Systems and methods for localizing epileptogenic zones from high frequency oscillation sequences

Abstract

Disclosed herein is a method for predicting the location of the epileptogenic zone in an epileptic subject using high-frequency oscillations (HFOs) as biomarkers. The method first detects and discriminates on HFOs in EEG and iEEG recordings. Next, sequences of HFO are detected and filters enabling the method to identify the spread of HFOs to localize the EZ in the brain. The disclosed method automatically identifies HFO sequences and determines and visualizes the location of the epileptic tissue.

Description

SYSTEMS AND METHODS FOR LOCALIZING EPILEPTOGENIC ZONES FROM HIGH FREQUENCY OSCILLATION SEQUENCESRelated Applications

[0001] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 641,740, filed May 2, 2024, the contents of which are incorporated herein in their entirety.Government Interest

[0002] This invention was made with United States government support under contract NS096761 awarded by the National Institutes of Health. The U.S. Government has certain rights in the invention.Background

[0003] Epilepsy, a condition characterized by unprovoked seizures, is one of the most common neurological disorders in the world, affecting around 70 million people globally. The lives of epilepsy patients are negatively affected, specifically when seizures are not able to be controlled. Unfortunately, around 30% of the patients are refractory to medications. These patients with medically refractory epilepsy (MRE) could suffer from various cognitive impairments induced by incessant seizures (e.g., losing consciousness, and even a raised level of mortality compared to the general population). Forthese patients, surgical interventions, typically resective surgery or intracranial neurostimulation, may be viable options to control or even stop the seizures. In this case, localizing the epileptogenic zone (EZ), the cortical area that is responsible for seizure generation, is of great importance for guiding the brain stimulation or achieving postsurgical seizure freedom in resective surgery.

[0004] Localizing the EZ is challenging, as no currently available diagnostic tool can directly measure the EZ. Clinically, the EZ can be approximated post-surgically by the resection in patients who become seizure-free, however, because surgical resection is often conducted large enough to ensure that all epileptogenic tissue is removed, the surgical area would most likely be an overestimate of the EZ. Another common practice is to determine the seizure-onset zone (SOZ) from intracranial EEG. Although successful in many patients, about 40-50% of patients do not achieve a favorable surgical outcome, which suggests that the current clinical procedure is suboptimal. Thus, there is a need for investigating accurate biomarkers to localize the epileptogenic tissues.

[0005] Recent studies in clinical research have shown that high-frequency oscillations (HFOs), including ripples (80 - 250 Hz) and fast ripples (250 - 500 Hz), are a promising biomarker of the EZ and SOZ. HFOs are reported to be increased in an epileptogenic brain and correlate with ictogenesis. The removal of brain areas containing HFOs is found to improve surgicaloutcomes which may be more accurate than the SOZ. Therefore, it is necessary to detect HFOs as a biomarker in localizing the epileptogenic brain with the goal of revealing the underlying EZ prior to surgery to achieve favorable treatment outcomes.

[0006] Currently, intracranial electroencephalogram (iEEG) recording using electrocorticogram (ECoG) and / or stereo-EEG (sEEG) is the current state-of- the-art practice to determine pathological and physiological electrophysiological activities for localizing the epileptogenic brain for pre- surgical planning. Candidates of epilepsy surgery are implanted with iEEG electrodes to record multiple seizures and allow for electrical stimulation studies to map the eloquent cortex which is required for safe epilepsy surgery.

[0007] This process typically takes 5-10 days with medications reduced or ceased for seizures to occur prior to the surgical treatment. The recorded seizures and other typical epileptiform activities are studied to roughly estimate the anatomic site of the EZ, even though it is not always conclusive. Although iEEG studies to determine EZ from seizure recordings are useful and successful in many patients, due to its nature of lengthy invasive recordings, patients experience discomfort and increased risk of serious infection associated with chronically implanted arrays during multi-days invasive monitoring. Consequently, the limitations of current practice in invasive methods are well recognized. Accurate electrophysiological biomarkers withinnovative clinical paradigms are thus highly necessary. New understandings and insights about the underlying epileptic network may provide innovative technologies for localization of the epileptogenic brain, which will substantially reduce the duration of invasive procedures and will have a significant clinical impact on the management of epilepsy.

