What Is Interictal Spike Detections Were Reviewed and Determined to Be Artifactural

• Loading metrics

Open Access

Peer-reviewed

Inquiry Article

A predictive epilepsy index based on probabilistic nomenclature of interictal spike waveforms

• Jesse A. Pfammatter,
• Rachel A. Bergstrom,
• Eli P. Wallace,
• Rama M. Maganti,
• Mathew V. Jones

x

• Published: November six, 2018
• https://doi.org/10.1371/periodical.pone.0207158
• See the preprint

Figures

Abstract

Quantification of interictal spikes in EEG may provide insight on epilepsy disease burden, but manual quantification of spikes is time-consuming and subject to bias. Nosotros present a probability-based, automatic method for the classification and quantification of interictal events, using EEG data from kainate- and saline-injected mice (C57BL/6J background) several weeks post-handling. We first detected high-aamplitude events, then projected upshot waveforms into Principal Components space and identified clusters of spike morphologies using a Gaussian Mixture Model. We calculated the odds-ratio of events from kainate- versus saline-treated mice within each cluster, converted these values to probability scores, P(kainate), and calculated an Hourly Epilepsy Index for each animate being by summing the probabilities for events where the cluster P(kainate) > 0.5 and dividing the resultant sum by the record duration. This Index is predictive of whether an animal received an epileptogenic treatment (i.e., kainate), even if a seizure was never observed. We applied this method to an out-of-sample dataset to assess epileptiform spike morphologies in five kainate mice monitored for ~1 month. The magnitude of the Index increased over fourth dimension in a subset of animals and revealed changes in the prevalence of epileptiform (P(kainate) > 0.5) spike morphologies. Importantly, in both information sets, animals that had electrographic seizures also had a high Index. This analysis is fast, unbiased, and provides information regarding the salience of spike morphologies for illness progression. Future refinement will allow a amend understanding of the definition of interictal spikes in quantitative and unambiguous terms.

Commendation: Pfammatter JA, Bergstrom RA, Wallace EP, Maganti RK, Jones MV (2018) A predictive epilepsy index based on probabilistic nomenclature of interictal spike waveforms. PLoS One 13(11): e0207158. https://doi.org/10.1371/journal.pone.0207158

Editor: Giuseppe Biagini, Academy of Modena and Reggio Emilia, Italy

Received: August 1, 2018; Accustomed: Oct 25, 2018; Published: November vi, 2018

Copyright: © 2018 Pfammatter et al. This is an open admission article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Data Availability: Data are available in the BioStudies database (http://www.ebi.air conditioning.britain/biostudies) under accession number Due south-BSST180.

Funding: This piece of work was funded in part by the National Institute for Health RO1 NS075366 (MVJ) and the Section of Defense PR161864 (RKM).

Competing interests: The authors take alleged that no competing interests exist.

Introduction

The electroencephalogram (EEG) is an essential tool for monitoring seizures, diagnosing epilepsy and understanding epileptogenesis. Most EEG analysis, including the scoring of seizures and interictal epileptiform events, is carried out manually by trained scorers. Transmission scoring is labor-intensive and decumbent to low interobserver agreement [1] which ofttimes precludes fine-scale and quantitative analysis of EEG, especially for chronic (>24 hours) recordings commonly employed in clinical and enquiry settings [2,3].

Adding to the claiming of analyzing EEG for epilepsy is that often convulsive seizure frequency is relatively low, oft much less than one/twenty-four hours, making capturing confirmatory seizure events on EEG rare [four]. Thus, the identification and analysis of subclinical, interictal events such as spikes may exist of import for diagnosis, identification of epileptic foci and understanding the progression of epilepsy [three–10]. But the definition of spike-like events for subsequent manual scoring remains an expanse of some disagreement among clinicians [2] and researchers [1,eleven]. Together with the laborious nature of transmission EEG analysis, this ways that the frequency of interictal spiking is not typically quantified for diagnostic purposes in the clinic [12], though there are special cases [13], and is underutilized in epilepsy research.

There are many automated approaches for detecting interictal events, including template-matching [14], not-linear energy operator (NEO) detection [15,16] wavelet filtering and line length [17], and neural networks [eighteen] amongst others [nineteen]. These algorithms rely in part upon a spike definition that is based on proficient visual analysis and are rooted in some version of the definition by Chatrian et al 1974, where a spike is a transient that is conspicuously distinguishable from background action, with a duration from twenty to seventy ms [20–22]. Still, a universal basis truth definition of an interictal fasten is defective every bit demonstrated in many cases where interobserver agreement may exist relatively low [1,23]. Further careful analysis and refined detection of spike and spike-like event morphology is needed to refine this definition and to improve the utility of spike analysis in epilepsy.

Studies in rodents have shown that spiking can exist identified inside the EEG from both normal and epileptic brains [2,4,24,25]. Thus, it is important to identify aspects of spike-similar morphologies that are specifically associated with epilepsy versus those that may be considered artifact related to EEG implantations, etc. Presumably, artifactual spikes may exist found in any animal while epileptiform spikes should only exist found in epileptic animals. Nevertheless, differentiation of the many spike types and shapes past visual or algorithmic analysis is a difficult prospect. Rigorous differentiation and quantification of spikes in epileptic versus normal animals could elucidate patterns of activity that shed light on spike-related mechanisms in epilepsy and identification of epileptic foci while likewise contributing to an unbiased definition of true interictal spiking activity across models of epilepsy. Given the links between interictal spiking and illness progression in epilepsy [26,27], quantitative understanding and differentiation between normal and epileptiform spike-similar activity is essential for time to come progress in epilepsy research; studies in animal models are essential to this progress.

Classification of spikes based on waveform morphology is a promising approach for differentiating events that are normal versus those that are relevant to epilepsy. Indeed, spike sorting past waveform clustering has been used to tune fasten-detection and quantification algorithms [28], to choose spike templates [29] and to identify or predict epileptic foci [30]. A very well-studied use of waveform clustering is to distinguish individual extracellularly recorded action potential waveforms in order to place neurons in single-unit recordings [31]. Several clustering approaches have been used beyond these applications, including K-ways clustering and affinity propagation [29], graph theoretical algorithms [30], and Gaussian Mixture Models (GMM) [31].

Hither nosotros present a novel waveform-based analysis to compare spike-like events recorded from mice treated with an epileptogenic insult (i.e., kainate, a model of temporal lobe epilepsy [32]) with spike-like events recorded from control mice treated with saline. Nosotros identified spike-like events in both groups, then projected these waveforms into a low-dimensional (3D) space using Main Components Analysis (PCA), and amassed these waveforms using a GMM. This allowed us to estimate the probability distributions that narrate normal versus epileptiform fasten-similar waveforms and to compute the probability that specific spike waveforms reflect epileptiform activeness. By assigning a probability score to each event type, nosotros provide a novel measure out (the Hourly Epilepsy Index) for estimating the probability that an animal is epileptic, even in the absence of observing electrographic or behavioral seizures. The ability to assess epilepsy with a loftier-resolution metric in every animal is extremely of import in animal models with depression seizure burden such as repeated depression-dose kainite in C57Bl6 mice (used in this study) and could result in the more efficient utilise of animals and EEG recording fourth dimension in future studies. This novel analysis thus provides rapid and unbiased insight into differences amid spike types in normal and epileptic animals that can exist used for early diagnosis of risk and serve as a biomarker of disease progression.

