Scott B. Wilson

Spike detection: a review and comparison of algorithms

For algorithm developers, this review details recent approaches to the problem, compares the accuracy of various algorithms, identifies common testing issues and proposes some solutions. For the algorithm user, e.g. electroencephalograph (EEG) technician or neurologist, this review provides an estimate of algorithm accuracy and comparison to that of human experts. Manuscripts dated from 1975 are reviewed. Progress since Frosts 1985 review of the state of the art is discussed. Twenty-five manuscripts are reviewed. Many novel methods have been proposed including neural networks and high-resolution frequency methods. Algorithm accuracy is less than that of experts, but the accuracy of experts is probably less than what is commonly believed. Larger record sets will be required for expert-level detection algorithms.

Spike detection II: automatic, perception-based detection and clustering.

OBJECTIVES We developed perception-based spike detection and clustering algorithms. METHODS The detection algorithm employs a novel, multiple monotonic neural network (MMNN). It is tested on two short-duration EEG databases containing 2400 spikes from 50 epilepsy patients and 10 control subjects. Previous studies are compared for database difficulty and reliability and algorithm accuracy. Automatic grouping of spikes via hierarchical clustering (using topology and morphology) is visually compared with hand marked grouping on a single record. RESULTS The MMNN algorithm is found to operate close to the ability of a human expert while alleviating problems related to overtraining. The hierarchical and hand marked spike groupings are found to be strikingly similar. CONCLUSIONS An automatic detection algorithm need not be as accurate as a human expert to be clinically useful. A user interface that allows the neurologist to quickly delete artifacts and determine whether there are multiple spike generators is sufficient.

Seizure detection: evaluation of the Reveal algorithm

OBJECTIVE The aim of this study is to evaluate an improved seizure detection algorithm and to compare with two other algorithms and human experts. METHODS 672 seizures from 426 epilepsy patients were examined with the (new) Reveal algorithm which utilizes 3 methods, novel in their application to seizure detection: Matching Pursuit, small neural network-rules and a new connected-object hierarchical clustering algorithm. RESULTS Reveal had a sensitivity of 76% with a false positive rate of 0.11/h. Two other algorithms (Sensa and CNet) were tested and had sensitivities of 35.4 and 48.2% and false positive rates of 0.11/h and 0.75/h, respectively. CONCLUSIONS This study validates the Reveal algorithm, and shows it to compare favorably with other methods. SIGNIFICANCE Improved seizure detection can improve patient care in both the epilepsy monitoring unit and the intensive care unit.

Seizure detection: correlation of human experts

OBJECTIVE The description and application of a new, overlap-integral comparison method and the quantification of human vs. human accuracies that can be used as goals for algorithms. METHODS Four human experts marked ten 8 h electroencephalography (EEG) records from seizure patients. The seizures varied in origin and type, including complex partial, generalized absence, secondarily generalized and primary generalized tonic-clonic. The traditional any-overlap comparison method is used in addition to the overlap-integral method, which is sensitive to the correct placement of the seizure endpoints. RESULTS The number of events marked by each reader ranged from 57 to 77. The average any-overlap sensitivity and false positives per hour rate are 0.92 and 0.117. The average overlap-integral correlation, sensitivity and specificity are 0.80, 0.82 and 0.9926. As expected, the correspondence between readers is high, but confounding issues resulted in overlap-integral sensitivities less than 0.5 for 10% of the records. Seven percent of the any-overlap sensitivities are less than 0.5. A comparison of the methods by record shows that the overlap-integral specificity and the any-overlap false positive rate measure different features. CONCLUSIONS There was little variation between readers and they were essentially interchangeable. High seizure rate (many per hour), short seizure durations (<10 s) and long seizure durations (approximately 10 min) with ambiguous offsets can complicate the analysis and result in poor correlation. There may be any number of unmarked events in rigorously marked records and it may be preferable to use records from non-epilepsy patients to compute the false positive rate. The any-overlap and overlap-integral comparison methods are complementary. SIGNIFICANCE Correlation between expert human readers can be low on some records, which will complicate testing of seizure detection algorithms.