[0008] In general, intracranial EEG (iEEG) provides an excellent tool to record and localize the epileptic brain, given its high signal-to-noise ratio (SNR), low signal attenuation, and spatial coverage of both the superficial and deep brain structures. The current process of pre-surgical investigation mainly focuses on the conventional biomarkers, such as the SOZ depicting the origin of ictal activities (seizures) and the irritative zone (IZ) which corresponds to the interictal epileptiform discharges (lEDs, or spikes). Although these activities have been shown to have significant pathological significance, they are limited, both spatially and temporally, in their ability to delineate the EZ with high specificity. Ictal signals propagate rapidly from onset to nearby and even farther tissues in a short time, therefore the SOZ is difficult to localize and may take multiple days to capture.

[0009] In contemporary clinical practice, there is a pressing need for reliable methods to identify pathological HFOs and accurately delineate the epileptogenic network at the individual patient level. Therefore, it is desirable to develop a valid and reliable biomarker based on HFOs and to establish the efficacy of HFOs in association to the epileptogenic brain.Summary of the Invention

[0010] Disclosed herein is a method for the detection and discrimination of high- frequency oscillations (HFOs) in electroencephalogram (EEG) and iEEG recordings and to systematically sequence and analyze propagational HFOs over extensive EEG and iEEG recordings. Spontaneous HFO sequences are a reliable marker of the EZ. The method detects and discriminates HFO sequences, enabling the method to identify the spread of HFOs to localize the EZ in the brain. The disclosed method automatically identifies HFO sequences and determines the location of the epileptogenicic tissue. This information can be used by clinicians to make a pre-surgical diagnosis and evaluate the results of a surgical procedure. The disclosed method provides a more accurate and efficient means of diagnosing and managing epilepsy, ultimately leading to improved patient outcomes.Brief Description of the Drawings

[0011] By way of example, specific exemplary embodiments of the disclosed system and method will now be described, with reference to the accompanying drawings, in which:

[0012] FIG. 1 is a schematic illustration of an overview of the disclosed method.

[0013] FIG. 2 is a flowchart of the overall method for localizing and visualizing theEZ.

[0014] FIG. 3 is a flowchart showing the process for detecting and discriminating HFOs of interest.

[0015] FIG. 4 is a histogram of detected HFOs in a recording from a representative subject.

[0016] FIG. 5 shows clusters of HFOs in each grouping.

[0017] FIG. 6 is a flowchart showing the detection and filtering of HFO sequences.

[0018] FIG. 7 shows examples of identified HFO sequences.

[0019] FIG. 8 shows two example HFO sequences with large extent.

[0020] FIG. 9 shows two HFO sequences with localized margin.

[0021] FIG. 10 is a flowchart showing the process of performing a spatial mapping of the HFO sequences.

[0022] FIG. 11 shows the localization of SOZ and estimated resection zone using HFO sequences (HFO-seq) and all HFOs (aHFOs).

[0023] FIG. 12 shows the spatiotemporal dynamics and stability of HFO sequences.Detailed Description

[0024] Disclosed herein is system and method for detecting HFO sequences that are indicative of a localization of the brain area and connections responsible for seizure generation (i.e., the epileptogenic zone). Note that references to "iEEG data" are meant to include a time series of data collected via iEEG, EEG or via magnetoencephalogram.

[0025] HFOs are defined as events that stand out of the background with an approximately sinusoidal shape and a duration of at least four cycles. Sequences of HFOs are concurrently propagating activities that are highly associated with the pathological zones of epilepsy, by which the underlying tissue epileptogenicity can be localized and delineated. Multiple spatiotemporal divarication may exist, while the highly recruited core region of HFO sequences provides precise and stable mapping of the epileptogenic networks.

[0026] As an overview of the method, high-frequency oscillation (HFO) sequences are saved over a 256 ms window centered on the peak, extracted from the iEEG recordings and mapped as spatial trajectories across the co-registered electrode positions, representing the propagation of HFO events over time and space.