Materials and methods

Animals

All apply of animals in this manuscript conformed to the Guide for the Care and Use of Laboratory Animals [33] and was canonical past the Academy of Wisconsin-Madison Institutional Animal Care and Use Committee (IACUC, protocol M005629).

Animals (male C57BL/6J background, ~5 weeks old) were received from Harlan (Madison, WI) and housed in clear acrylic cages ~7 L cages with filtered airflow and corn cob bedding with paper nesting material. Animals were housed in groups of five or less with rodent chow and water available advertizement libitum. After ~2 weeks of acclimation, animals underwent repeated low-dose kainate (KA) or saline (SA) injections (~7 weeks quondam) and EEG implantation and recording (~18–20 weeks). All animals were checked daily for food/water/health past laboratory personnel and/or the University of Wisconsin-Madison Biomedical Research Model Services. Staff veterinaries were consulted in the event of affliction (post-injection) or severe injury (every bit a result of fighting mail service-injection) and their recommendations for animal care were followed.

The data used in this study represents a portion of information collected as a result of an additional study (IACUC canonical) investigating pneumonic bigotry during behavioral trials and computational design separation in slice (In Preparation).

Epilepsy consecration

During the injection process, when not handled, mice were individually housed in enclosed ~150 cm³ acrylic cubicles with opaque sides and clear forepart portals with holes to allow air exchange, and equipped with corn cob bedding and rodent grub. Animals were so randomly assigned to SA or KA treatment, ear punched or tagged for identification and weighed. Mice received a serial of intraperitoneal (IP) injections based on the following schedule (herein referred to as "repeated low-dose kainate"). Mice in the KA group first received a x mg/kg (5.5 mM) dose of kainate (Tocris Bioscience, UK) delivered in 1x phosphate-buffered saline (PBS) prepared from PBS tablets (Dot Scientific, Michigan) dissolved in deionized distilled water (DDW) and filter sterilized (0.22 μm, Corning, MI). Mice continued to receive injections at 2.5–5 mg/kg every twenty minutes until status epilepticus (SE) occurred (4–ix injections) [34,35]. Mice in the SA group received the same schedule of sterile PBS injections. Mice were considered to be in SE when displaying behavioral seizures of level 4–v on the Racine Calibration [36] for a minimum of 30 minutes. Mice usually experienced SE for 90–180 minutes and were allowed to recover spontaneously. Mice were treated in cohorts of ~8 at a time, with a ratio of ~2:1, KA:SA animals. The outset cohort of mice received alternating 5 mg/kg and 2.5 mg/kg injections after the initial 10 mg /kg dose, but this schedule took longer than merely v mg/kg injections without any noticeable differences in eventual efficacy or survival, and then subsequent cohorts received five mg/kg injections. During chemoconvulsant assistants, if an animal appears extremely distressed and immobile such that it enters stage three status epilepticus (continual ictal behavior) or remain in condition at the 4 hr time indicate, it was euthanized. Throughout the study, poor body status (due east.g. fighting wounds or dehydration), immobility, and hunched posture were too criteria for the consideration of euthanasia. Under these circumstances a vet was consulted. Nosotros euthanized one animal due to dehydration and hunched posture three days subsequently kainite injection. I animate being died during the kainite injection procedure as a result of status epilepticus. No other mortality occurred during this study.

Subsequently the injection schedule, animals were given fresh apple slices, monitored until SE ceased (assessed by normal posturing), and returned to group housing. They were monitored, weighed, and given 0.4 mL of sterile 1x PBS via IP injection on a daily ground equally necessary until their weight returned to pre-injection level. No fauna injected with SA ever experienced condition epilepticus, lost weight, or required recovery injections of 1x PBS. Mice remained in the vivarium for seven–9 weeks afterward injections, allowing time for the evolution of epilepsy in KA mice. During this time some KA mice became aggressive and were removed from group housing and caged individually.

EEG surgery and recording

Animals were implanted with EEG headcaps while under ane–ii% vaporized isoflurane anesthesia (see [37] for detailed methods) at ~eighteen–20 weeks old (~xi–13 weeks post injection), immune to recover for 3–four days, and then recorded continuously at 256, 512, or 1024 Hz sampling rate for three days. All information not collected at 512 Hz were either resampled (1024 Hz) or linearly interpolated (256 Hz) to 512 Hz for analysis. Information were so notch filtered at sixty Hz using a Chebyshev Type 2 digital filter and high-pass (>0.v Hz) filtered with a Chebyshev Type I digital filter. All analyses presented in this manuscript used a right-frontal EEG channel. All determinations of electrographic seizures in this manuscript are based solely on EEG and EMG traces.

High-amplitude result detection

To assess the severity of epilepsy in animals injected with kainate, we developed a probability-based approach for the detection of epilepsy-related events (e.g. interictal abnormal events like EEG spikes, seizures) and the calculation of an epilepsy severity index.

Filtered EEG signals were normalized to using a variation of z-score normalization that we term Gaussian normalization. Kickoff, nosotros fit a Gaussian distribution to the all-points histogram (calculated using the histfit() function with bins equal to the flooring of the square-root of the number of information points in each EEG signal) of each 24-hour record using least-squares minimization with Nelder-Mead Simplex [38] optimization implemented by the fminsearch() part (with 'fit_gauss' as an option) in Matlab. We so normalized the 24-hour record by subtracting the mean from each bespeak and dividing by the standard difference of the model fit. The resultant normalization is unlike from the standard z-score normalization in that the hateful and standard deviation components are calculated from a model estimating the Gaussian components of the EEG rather than empirically calculating the mean and standard difference. We employed this normalization because it adjusts the data based on the normal component of the point and is less influenced by very large and brief artifacts.

Using normalized EEG we detected high-aamplitude, spike-like events using a two-threshold strategy [39]. The final thresholds used in this assay were optimized to yield the largest predictive upshot size (Fig 1D). The offset of an effect was triggered when the EEG signal rose v standard deviations above the mean of the Gaussian fit. An event ended when the signal passed 1 standard difference below the mean. We then combined any events occurring within 0.2 seconds of one another (end to starting time) into a single event [21]. This resulted in a list of events of varying duration (98% shorter than 1s, with some representing rapid fasten bursts) from each animal/day EEG record. We and then took 2 seconds of point centered on the offset of each upshot and collated events beyond 72 days of recording (iii–4 days from each of 15 KA and 7 SA animals). This resulted in a matrix with ~33000 upshot waveforms (1024 points each) beyond all treatments, recording days, and animals. We chose optimal high (consequence beginning) and low (upshot stop) threshold values afterward a grid search to investigate high values between 3 and 8, and depression values between -3 and 2 and then performing a t-test to compare the number of events per hour identified in SA vs. KA animals (Fig 1D).

