Visual evaluation of electroencephalogram (EEG) for Interictal Epileptiform Discharges (IEDs) as distinctive biomarkers of epilepsy has various limitations, including time-consuming reviews, steep learning curves, interobserver variability, and the need for specialized experts. The development of an automated IED detector is necessary to provide a faster and reliable diagnosis of epilepsy. In this paper, we propose an automated IED detector based on Convolutional Neural Networks (CNNs). We have evaluated the proposed IED detector on a sizable database of 554 scalp EEG recordings (84 epileptic patients and 461 nonepileptic subjects) recorded at Massachusetts General Hospital (MGH), Boston. The proposed CNN IED detector has achieved superior performance in comparison with conventional methods with a mean cross-validation area under the precision–recall curve (AUPRC) of 0.838[Formula: see text]±[Formula: see text]0.040 and false detection rate of 0.2[Formula: see text]±[Formula: see text]0.11 per minute for a sensitivity of 80%. We demonstrated the proposed system to be noninferior to 30 neurologists on a dataset from the Medical University of South Carolina (MUSC). Further, we clinically validated the system at National University Hospital (NUH), Singapore, with an agreement accuracy of 81.41% with a clinical expert. Moreover, the proposed system can be applied to EEG recordings with any arbitrary number of channels.

Objective The presence of interictal epileptiform discharges (IED) in the electroencephalogram (EEG) is a key finding in the medical workup of a patient with suspected epilepsy. However, inter-rater agreement (IRA) regarding the presence of IED is imperfect, leading to incorrect and delayed diagnoses. An improved understanding of which IED attributes mediate expert IRA might help in developing automatic methods for IED detection able to emulate the abilities of experts. Therefore, using a set of IED scored by a large number of experts, we set out to determine which attributes of IED predict expert agreement regarding the presence of IED. Methods IED were annotated on a 5-point scale by 18 clinical neurophysiologists within 200 30-s EEG segments from recordings of 200 patients. 5538 signal analysis features were extracted from the waveforms, including wavelet coefficients, morphological features, signal energy, nonlinear energy operator response, electrode location, and spectrogram features. Feature selection was performed by applying elastic net regression and support vector regression (SVR) was applied to predict expert opinion, with and without the feature selection procedure and with and without several types of signal normalization. Results Multiple types of features were useful for predicting expert annotations, but particular types of wavelet features performed best. Local EEG normalization also enhanced best model performance. As the size of the group of EEGers used to train the models was increased, the performance of the models leveled off at a group size of around 11. Conclusions The features that best predict inter-rater agreement among experts regarding the presence of IED are wavelet features, using locally standardized EEG. Our models for predicting expert opinion based on EEGer’s scores perform best with a large group of EEGers (more than 10). Significance By examining a large group of EEG signal analysis features we found that wavelet features with certain wavelet basis functions performed best to identify IEDs. Local normalization also improves predictability, suggesting the importance of IED morphology over amplitude-based features. Although most IED detection studies in the past have used opinion from three or fewer experts, our study suggests a “wisdom of the crowd” effect, such that pooling over a larger number of expert opinions produces a better correlation between expert opinion and objectively quantifiable features of the EEG.

Pathological slowing in the electroencephalogram (EEG) is widely investigated for the diagnosis of neurological disorders. Currently, the gold standard for slowing detection is the visual inspection of the EEG by experts, which is time-consuming and subjective. To address those issues, we propose three automated approaches to detect slowing in EEG: Threshold-based Detection System (TDS), Shallow Learning-based Detection System (SLDS), and Deep Learning-based Detection System (DLDS). These systems are evaluated on channel-, segment-, and EEG-level. The three systems perform prediction via detecting slowing at individual channels, and those detections are arranged in histograms for detection of slowing at the segment- and EEG-level. We evaluate the systems through Leave-One-Subject-Out (LOSO) cross-validation (CV) and Leave-One-Institution-Out (LOIO) CV on four datasets from the US, Singapore, and India. The DLDS achieved the best overall results: LOIO CV mean balanced accuracy (BAC) of 71.9%, 75.5%, and 82.0% at channel-, segment- and EEG-level, and LOSO CV mean BAC of 73.6%, 77.2%, and 81.8% at channel-, segment-, and EEG-level. The channel- and segment-level performance is comparable to the intra-rater agreement (IRA) of an expert of 72.4% and 82%. The DLDS can process a 30 min EEG in 4 s and can be deployed to assist clinicians in interpreting EEGs.

No abstract

Automated annotation of electroencephalograms (EEG) of epileptic patients is important in diagnosis and management of epilepsy. Epilepsy is often associated with the presence of epileptiform transients (ET) in the EEG. To develop an efficient ET detector, a vast amount of data is required to train and evaluate the performance of the detector. Interictal EEG data contains mostly background waveforms, since ETs only occur occasionally in most patients. In order to detect ETs in an automated fashion, it is meaningful to first try to eliminate most background waveforms by means of simple, fast classifiers. The remaining waveforms can in a following step be processed by more sophisticated and computationally demanding classification algorithms, such as deep learning systems. In this study, we design a cascade of simple thresholding steps to reject most background waveforms in interictal EEG, while maintaining most ETs. Several simple and quick-to-compute EEG features are chosen. By thresholding these features in consecutive steps, background waveforms are rejected sequentially. In our numerical experiments, a cascade of 10 steps is able to reject 98.65% of all background segments in the dataset, while preserving 90.6% of the ETs.