[0027] FIG. 1 is a schematic illustration showing the overview of the method: (A) illustrates the iEEG data analysis workflow. Long-term iEEG data were collected and HFOs were detected and sequenced to identify the propagation of HFOs, which consists of a series of HFOs being captured across multiple channels spanning a tight time window in a relatively local cortical area; (B) illustrates the clinical evidence and modeling. MRI is used to build a realistic head model for individual patients, CT images are co-registered with the MRI and used to localize the iEEG implantation from which the SOZ is defined by the clinicians, and post-surgical MRI is for modeling the surgical resection zone (green); (C) illustrates the mapping of spontaneous HFO sequences. (D)illustrates connectivity antagonism between the HFOs and surrounding areas.The information flow between the HFO-zone and outside is associated with the propagation of the HFOs, and hence reflects the underlying pathology of the epileptic network.

[0028] FIG. 2 is a flowchart showing the steps of the method 200 for localizing and visualizing the EZ. At step 300, HFOs are first detected, then discriminated to discover HFOs of interest, that is, HFOs that will contribute to the localization of the EZ. At step 600, sequences of the HFOs of interest are detected and screened to remove sequences of lesser clinical significance. At step 1000, the HFOs of interest are spatially mapped to a model of the brain of the subject to localize and visualize the EZ. All steps of method 200 are explained in detail herein.

[0029] Detection and Discrimination of Interictal High-Frequency Oscillations

[0030] A flowchart of the method 300 for detecting and discriminating HFOs from a time series of EEG, iEEGs or via a magnetoencephalogram is shown in FIG. 3. First, a long-term time series of electroencephalography (EEG), intracranial electroencephalography (iEEG) or magnetoencephalogram recordings are obtained at 302 for the subject, during interictal periods. The recordings are captured at a sampling rate typically between 500 Hz and 3000 Hz or above. The length of data used for analysis varies, depending on the availability and quality of the continuous clinical monitoring.

[0031] Automated screening of extracted long-term iEEG data is then conducted using a method similar to a previously established method for high-frequency oscillations (HFOs) detection, disclosed in U.S. Pub. Pat. App. No. 2021 / 0106247. Each channel is processed independently.

[0032] Initially, a visual examination of the recordings is conducted to verify data quality and to ensure minimal contamination by artifacts. Data from ictal periods, as identified by clinicians, are excluded from analysis at 304. Any segments with evident artifacts or saturation are also removed. Subsequently, at 306, the raw iEEG data is high-pass filtered above 1 Hz to remove baseline drifts and notch filtered at 60 Hz, and its harmonics are applied to reduce baseline drift and powerline noise. In one embodiment, a FFT Hann filter with a slope of 2 Hz may be used for this purpose.

[0033] The filtering step is followed by an energy detector screening at 308 to identify potential high-frequency activities (HFAs) distinct from background noise. The data are re-referenced based on electrode type— common average referencing for ECoG and white matter referencing for sEEG. Initial HFO detection is performed on a band-pass filtered signal (80-1000 Hz), using a moving-window energy thresholding approach. Each channel is analyzed independently to identify segments that exceed the median amplitude by a predetermined amount (in preferred embodiments, by 5x) over a predetermined timed baseline (in preferred embodiments, 100ms), producing a set of high-sensitivity candidate events. Specifically, in eachchannel, a 100 ms moving window was used to calculate the standard deviation distribution of the signal amplitude. A baseline threshold is set at five times the median of the distribution. Any sample in the filtered signal exceeding this amplitude threshold is marked as an initial candidate for HFOs. Generally, a low baseline was employed to ensure high sensitivity in this phase.

[0034] Each candidate is then evaluated over a 256 ms window centered on its peak. The signal envelope is computed using the Hilbert transform, and events are retained if they contain at least eight crossings above three times the envelope baseline (to satisfy the morphological criteria of oscillatory HFOs). Additional filtering removes events with more than ten zero-crossings in the unfiltered signal, a proxy for transient noise. The remaining candidates undergo a feature extraction process incorporating temporal, spectral, and spectrogram descriptors from both raw and filtered data. These features are used for unsupervised clustering using a Gaussian Mixture Model (GMM) initialized by k-means. The elbow method is employed to determine the optimal number of clusters. Events grouped into clusters with consistent morphology and spatial plausibility are preserved as putative HFOs, while clusters dominated by noise or artifacts are excluded.