Fig ane. Effect detection by two-threshold approach.

Sample traces of events identified (red) using the two-threshold method (high threshold = v SD above Gaussian mean, depression threshold = ane SD beneath mean) in (A) SA- and (B) KA-injected animals. Each row in A and B shows an expanded view (indicated by the bluish lines) of the point straight above. (C) Boilerplate number of events per hour detected in SA (black) versus KA (reddish) animals (P = 0.166, t = 1.44, df = 19.83, two-sample T-test with unequal variances). Circles correspond the average number of events per hour for each creature over 3 days of recording. Airtight circles represent animals that were recorded having an electrographic seizure. (D) P-values comparison the average number of events per hour in SA and KA animals beyond a set of high and depression thresholds; the p-value is minimized at 5 and -1. The center mark of the box and whisker plot represents the median, while the lower and upper bounds of the box represent the 25^th and 75^th quartiles, respectively; the whiskers extend to the most extreme information points or one.5x the interquartile range.

https://doi.org/10.1371/journal.pone.0207158.g001

Epileptiform outcome confirmation

We compared the output of the two-threshold detection method with output from some other method that has been highly cited for seizure and interictal spike detection, and that also performs similarly to manual man scoring [17]. Spikes, seizures, and other abnormal epileptiform activity were detected in normalized EEG according to Bergstrom et al. [2013]. The baseline was determined using the whole bespeak. The effect threshold was set up at iv.0, where spikes were defined using a modified aamplitude of 15, and seizures were defined as events greater than 1 2nd in duration. To compare the two-threshold method with that of Bergstrom et al. [2013], all records were represented every bit logical vectors where 1 indicates the presence of an event during each 2d of recording and 0 indicates the absence of an result. Confusion matrices [40] betwixt these vectors were used to compute truthful-positive and false-negative rates, by provisionally assuming that Bergstrom et al. [2013] detected "true" events.

PCA, clustering, and P(KA)

Nosotros performed Main Components Assay (PCA) [41,42] on the matrix (~33000 x 1024) of waveforms detected by the two-threshold method and projected each original waveform into the infinite spanned by the first three Master Components (PCs). In such 3D projections, each point represents a detected waveform and the distance between any two points represents the deviation betwixt their original waveform shapes. Three dimensions were chosen for ease of brandish, simply higher or lower dimensional projections could exist used throughout this analysis in principle.

Waveforms that are closely intertwined with epilepsy should be a) more common in KA than in SA and b) like to other waveforms in epileptic (KA) mice and dissimilar from those of SA mice. Therefore, the densities of PCA-projected points for KA and SA events are expected to overlap somewhat but, by examining quantitatively their overlap in specific regions of PCA space, we can compute the likelihood of a waveform in any region having come from a KA vs SA creature.

The resulting ensemble of points was fit with a Gaussian Mixture Model (GMM) with expectation maximization using the function fitgmdist() [43] with nine components and and so clustered based on the GMM coefficients. The GMM defines a probability for every betoken in the space, and in this case reflects the prior probability distribution that any waveform would be detected because both KA and SA were included in the GMM fit. We then calculated the conditional probability, P(KA) that events in each cluster are related to an epileptogenic handling past calculating the ratio of KA (epileptogenic) to SA (control) events in each cluster (Odds ratio) and converting it to a probability, P, using .

Hourly Epilepsy Index

Given that there is some disagreement between manual scorers, we are attempting to provide a statistical and unbiased measure out of epileptiform event frequencies. We commencement defined epileptiform event morphologies equally events with a P(KA) > 0.fifty, because they are more than probable to be establish in KA than SA animals and are therefore more than relevant to epilepsy. The range of P(KA) values was rescaled from 0.v–1.0 to 0–1, where scaled P(KA) = (P(KA) − 0.5) × 2 to reverberate the relative contribution of each event morphology to the KA-specific ready. Our method is probabilistic, so it can only be interpreted as an "index" rather than an actual count of interictal events. Therefore, we define the Hourly Epilepsy Index = ∑(freq.events in each cluster with P(KA) > 0.v × associated scaled P(KA)). Theoretically, the possible Hourly Epilepsy Index for any given EEG recording ranges between 0 and the total number of loftier amplitude events detected from that record.

Parameter optimization

I tunable parameter in this analysis is the pick of how many clusters to utilise in modeling the data (i.due east., how fine-grained should our analysis of waveform shapes be?). To investigate this parameter, we built a fix of models varying the full number of GMM clusters. Additionally, because creation of a GMM in our implementation is a stochastic process, we rebuilt each model at 25 unique random seeds (seeds = 1 through 25). In total, nosotros generated 600 Gaussian mixture models and related cluster profiles, P(KA)s, and Hourly Epilepsy Indices. From each of these model sets we calculated: 1) the Negative Loglikelihood of the information fit to the GMM, 2) The boilerplate belongingness of each event to its assigned cluster (Cluster Belongingness, divers below), 3) the outcome size between SA and KA treated animals and iv) P value comparing the Hourly Epilepsy Index between SA and KA treated animals. Cluster Belongingness was calculated past taking the sum of squares of the set of posterior probabilities, calculated with the posterior() function in Matlab (which calculates posterior probabilities for Gaussian mixture components) for each event and then averaged across all events. For each event, these posterior probabilities (every bit many probabilities every bit in that location are clusters) provide the likelihood of belonging to each cluster in the model and these probabilities sum to one. For example, in the case of a 2 cluster model, data signal A may belong every bit to each cluster P = (0.5, 0.5) where its Cluster Belongingness equals 0.v. In dissimilarity, information point A may belong much more strongly to i cluster over another P = (0.9, 0.1) with a cluster belongingness of 0.82. Thus, a clustering model with poor fit will have a low value for Cluster Belongingness.

Computer programming

All analyses, data processing, and event classification was done in Matlab (Mathworks, Natick, MA) using domicile-written scripts. The final version of this algorithm was executed with Matlab 2017b.

Statistical analysis

The usage of statistical tests and procedures used to estimate model fit are indicated throughout the methods section of this manuscript. All tests and statistical applications were called after careful consideration of data distribution (using the Kolmogorov-Smirnov (KS) tests in addition to Quantile-Quantile plots to assess normality and visual assessment of variance with boxplots) and independence. In the consequence that data were considered non-usually distributed according to the KS test at alpha = 0.05, a non-parametric Wilcoxon Rank Sum (using ranksum()) examination was performed in place of a two-sample t-test. All statistical tests were two-tailed. The researchers were non blinded or randomized to the treatments given to animals, but rather treatments (KA and SA) known to cause the evolution of epilepsy [34,35] were used to calculate probabilistic information that went into creating the Hourly Epilepsy Index. Indeed, the design of this approach was to optimize for differences betwixt the treatments. However, nosotros feel that this approach is warranted and importantly, the clustering of the events morphologies was unsupervised and thus 'blinded' to the handling from which each result originated.