Algorithm architectures for patient dependent seizure detection

OBJECTIVE The goal of this work is to determine whether improved performance (compared to patient independent algorithms) can be achieved by an algorithm, developed on the fly, that requires no user input beyond the identification of the first one or two seizures in the record. METHODS The previously developed AutoLearn algorithm, which employs the probabilistic neural network (PNN), is tested on 209 seizures obtained from the epilepsy monitoring unit (EMU) or ambulatory recordings. A construction algorithm is used to compare a variety of algorithm architectures and factors. The Taguchi design of experiments (DoE) method is employed find the significant factors without resorting to a full factorial design. RESULTS Architectures that train a single PNN per channel and use segmentation to identify ranges of similar activity are preferred. The two best architectures are insensitive to the levels of any of the other factors tested. The training time for the algorithm is less than 1s, and approximately 2 min are required to find the seizures in an 8 h record. CONCLUSIONS The final algorithm, which requires no input from the user other than the marking of the first seizure in a record, performs as well or better than the 3 seizure detectors on EMU and ambulatory records. The algorithm performs nearly as well as human experts on the EMU records. SIGNIFICANCE The described method can be used to identify unusual seizures (or other patterns) that will be missed by the current generation of seizure detectors. We expect that the methods developed here will also aid the development of patient independent seizure detectors that can improve their performance over time by incorporating new examples.

Systematic approach to dipole localization of interictal EEG spikes in children with extratemporal lobe epilepsies

OBJECTIVES To assess the reliability of dipole localization based on residual variances (RV), using equivalent current dipole analysis of interictal EEG spikes in children with extratemporal lobe epilepsy. METHODS Four pediatric patients with extratemporal lobe epilepsy were studied. Digital EEG was recorded from 19 scalp electrodes. Computer programs for spike detection and clustering analysis were used to select spikes. Dipoles were calculated 5 times for each spike using different initial guesses by the moving dipole model. Standard deviation (SD) of the dipole positions was calculated at each time point in the 5 trials. RESULTS We analyzed the dipoles at 1097 time points from 4 patients. Among 106 time points with RV < 2%, the SD was < 1 mm in 78 (74%), while in those with SD > 1 mm the dipole positions varied between 2.8 and 52.6 mm. Of dipoles with RV < 1%, 26 of 27 (96%) had an SD < 1 mm; the one dipole with SD > 1 mm varied within 2.5 mm. The dipole localizations with RV < 2% corresponded to the epileptogenic zones identified on intracranial invasive video EEG and intraoperative ECoG. CONCLUSIONS The systematic approach of equivalent current dipole analysis using spike detection, clustering analysis, and an RV < 2% as a standard is useful for identifying extratemporal epileptic regions.

A neural network method for automatic and incremental learning applied to patient-dependent seizure detection.

OBJECTIVE Describe and evaluate a neural network method for automatic and incremental learning applied to patient-dependent seizure detection. Compare the classification ability of various time-frequency methods including FFT spectrogram, spectral edge frequency and bicoherence. METHODS 57 seizures from 10 epilepsy patients are used. A probabilistic neural network (PNN) is trained and incrementally updated in a novel fashion. The speed and accuracy of the method is evaluated with different training parameters and time-frequency methods. RESULTS Training the PNN on a single seizure from each record offers better performance (sensitivity = 0.89 and false-positive-rate = 0.56/h) than 3 patient-independent seizure detection algorithms. The method is virtually unaffected by the settings of various training parameters. Training is very fast (0.9 s), and the accuracy improves as more examples are added incrementally (without retraining). The overall best time-frequency method was the FFT spectrogram. The bicoherence plus the FFT spectrogram was the best method on 4 records, improving the correlation from 0.111 to 0.940 on one and from 0.288 to 0.612 on another. CONCLUSIONS The proposed method offers accurate, robust and virtually instantaneous training and incremental learning when applied to patient-dependent seizure detection. SIGNIFICANCE Accurate seizure detection can improve patient care in both the epilepsy monitoring unit and the intensive care unit. Future applications include patient-independent algorithms that continue to learn as new examples are encountered.

Spike detection: Inter-reader agreement and a statistical Turing test on a large data set

OBJECTIVE Compare the spike detection performance of three skilled humans and three computer algorithms. METHODS 40 prolonged EEGs, 35 containing reported spikes, were evaluated. Spikes and sharp waves were marked by the humans and algorithms. Pairwise sensitivity and false positive rates were calculated for each human-human and algorithm-human pair. Differences in human pairwise performance were calculated and compared to the range of algorithm versus human performance differences as a type of statistical Turing test. RESULTS 5474 individual spike events were marked by the humans. Mean, pairwise human sensitivities and false positive rates were 40.0%, 42.1%, and 51.5%, and 0.80, 0.97, and 1.99/min. Only the Persyst 13 (P13) algorithm was comparable to humans - 43.9% and 1.65/min. Evaluation of pairwise differences in sensitivity and false positive rate demonstrated that P13 met statistical noninferiority criteria compared to the humans. CONCLUSION Humans had only a fair level of agreement in spike marking. The P13 algorithm was statistically noninferior to the humans. SIGNIFICANCE This was the first time that a spike detection algorithm and humans performed similarly. The performance comparison methodology utilized here is generally applicable to problems in which skilled human performance is the desired standard and no external gold standard exists.