[0035] To discriminate the putative HFOs with high precision, at 310, a comprehensive set of features is employed, incorporating aspects of temporal, spectral, and spectrogram analysis. Features are extracted fromcandidate HFOs using trained feature extractors. These features were extracted from each event that survived initial screening, in both unfiltered and filtered data (above 80 Hz). Time-frequency representation of these detected high-frequency events in the raw data was utilized to define the characteristics of each event across both low and high-frequency domains. The inventive method includes three categories of patterns: temporal, spectral, and spectrogram.

[0036] At 312, the extracted features facilitate the segregation of detected events into distinct clusters using, for example, an unsupervised Gaussian Mixture Model, initialized with k-means clustering. The optimal number of clusters was identified using the elbow method. Upon clustering, the characteristics within each cluster were visualized through piled waveforms of both unfiltered and filtered data, along with the spatial distribution of channels detecting these events. The clustering separates candidate HFOs into groups and distinguishes HFOs of interest from noisy groups. The noisy or artifact groups are then excluded and, at 314, the features from each candidate HFO are classified at using a trained classifier to identify HFOs of clinical interest.

[0037] The clustering of detected HFOs events serves a critical role in enhancing the specificity and clinical relevance of the overall mapping procedure. After initial automated detection, HFO candidates may include a mix of true neural oscillations, noise, and non-epileptic high-frequency transients. To differentiate these components, the previously-described comprehensivefeature extraction process is applied to each candidate event, capturing temporal, spectral, and time-frequency characteristics. These features are used in an unsupervised clustering procedure— implemented using a Gaussian Mixture Model— to segregate detected events into distinct groups based on morphological similarity and spectral profiles.

[0038] The purpose of this clustering step is to isolate biologically meaningful, repetitive, and spatially coherent HFO patterns from non-repetitive, artifactlike, or physiologically irrelevant signals. Clusters exhibiting consistent waveform morphology and plausible anatomical localization are retained as putative HFOs, while clusters dominated by high-amplitude noise, sharp transients, or irregular patterns are excluded. This refinement ensures that only high-quality, reproducible HFO events are used in downstream analyses, including sequence formation and spatial mapping, thereby improving the reliability of epileptogenic zone (EZ) localization.

[0039] FIG. 4 is a histogram of detected HFOs across an approximate full-day long recording from a representative subject. FIG. 5 shows clustered HFOs in each group are displayed with the raw data, high-pass filtered (>80 Hz) data, timefrequency representations, and spatial distribution over the iEEG electrodes located in a two-dimensional space modeled from post-implantation CT. The signal traces are piled with the mean in bold and colors correspond to each group. Note that the first three clusters show clear HFOs in the filtered signals (lower panels) and unique low-frequency profile in the raw signals(upper panels), with the first cluster exhibiting potential interictal epileptiform discharges in the raw signal, and the last two clusters showing obvious groups of artifacts.

[0040] The morphology of traces, both unfiltered and filtered, aid in differentiating clusters of putative HFOs from artifacts. As illustrated, various HFO clusters can be observed with distinct repetitive patterns, aligning with the previous findings. The consistency of these clustered events can be verified by examining the repetitiveness in the stacked signal waveforms and the spatial channel distribution. Those clusters containing evident artifacts, such as strong transient or high amplitude artifacts, as shown in FIG. 5, are excluded. The remaining events are preserved as putative HFOs for further analysis.

[0041] Identification of HFO Sequences

[0042] FIG. 6 is a flowchart showing the detection and filtering of HFO sequences. All detected HFOs from the initial phase are aggregated and categorized based on their temporal and spatial characteristics, namely, their occurrence time and channel location. An "HFO sequence" (HFO-seq) is a spatiotemporal cluster of HFOs that represent coordinated neural propagation. Retained HFOs are grouped based on their temporal proximity and anatomical location. More specifically, an HFO Sequence is defined as a series of HFO events spanning multiple electrodes within a brief timeframe, representing a spatiotemporal co-activation or propagation of HFO activities. These sequences are identified by analyzing the spatiotemporal patterns of theaggregated HFOs during long-term monitoring using the method described herein.