Results

EEG recording and event detection

We collected iii days of continuous, 24-60 minutes EEG for 15 KA and 7 SA animals and normalized the records as described above. The two-threshold effect detection algorithm identified 25681 and 5764 events in KA- and SA-treated animals, respectively (Fig i; hateful ± SD event duration for KA animals = 55 ± 148 ms; for SA animals = 169 ± 380 ms). The average number of events per hour detected in KA (23.three ± 27.four) and SA (xi.five ± 11.0) animals using this detection method was not significantly unlike (Fig 1C, P = 0.166, t = one.44, df = 19.83, two-sample T-test with unequal variances). The p-value for comparison of KA and SA animals was minimized by the utilise of the loftier threshold of +5 and low threshold of -one, though combinations of loftier thresholds of 4 or five and low thresholds of 1 to -2 showed low p-values relative to all other combinations (Fig 1D).

High-amplitude events are epileptiform

To confirm that the events identified by the ii-threshold method include epileptiform events, the normalized EEG signal was analyzed using an event detection and sorting algorithm that is validated with human fasten scoring [17]. A comparison between events detected by the two-threshold method and spikes from Bergstrom et al. [2013] resulted in a true positive rate (TPR) of 0.781 and a imitation positive charge per unit (FPR) of 0.005 (Fig 2A and 2C). Thus, our initial detection of spikes agrees reasonably well with a previous method that itself agrees well with man scoring. In a comparison of the two-threshold detection method to all epileptiform events (spikes, seizures, and other low-amplitude, short duration abnormal) from Bergstrom et al. [2013], the understanding between the algorithms drops precipitously (Fig 2B and 2D, TPR and FPR for all events are 0.038 and 0.001, respectively), suggesting that the 2-threshold method is selective for spike-blazon events and does not detect low-aamplitude abnormal or some seizure-like events. We illustrate specific examples of agreement (Fig 2E) and disagreement (Fig 2F) between the two detection methods, highlighting that the two-threshold detection method selects for loftier-amplitude fasten- and seizure-like activity.

Fig 2. Interictal events correspond to epileptiform activity.

We compared events from the two-threshold crossing method to events detected past a validated outcome detection and nomenclature algorithm [17]. Second-by-second, fauna-specific (SA = black, KA = scarlet) event classification agreement (Truthful positive rate, TPR) and the rate of detection of additional events non identified in the validated method (Additional events, faux positive rate, FPR) for spike-type events only (A) and all event types (seizures, spikes, and other low-amplitude abnormal and epileptiform events, (B), with corresponding confusion matrices (spikes only C, all events D). True positives are events detected by both the validated algorithm and the ii-threshold crossing method. (E) Sample true positive events for each consequence blazon, where gray is the original EEG trace, green is two-threshold detection output, and black is Bergstrom et al. [2013] detection output. (F) Sample events for each detection method that were not detected past the other method.

https://doi.org/10.1371/journal.pone.0207158.g002

Measures predictive of an epileptogenic insult

Events from KA- and SA-treated animals were projected into Principal Components infinite (Fig 3A, first three PCs shown). We used the outset iii PC (76.7% of the overall variance) and clustered the data into ix groups using a GMM (Fig 3B). The probability, P(KA), that a cluster morphology is relevant to epilepsy is equal to the probability of that morphology being found in KA but not SA records and was calculated as the odds ratio of the number of KA to SA events for each cluster. The P(KA) for four of ix clusters is in a higher place 0.5 (Fig 3C, range 0.73–0.99), indicating that issue morphologies in these clusters may be specific to the KA treatment and thus related to epilepsy. On average, KA animals exhibit a significantly higher boilerplate Hourly Epilepsy Index than do SA animals (11.half-dozen ± xiv.half dozen for KA, 3.0 ± two.0 for SA; P = 0.039, t = ii.26, df = xv.04, two-sample t-examination with unequal variances). Two of fifteen KA animals displayed electrographic seizure activity in at least one recording (Fig 3D, magenta filled circles); the average Hourly Epilepsy Index for these animals falls above the 75th percentile (top line of box plot) of the SA hourly alphabetize. Nosotros found that there was a wide spread of Hourly Epilepsy Indices beyond the 17 KA mice, where eight KA mice had Hourly Epilepsy Indices similar to SA mice and 9 KA mice had Hourly Epilepsy Indices to a higher place the 95% confidence interval of the SA Alphabetize (95% confidence interval for SA = i.14–4.47). This interesting result is in line with published reports of varying sensitivity to systemic KA injection in epileptogenesis [34,35].

Fig 3. Event sorting by clustering and Hourly Epilepsy Index.

(A) Events detected with the two-threshold crossing method, from all animals and records, and projected into Chief Components space (first three PCs shown, red = KA, blackness = SA). (B) Events in A amassed into ix groups using a GMM with Expectation Maximization. The upper inset shows a rotated and expanded view of the aforementioned plot, allowing a view of cluster 5 (embedded within cluster 6, which is prepare to x% transparency for visualization). (C) P(KA) for each cluster identified in B calculated from the relative proportion of KA to SA events, presented in descending club with examples of events. (D) The boilerplate Hourly Epilepsy Index (fifteen KA, 7 SA animals) is much higher for animals treated with KA (P = 0.039, t = 2.26, df = 15.04, two-sample T-test with unequal variances). The heart mark of the box and whisker plot represents the median, while the lower and upper bounds of the box represent the 25^th and 75^th quartiles, respectively; the whiskers extend to the nigh extreme data points or 1.5x the interquartile range.

https://doi.org/10.1371/periodical.pone.0207158.g003

Optimization of the clustering algorithm

To make up one's mind the optimum number of clusters to differentiate between KA- and SA-specific fasten morphologies, we calculated 600 repetitions of the Gaussian Mixture Model used to cluster events at each unique value of 2 to 25 clusters and random seeds set from i to 25. The Negative Loglikelihood (Fig 4A) approaches an asymptote effectually eight to ten clusters, suggesting the reasonable model fit in this range. Cluster Belongingness (Fig 4B) decreases across the simulation range with no clear inflection point (i.e. elbow) in the information to indicate diminishing returns of increasing cluster number. Consequence size and p-value for the Hourly Epilepsy Index (Fig 4C and 4D) models signal that the effect size and the p-value reach an asymptote in a range similar to that indicated by the Negative Loglikelihood results. Here, we accept a fuzzy problem in terms of cluster selection. We selected nine clusters for our purposes and hash out cluster option in depth in the word section.

Fig 4. Selecting an appropriate number of clusters.

