estimates the signal envelope distribution to discriminate IEDs from background activity. Results Readers together labelled 6518 IEDs (53 ± 21% obvious, 47 ± 21% ambiguous). The reader’s maximal match was 58% in pair and agreement of all three readers was only 30% (Cohen’s kappa 0.14 ± 0.11). Detector’s performance was characterized by sensitivity 91 ± 12% and 8 ± 7 false positives per min and per channel. Its performance was 1.4× better than published detector. Examination of false positives revealed that substantial proportion had shape of reminiscent of IEDs, but with lower amplitude. More than 50% false positives were reclassified by readers as IEDs. In addition, regression analysis showed positive relationship between IEDs marked by readers and number of false positives. Conclusion The inter-reader agreement in visual IED evaluation is poor. Even experienced readers can identify approximately 40% of IED, especially those with high signal-to-noise ratio. In contrast, automatic detector is 2.5× more sensitive and can identify also low-amplitude IEDs. Areas generating not only high- but also low-amplitude IEDs can be crucial for epileptogenic zone localization. Supported by grants from IGA NT11460, NT13357, NT14489, GACR 14-02634S and Neuron Fund (NFKJ 001/2012).