[0043] The identification process commences at 602 by selecting an initial HFO event as the "leading event" of a potential sequence. At 604, Subsequent HFO events occurring within a temporal window (in preferred embodiments, 150 ms) of this leader are incorporated into the sequence. Any events within 15 ms of an existing sequence member were also appended. This grouping process iterates through the entire dataset to extract all candidate sequences.

[0044] At 606, several biologically informed criteria are then applied to filter out spurious or implausible patterns. In one embodiment, the following filters are applied: (1) Sequences with fewer than a predetermined number of HFOs are discarded. In a preferred embodiment, sequences with fewer than three HFOs are discarded. (2) A propagation speed threshold (for example, of 10 m / s) is imposed between adjacent events to reflect plausible neural conduction velocities, unless supported by recurrent fast co-occurrence patterns (e.g., gap junction propagation). (3) Sequences wherein more than a predetermined percent of the events cluster within a predetermined time window (preferably a 2 ms time window) are removed to avoid contamination by cross-channel artifacts. In a preferred embodiment, the predetermined percentage is 50%. Finally, (4) a trajectory-based clustering refinement is performed to eliminate outlier sequences that are dissimilarfrom others in the same subject. Any combination of these filters, or other filters, may be used to determine whether of not the sequence is a sequence- of-interest.

[0045] The groupings that pass each of the filtering criteria are retained as HFO sequences of interest at 614, while those that fail any of the filtering criteria are discarded at 610. Process 600 iterates at 612 until the end of the data is reached. The sequences reflect coordinated pathological activation and are subsequently mapped across electrodes to form spatial trajectories— capturing not just isolated HFOs but their dynamic propagation across the cortical network.

[0046] Sequence identification process 600 plays a key role as a discerning tool for pathological HFOs. It is based on the observation that when HFOs occur concurrently within a network in a coordinated spatiotemporal manner, they are more likely to be closely associated with underlying pathological processes. The process repeats at 608 until the end of the collected iEEG data is reached.

[0047] FIG. 7 shows examples of identified HFO sequences. (A, B) Two HFO sequences from subject P3 (C,D) and two HFO sequences from P34. The examples showcase distinct HFOs morphology between sequences. The left panel of each row shows the raw signal data, and the middle panel displays the signals after high-pass filtering above 80 Hz. The red line indicates SOZ channels, and blue as non-SOZ channels. Onset and offset of events aredenoted by left and right-pointing triangles, respectively. The rightmost panel for each row visualizes the spatial positioning of iEEG channels (dark) on the cortical surface— channels participating in the sequences are highlighted in red, and the SOZ is marked with a green hexagon.

[0048] FIG. 8 shows two example HFO sequences with large extent, associated with cortical regions encompassing the amygdala, hippocampus, and middle temporal region. FIG. 9 shows two HFO sequences with localized margin, involving mainly the hippocampus region. Note that within each group, (FIG.8 and FIG. 9), the HFO sequences (upper and lower panel) exhibit similar morphological characteristics, indicating a consistent pattern of neural activity. However, when comparing across the groups, while there is a high spatial overlap in the involved electrodes (both recruiting the right hippocampus), the morphology and sequential order of the HFOs vary, underscoring distinct neural dynamics between these cortical areas. Bold red lines mark the SOZ channels, and left / right triangle indicate the onset / offset of the HFOs events.

[0049] Mapping of HFO Sequences and Performance Assessment

[0050] FIG. 10 is a flowchart showing the process of performing a spatial mapping of the HFO sequences. First, an individualized cortical model is constructed for each subject at step 1002, using segmentation of gray matter from their presurgical magnetic resonance imaging (MRI). This model serves as the anatomical foundation for subsequent spatial analyses. At 1004, three-dimensional electrode positions, including both grid and depth electrodes, are reconstructed from post-implantation imaging data and co-registered to the pre-surgical MRI. These electrode locations are then projected onto the cortical surface model to provide spatial references for intracranial EEG (iEEG) data. If surgery has been performed on the patient, the surgical resection area may also be segmented from post-surgical MRI scans and aligned with the pre-surgical MRI space to enable accurate localization of resected tissue relative to the pre-implantation brain anatomy.