We repeatedly generated a set of models (n = 25, random seeds prepare from ane:25) specifying a range of 2 to 25 clusters (totaling 600 models) and calculated the (A) Negative Loglikelihood, (B) Cluster Belongingness, the (C) Effect size and (D) P value of the Hourly KA Index comparison KA to SA treated animals (calculated using merely events where P(KA) > 0.five). These values offer guidelines for model selection via 'elbows' in the data which are indicative of diminishing returns in model fit. It does not announced that there is a clear best value for clustering these data into detached groups. Therefore, the cluster number must exist selected and evaluated based on the goal of the cluster assay.

https://doi.org/x.1371/journal.pone.0207158.g004

Cluster morphologies match published characteristics of interictal spikes

We performed a preliminary waveform shape analysis by plotting individual event elapsing vs aamplitude for each cluster. In clusters where P(KA) > 0.five (clusters one through four, Fig 5A) many events from KA-treated animals have a college aamplitude than do events from the aforementioned cluster institute in SA-treated animals. The bulk (92%) of events in these clusters for both treatment groups are shorter than 200 ms, consistent with published spike definitions. This trend is non every bit distinct in clusters five through nine, where P(KA) < 0.5. Farther inspection of events in clusters one through four reveals that the morphologies of these upshot types are more than classically "spikey" (brusk-duration events with precipitous waveforms) than are events found in the remaining clusters. Finally, the majority of the events identified in KA animals (88%) were constitute in clusters one through 4, whereas events in SA animals were equally distributed among the clusters with P(KA) in a higher place (49%) and beneath 0.five. The boilerplate number of events per hour for each cluster in SA and KA shows separation between outcome frequency in clusters one through four but not clusters 5 through 9 (Fig 5B, p-values reported on each graph).

Fig v. Characterization of individual clusters and contribution to total event count.

(A) Graphical distribution of each cluster'southward events in EEG amplitude (x-axis) past upshot elapsing (y-centrality, time (s)) for events belonging to both KA (red) and SA (black). Example events from each cluster type are displayed in the upper right corner of each plot. For improved visualization of the majority of data, nosotros have set up the axis to exclude a small amount of the most extreme data points (nigh notably from cluster 7). (B) The average number of events per hour within each cluster for SA (black) and KA (crimson) animals. P-values indicate the results of t-tests or rank sum tests betwixt KA and SA for each cluster of waveform shapes.

https://doi.org/10.1371/journal.pone.0207158.g005

Analysis of out-of-sample data supports the methodology

We applied the clustering approach to events detected in a novel dataset of longitudinal recordings from five KA-treated animals (Fig 6A–6D, recorded on days 30–34 and 44–51 postal service-KA). The longitudinal data were projected into the same PC space as the SA and KA dataset from Fig 3A and clustered according to the same parameters as Fig 3 (Fig 6A and 6B). All but i animate being (Animal Eastward) experienced similar, only not identical, increases over time in total upshot charge per unit (average for all events per hour, Fig 6C) and Hourly Epilepsy Alphabetize (clusters i through iv, Fig 6D). Electrographic seizures were only detected in one animate being (Animal B) at days 44 and 45; this animal displayed a trend toward increasing event rate and Hourly Epilepsy Index (Fig 6C–6E, magenta-filled circles).

Fig 6. Detection and categorization of interictal events in longitudinal data.

We applied our detection and categorization algorithm to longitudinal EEG records (recorded days 30–34 and 44–51 post-KA) from 5 KA mice. (A) Events identified from these v animals (majestic circles) projected into PCs with data from Fig iii (black circles, SA and KA data). (B) Clustering of these new data points using the same GMM coefficients and clustering parameters used in Fig 3 (cluster colors correspond to those throughout manuscript). The upper inset shows a rotated and expanded view of the same plot, assuasive a view of cluster v (embedded within cluster 6, which is set to x% transparency for visualization). (C) Events per hr equally detected by the two-threshold method in each of the 5 KA animals over time. (D) Hourly KA Alphabetize for longitudinal data set. E) Average events per 60 minutes as detected by the two-threshold method inside each cluster group. In panels C-E, records with electrographic seizures are shown every bit filled in circles. An example seizure is shown in the inset of panel D.

https://doi.org/ten.1371/journal.pone.0207158.g006

Discussion

Interictal spikes are a authentication of the epileptic brain [ii]. Yet the definition of an interictal spike remains somewhat nebulous, potentially biasing spike-detection strategies and algorithms. Previous analyses of spikes in kainate and other models of epilepsy have relied upon scoring of spikes and other interictal events to differentiate between KA-treated and control animals, where more spikes are observed in epileptic than in control animals [four,17]. In these studies, control animals likewise exhibit fasten-type events in the EEG; however, given that control (and many KA animals) animals practice non proceed to take seizures, information technology is probable that some spikes arise past a non-epileptic machinery, including spike-like events in normal brain (eastward.g., sharp moving ridge ripples [2], electrographic noise, artifact, or damage associated with electrode implantation. That not-epileptiform spikes may likewise exist present in both normal and kainate-treated animals further complicates the quantification of interictal spikes and thus the interpretation of nonconvulsive EEG events in epilepsy and epileptogenesis.

Hither nosotros nowadays a novel automatic algorithm for probability-based differentiation of epileptiform versus non-epileptiform spike and interictal EEG event morphologies in the kainate model of TLE. Our algorithm detects interictal events and weights them in an Hourly Epilepsy Alphabetize based on the likelihood that they are, indeed, relevant to the injection of kainic acid and thus TLE (i.e., related to an epileptogenic insult [32]). Interictal events and seizures projected into PC space were clustered using a Gaussian Mixture Model. We used the probability P(KA) that an event cluster was KA-specific to calculate an Hourly Epilepsy Index as a weighted boilerplate of the number of events with a P(KA) > 0.5. This value successfully estimates the probability that a subject had received an epileptogenic insult, in this case a KA injection. This algorithm can, in principle, be mostly applied to chronic EEG recordings from either experimental animals or human patients, if a control data ready is bachelor for comparison. It can also provide rapid and quantitative scoring of interictal events to predict the likelihood that a bailiwick is at risk for developing spontaneous seizures. Importantly, the awarding of this algorithm to EEG recordings from animal models of epilepsy with low seizure brunt (e.g. repeated low-dose kainate) may result in a reduction in the number of animals sacrificed in future studies.

High-aamplitude event detection is an inclusive starting betoken

Our approach to event detection using a two-threshold algorithm selected events of high aamplitude compared to background betoken, but did not, at the outset, impose limitations on effect elapsing or morphology. This method is therefore intentionally inclusive of spike and other high-amplitude, non-spike-type events that may be physiologically relevant to epilepsy. Upon visual inspection of the signal and in comparison to a validated seizure- and spike-detection algorithm, we found that this method preferentially detects spike-type and seizure-like activity in the EEG signal (Figs ane and 2), as well equally other event morphologies, including artifact (Fig 3, cluster 8). That our detection method primarily selects for high-amplitude events does not negate the possibility that low-amplitude interictal events are relevant to epilepsy and epileptogenesis. Detection and analysis of low amplitude epileptiform events is an area for farther inquiry [17].

Selecting the appropriate number of clusters is flexible and awarding-dependent

We built on prior clustering approaches for EEG by considering how spike sorting may elucidate fasten morphologies relevant to epilepsy and to differentiate these morphologies from non-epileptiform spike-like events observed in normal rodent brains [2,iv]. The high-amplitude events identified in records from SA- and KA-treated animals were projected into PCA space and then clustered using a GMM with expectation maximization. Similar to spike sorting in extracellular recordings [31], we establish that the GMM provided a strong fit for the clustering of high-amplitude events in our dataset.

Nine clusters provided adequate separation between SA and KA-treated animals, though our analyses as well found that this is, indeed, a fuzzy problem with many different options for cluster number selection depending on the features to be extracted. Spike morphology in extracellular recording is reflective of the geometry and physiological properties of private neurons. Spike morphology in EEG may be reflective of a number of unlike parameters, including the underlying pathology of the spike, the spike location, the recording montage/referencing, private patient and recording characteristics, and other features. Defining the number of clusters therefore requires careful consideration to forestall over- or nether-estimation of the true number of spike morphologies.

We found that Cluster Belongingness and p-values for the separation betwixt KA vs not-KA consequence morphologies ameliorate as the number of clusters increase. However, very higher cluster numbers may exist sectionalisation spike morphologies (clusters) based on individual animal or recording characteristics, rather than differentiating between KA and non-KA events. Selecting a smaller number (e.k. iii rather than ix) of clusters for analysis does not provide statistically significant separation between KA and SA animals, and then information technology is important to consider the tradeoffs in increasing and decreasing the cluster number for analysis (Fig iv). Through modeling using many different cluster numbers (from two to 25), we find that there is no clear "elbow" in the Negative Loglikelihood or Cluster Belongingness scores across our simulations. This indicates that there is non a hard and fast cluster number for this assay. Yet, we found that the p-value and upshot size for the epilepsy index, and the Negative Loglikelihood values seemingly approached an asymptote past nine clusters (i.eastward. no major improvements at higher cluster numbers). Thus, nosotros selected ix clusters for our analysis because information technology effectively separated KA from SA events while minimizing animal-specific effects.

Nosotros further explored upshot morphologies of private clusters by specifically focusing on upshot amplitude and duration (Fig v), as instructed by the descriptive fasten definition of Chatrian et al. [1974]. Events with a P(KA) > 0.5 (clusters one–four, more than probable to be establish in KA than SA animals) exist inside a space consistent with general definitions of spikes, that is events with high aamplitude and brusk duration (the far lower left portion of the graphs in Fig 5A). Every bit the P(KA) drops beneath 0.five (clusters 5 through ix), the events brainstorm to shift toward longer, lower-amplitude characteristics. It must be noted, nonetheless, that there is non a sharp demarcation in the amplitude and duration characteristics of spikes above and below the cutoff of P(KA) = 0.5.

In applying this analysis to other models of epilepsy, we see ii major approaches to selecting an appropriate number of clusters, dependent upon the end goal of the analysis. One approach, similar to the approach that nosotros have taken, is to identify the minimum number of clusters that provide significant separation betwixt experimental weather or a minimum cluster number that yields clusters of visually similar morphologies. In add-on to differentiating betwixt normal and epileptiform activeness, this approach should ideally separate noise and antiquity as we see in cluster eight (Fig iii). Quantification of events could thus proceed relatively noise-free. This is similar to only more specific than an arroyo taken by (28), where spike morphologies in clusters that contain less than v% of the total fasten count are removed from analysis. The 2d arroyo is to maximize the number of clusters, finely parcellating event into clusters for a college-resolution model of epilepsy. Intendance must be taken, however, not to parcellate the data likewise far, resulting in over-plumbing fixtures of the model.

The Hourly Epilepsy Index: What it is, what information technology is not

We have introduced an Hourly Epilepsy Alphabetize as a summary descriptive variable of the likelihood of epilepsy in an private animal, calculated as an boilerplate hourly event rate weighted based on the probability that an effect morphology is plant in KA- but not SA-treated animals. Given that kainate injection is merely one model of epilepsy, we cannot still claim that this is a true model that is applicable to all epilepsy models or etiologies. Yet, this index does begin to address some of the challenges of rodent models of epilepsy, namely that not all mice that receive KA-treatment will go along to develop epilepsy [32] and low seizure frequency may make it difficult to identify all epileptic mice, by providing relevant probability-based estimates of likelihood of an animal being epileptic. Previous studies have demonstrated that spike frequency during the latent period and in interictal periods may be correlated with, or predictive of, future development of seizures [2, 4]. These studies have non differentiated between epileptiform and non-epileptiform spiking, thus they exercise not fully narrate the contribution of truly epileptiform spiking to seizure evolution and epileptogenesis. Others [32] have also shown that some animals accept different susceptibility to epilepsy in the KA model of TLE; resistant animals may be expected to have a different epileptic spike frequency during the latent catamenia than animals that will develop epilepsy. The Hourly Epilepsy Index may thus be almost useful as a mensurate during the latent period to predict the likelihood of developing epilepsy. This is a possibility that remains to be further investigated via longitudinal studies and using video EEG and an Hourly Epilpesy Alphabetize-based assay of the latent period through the development of spontaneous seizures.

There are several features of the Hourly Epilepsy Alphabetize that indicate that this may be a valuable model by which to analyze EEG data in epilepsy studies. At its simplest level, the Hourly Epilepsy Index provides a more than robust separation between KA and SA animals than does raw event count (Fig 1C vs Fig 3D) by excluding events that are not likely to be epileptiform (events where P(KA) < 0.5) and weighting the remaining events (where P(KA) > 0.v) according to the probability that they are, indeed, epileptiform. The relevance of this new metric is further supported by data that reveals that the Hourly Epilepsy Index is significantly college for the few KA animals with electrographic seizures than the average epilepsy index of SA-treated animals (Fig 3D). In further confirmation of this, we did not observe electrographic seizure activity in SA animals or KA-treated animals that autumn inside the 95% CI range of the Hourly Epilepsy Alphabetize for SA animals. While we take shown that animals with electrographic seizures do have a higher Hourly Epilepsy Index and that individual spike morphologies change over time in a manner seemingly related to epileptogenesis, we have not, in this present study, fully assessed the relative contribution of each spike waveform morphology to the evolution of epilepsy. We plan further investigation of the Hourly Epilepsy Index and the contribution of specific spike effect morphologies to epileptogenesis past monitoring mice straight after SE through the development of spontaneous seizures and beyond.

Cluster ane (Figs 3C and 5A) is of detail interest in considering the power of spike sorting by clustering for epilepsy assay. These events are relatively infrequent across records, only are highly correlated with KA handling (P(KA) = 0.99) and are very rare in SA animals. Visual inspection of these events confirms that they are, indeed, spike-like, fitting into a more classical definition of an interictal spike [20]. Because these event types are almost exclusively seen in records from KA-treated animals, they are weighted more strongly in the epilepsy index than are events from cluster four, which show features of spiking (loftier amplitude, fast elapsing) just may also cause an proficient scorer to pause on their definitive identity as an interictal fasten.

The Hourly Epilepsy Alphabetize is not without caveats, however. Kainate is a widely-used model of TLE and epileptogenesis [32], but it is only ane of many models. Other models, including pilocarpine, traumatic encephalon injury (TBI), genetic or viral models of epilepsy, and kindling, rely upon different molecular mechanisms of seizure consecration and may therefore produce interictal events of dissimilar morphologies. Ultimately, this approach may be used to produce a library of fasten templates across epilepsy models to further expand and refine the working definition of an interictal spike.

Decision

Clustering of interictal spike provides insight into event morphologies relevant to epilepsy and epileptogenesis. Using this clustering method, we isolated epiliptiform event morphologies that were and so quantified to decide the probability that a item animal is at risk for epilepsy or not, a variable which nosotros have called the Hourly Epilepsy Index. This probabilistic nomenclature method for interictal issue waveforms may provide a biomarker for hazard and development of epilepsy, fifty-fifty in the absenteeism of observing electrographic or behavioral seizures. By distinguishing spike morphologies that are preferentially present in the epileptic condition, we may further investigate the contribution of different spike morphologies to the evolution of spontaneous seizures, thereby contributing to an unbiased understanding of the definition of interictal spiking in epilepsy and epileptogenesis.

Supporting data

Acknowledgments

We would similar to thank Antoine Madar, Rebecca Willet, and Jun Zhu for helpful conversations during algorithm development. We too give thanks the two reviewers who provided valuable feedback which improved this manuscript.

References

1. ane. Webber WR, Litt B, Lesser RP, Fisher RS, Bankman I. Automated EEG spike detection: what should the computer imitate? Electroencephalogr Clin Neurophysiol. 1993 Dec i;87(6):364–73. pmid:7508368
   • View Commodity
   • PubMed/NCBI
   • Google Scholar
2. 2. Fisher RS, Scharfman HE, deCurtis Thousand. How can we identify ictal and interictal abnormal activity? Adv Exp Med Biol. 2014;813:3–23. pmid:25012363
   • View Article
   • PubMed/NCBI
   • Google Scholar
3. 3. Boly 1000, Maganti R. Monitoring epilepsy in the intensive care unit: current land of facts and potential interest of loftier density EEG. Brain Injury. 2014 Aug ane;28(9):1151–v. pmid:25099019
   • View Article
   • PubMed/NCBI
   • Google Scholar
4. four. White A, Williams PA, Hellier JL, Clark S, Dudek FE, Staley KJ. EEG spike action precedes epilepsy later kainate-induced status epilepticus. Epilepsia. 2010 Mar;51(3):371–83. pmid:19845739
   • View Commodity
   • PubMed/NCBI
   • Google Scholar
5. v. Barkmeier DT, Senador D, Leclercq Thou, Pai D, Hua J, Boutros NN, et al. Electrical, molecular and behavioral furnishings of interictal spiking in the rat. Neurobiol Dis. 2012 Jul;47(1):92–101. pmid:22472188
   • View Article
   • PubMed/NCBI
   • Google Scholar
6. 6. Cascino GD, Kelly PJ, Sharbrough FW, Hulihan JF, Hirschorn KA, Trenerry MR. Long-term follow-up of stereotactic lesionectomy in partial epilepsy: predictive factors and electroencephalographic results. Epilepsia. 1992 Jul;33(four):639–44. pmid:1628577
   • View Commodity
   • PubMed/NCBI
   • Google Scholar
7. 7. Krendl R, Lurger S, Baumgartner C. Absolute spike frequency predicts surgical event in TLE with unilateral hippocampal atrophy. Neurology. 2008 Aug 5;71(6):413–8. pmid:18614768
   • View Article
   • PubMed/NCBI
   • Google Scholar
8. 8. Miller JW, Gotman J. The meaning of interictal spikes in temporal lobe epilepsy: Should we count them? Neurology. 2008 Aug 5;71(6):392–iii. pmid:18678822
   • View Article
   • PubMed/NCBI
   • Google Scholar
9. nine. Avoli M, de Curtis M, Köhling R. Does interictal synchronization influence ictogenesis? Neuropharmacology. 2013 Jun;69:37–44. pmid:22776544
   • View Article
   • PubMed/NCBI
   • Google Scholar
10. 10. Tomlinson SB, Bermudez C, Conley C, Chocolate-brown MW, Porter BE, Marsh ED. Spatiotemporal mapping of interictal spike propagation: a novel methodology practical to pediatric intracranial EEG recordings. Front Neurol. 2016 Dec xix;seven–229.
    • View Commodity
    • Google Scholar
11. 11. Webber WRS, Bottom RP. Automated spike detection in EEG. Clin Neurophysiol. 2017 Jan;128(1):241–two. pmid:27940048
    • View Commodity
    • PubMed/NCBI
    • Google Scholar
12. 12. Gotman J, Koffler DJ. Interictal spiking increases afterwards seizures but does non after decrease in medication. Electroencephalogr Clin Neurophysiol. 1989 Jan;72(one):7–15. pmid:2464478
    • View Article
    • PubMed/NCBI
    • Google Scholar
13. thirteen. Tassinari CA, Rubboli G, Volpi L, Meletti Due south, d'Orsi One thousand, Franca 1000, et al. Encephalopathy with electrical status epilepticus during slow sleep or ESES syndrome including the caused aphasia. Clin Neurophysiol. 2000 Sep i;111:S94–S102. pmid:10996561
    • View Article
    • PubMed/NCBI
    • Google Scholar
14. 14. Lodder SS, Askamp J, van Putten MJAM. Inter-ictal spike detection using a database of smart templates. Clin Neurophysiol. 2013 Dec;124(12):2328–35. pmid:23791532
    • View Article
    • PubMed/NCBI
    • Google Scholar
15. fifteen. Mukhopadhyay S, Ray GC. A new interpretation of nonlinear energy operator and its efficacy in spike detection. IEEE Trans Biomed Eng. 1998 Feb;45(2):180–7. pmid:9473841
    • View Commodity
    • PubMed/NCBI
    • Google Scholar
16. 16. Malik MH, Saeed M, Kamboh AM. Automated threshold optimization in nonlinear energy operator based fasten detection. In Technology in Medicine and Biology Club (EMBC), 2016 IEEE 38th Almanac International Conference. 2016; pp. 774–7.
17. 17. Bergstrom RA, Choi JH, Manduca A, Shin H-S, Worrell GA, Howe CL. Automated identification of multiple seizure-related and interictal epileptiform event types in the EEG of mice. Sci Rep. 2013 Mar 21;3(one):1483.
    • View Article
    • Google Scholar
18. 18. Scheuer ML, Bagic A, Wilson SB. Spike detection: Inter-reader agreement and a statistical Turing examination on a big data set. Clin Neurophysiol. 2017 Jan;128(1):243–50. pmid:27913148
    • View Article
    • PubMed/NCBI
    • Google Scholar
19. 19. Wilson SB, Emerson R. Fasten detection: a review and comparison of algorithms. Clin Neurophysiol. 2002 December;113(12):1873–81. pmid:12464324
    • View Article
    • PubMed/NCBI
    • Google Scholar
20. xx. Chatrian GE, Bergamini L, Dondey M, Klass DW, Lennox Buchthal M, Petersen I. A glossary of terms most ordinarily used by clinical electroencephalographers. Electroenceph Clin Neurophysiol. 1974;37:538–48. pmid:4138729
    • View Article
    • PubMed/NCBI
    • Google Scholar
21. 21. Kane N, Acharya J, Benickzy Southward, Caboclo L, Finnigan Southward, Kaplan Prisoner of war, et al. A revised glossary of terms most commonly used past clinical electroencephalographers and updated proposal for the study format of the EEG findings. Revision 2017. Clin Neurophysiol Pract. Elsevier; 2017 January one;2:170–85.
    • View Article
    • Google Scholar
22. 22. Noachtar Due south, Schmidt D. Introduction. Epilepsy & Behavior. 2010 Oct;19(2):95–5.
    • View Article
    • Google Scholar
23. 23. Gerber PA, Chapman KE, Chung SS, Drees C, Maganti RK, Ng Y-T, et al. Interobserver understanding in the interpretation of EEG patterns in critically ill adults. J Clin Neurophysiol. 2008 Oct;25(5):241–9. pmid:18791475
    • View Article
    • PubMed/NCBI
    • Google Scholar
24. 24. Pearce PS, Friedman D, LaFrancois JJ, Iyengar SS, Fenton AA, MacLusky NJ, et al. Fasten—wave discharges in adult Sprague—Dawley rats and their implications for animal models of temporal lobe epilepsy. Epilepsy & Behavior. 2014 Mar;32:121–31.
    • View Commodity
    • Google Scholar
25. 25. Twele F, Schidlitzki A, Töllner M, Löscher Due west. The intrahippocampal kainate mouse model of mesial temporal lobe epilepsy: Lack of electrographic seizure-similar events in sham controls. Epilepsia Open. 2017 Jun;2(ii):180–vii. pmid:29588947
    • View Article
    • PubMed/NCBI
    • Google Scholar
26. 26. Staley KJ, White A, Dudek Atomic number 26. Interictal spikes: Harbingers or causes of epilepsy? Neurosci Lett. 2011 Jun 27;497(3):247–50. pmid:21458535
    • View Article
    • PubMed/NCBI
    • Google Scholar
27. 27. Goodin DS, Aminoff MJ. Does the interictal EEG have a part in the diagnosis of epilepsy? Lancet. 1984 Apr 14;ane(8381):837–ix. pmid:6143148
    • View Article
    • PubMed/NCBI
    • Google Scholar
28. 28. Nonclercq A, Foulon Yard, Verheulpen D, De Cock C, Buzatu M, Mathys P, et al. Cluster-based spike detection algorithm adapts to interpatient and intrapatient variation in spike morphology. J Neurosci Meth. 2012;210(2):259–65.
    • View Article
    • Google Scholar
29. 29. Thomas J, Jin J, Dauwels J, Cash SS, Westover MB. Clustering of interictal spikes by dynamic fourth dimension warping and affinity propagation. Proc IEEE Int Conf Acoust Speech Signal Procedure. IEEE; 2016 Mar;2016:749–53.
    • View Article
    • Google Scholar
30. 30. Wahlberg P, Lantz One thousand. Methods for robust clustering of epileptic EEG spikes. IEEE Trans Biomed Eng. 2000 Jul;47(7):857–68. pmid:10916256
    • View Article
    • PubMed/NCBI
    • Google Scholar
31. 31. Souza BC, Lopes-dos-Santos V, Bacelo J, Tort AB. Spike sorting with Gaussian mixture models. bioRxiv. 2018 January 17;248864.
32. 32. Lévesque Chiliad, Avoli M. The kainic acrid model of temporal lobe epilepsy. Neurosci Biobehav Rev. 2013 Dec;37(10):2887–99.
    • View Article
    • Google Scholar
33. 33. National Enquiry Council. Guide for the Intendance and Use of Laboratory Animals. 2010.
34. 34. Umpierre AD, Bennett IV, Nebeker LD, Newell TG, Tian BB, Thomson KE, et al. Repeated low-dose kainate administration in C57BL/6J mice produces temporal lobe epilepsy pathology but exceptional spontaneous seizures. Exp Neurol. 2016 May;279:116–26. pmid:26896834
    • View Commodity
    • PubMed/NCBI
    • Google Scholar
35. 35. Tse K, Puttachary S, Beamer E, Sills GJ, Thippeswamy T. Advantages of repeated low dose against unmarried high dose of kainate in C57BL/6J mouse model of status epilepticus: behavioral and electroencephalographic studies. PLoS ONE. 2014 May 6;ix(5):e96622. pmid:24802808
    • View Article
    • PubMed/NCBI
    • Google Scholar
36. 36. Racine RJ. Modification of seizure activity by electrical stimulation. II. Motor seizure. Electroencephalogr Clin Neurophysiol. 1972 Mar;32(three):281–94. pmid:4110397
    • View Article
    • PubMed/NCBI
    • Google Scholar
37. 37. Wallace Eastward, Kim DY, Kim K-M, Chen Southward, Braden BB, Williams J, et al. Differential effects of duration of slumber fragmentation on spatial learning and synaptic plasticity in pubertal mice. Brain Res. 2015;1615:116–28. pmid:25957790
    • View Article
    • PubMed/NCBI
    • Google Scholar
38. 38. Nelder JA, Mead R. A Simplex Method for Office Minimization. The Computer Journal. 1965 Jan ane;7(4):308–13.
    • View Commodity
    • Google Scholar
39. 39. Twyman RE, Macdonald RL. Kinetic properties of the glycine receptor main- and sub-conductance states of mouse spinal cord neurones in culture. J Physiol (Lond). 1991 Apr;435:303–31.
    • View Article
    • Google Scholar
40. 40. Fawcett T. An introduction to ROC analysis. Design Recognit Lett. 2006 Jun 1;27(8):861–74.
    • View Article
    • Google Scholar
41. 41. Pearson 1000. LIII. On lines and planes of closest fit to systems of points in space. Lond Edinb Dubl Phil Mag. 2010 Jun 8;ii(eleven):559–72.
    • View Article
    • Google Scholar
42. 42. Hotelling H. Analysis of a complex of statistical variables into chief components. J Educ Psychol. 1933;24(6):417–41.
    • View Article
    • Google Scholar
43. 43. McLachlan G, Peel D. Finite Mixture Models. In Wiley series in probability and statistics: practical probabilities and statistics section. New York. Wiley & Sons Inc; 2000.

collinsquoinep.blogspot.com

Source: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0207158

Share this post