[0051] To capture the cumulative spatial footprint of these sequences over a defined time period, individual HFO events are aggregated to form a spatiotemporal density map at 1010. This map highlights regions of consistently elevated HFO activity, referred to as the HFO-core, which is hypothesized to correspond to the epileptogenic zone (EZ). To robustly identify these core regions while accounting for variability in HFO event distribution, a nonparametric bootstrap confidence thresholding procedure is used at 1012. Specifically, the set of identified HFO sequences is resampled with replacement 1,000 times to generate a distribution of spatial activity maps. For each resample, a new cumulative spatial map is computed, and the mean spatial density across these 1,000 iterations is calculated for each electrode or cortical location. From this distribution of means, the 10% confidence interval is estimated. For example, the lower 10th percentile value at each location is used as a conservative threshold. This approach helps distinguishlocations that are reliably and repeatedly involved in HFO sequences from those where activity may be due to noise or random fluctuations. Any location with a spatial activity value exceeding this threshold is considered significantly active and is included in the HFO-core. This bootstrap-based confidence thresholding approach provides a data-driven, patient-specific, and statistically grounded method for identifying the most relevant cortical regions implicated in HFO activity, reducing bias from arbitrary cutoffs and enhancing robustness in epileptogenic zone estimation.

[0052] In addition to HFO sequences, the same mapping and estimation procedures may be applied to two other biomarkers— namely, all detected HFOs (aHFOs) and the onsets of HFOs (oHFOs)— to assess their respective spatial distributions. These estimations are quantitatively compared against a clinically defined EZ, which integrates information from two sources: the seizure onset zone (SOZ), identified through expert analysis and coregistration of post-implantation CT with pre-surgical MRI; and the surgical resection area, segmented from post-surgical MRI and mapped to the same reference space.

[0053] FIG. 11 shows the localization of SOZ and resection zone using HFO-seq and aHFOs. Mapping of HFO sequences (HFO-seq) and all HFOs (aHFOs) to localize the SOZ are shown as the green stars. The iEEG electrodes are shown in dark, with one case for ECoG (left panel) and one for sEEG (right panel). Theestimations are color coded from light (minimal activity) to warm (maximal activity).

[0054] FIG. 12 shows the spatiotemporal dynamics and stability of HFO sequences.(A) Temporal distributions of HFOs onset timings across various channels relative to a specific SOZ channel. Each row represents the probability distribution for one recording channel, with a white dashed line marking the peak probability timing. Notable features include the concentration of channels exhibiting either leading (e.g., R-Mid-T) or lagging (e.g., initial channels in R-Hipp) tendencies, as well as the bimodal distribution consisting of both leading and lagging instances (e.g., later channels of R-Hipp). Also note a flatter, uniform-like histogram from R-Lat-Fr-T. (B) Clustering of HFO sequences based on spatiotemporal patterns. The cluster indices are displayed in the left, with varying colors encoding different groups. The corresponding sequences are enumerated with the sequence indices in the vertical axis and the horizontal axis lists channels engaged in the sequences. The color-coding scheme of HFO sequences represents the propagation tendency of activities within the channels, ranging from cool (early, or upstream) to warm (late, or downstream) colors. (C) Association of channels with specific spatiotemporal clusters. Clusters are color coded consistent to the colors in (B) on the left side with the vertical length of each color block proportional the size of the cluster. Channels involved in the sequences are displayed on the right, paired with the respective clusters, and the corticalstructures where the channels are located are labeled on the right axis. Note a high degree of spatial overlap in channel involvement across the different clusters. (D-l) Two representative and distinct sequence clusters, separating the bimodal channel behaviors noted in (A). (D,G) Temporal distributions of HFOs onset timings across various channels involved in the sequences of specific clusters. (E,H) Sequential interactions among channels within various cortical regions in the identified clusters. The co-activation matrix, color- coded to represent co-occurrence counts, highlights the significant recruitment of the amygdala, hippocampus, and middle temporal regions, aligning with the SOZ. (F,l) Examples of HFO sequences within each cluster, demonstrating different propagation orders, despite considerable overlap in channel recruitment across clusters. The colors of the signals depict the relative timing of onsets, from cool (early) to warm (late). (J) Propagation tendency of each sequence within cluster 1 (D-F), with colors encoding the preference for leading or lagging. (K) Sequence similarity across all patients against a shuffled control group. (L) Hubness measures of nodes in SOZ versus outside (non-SOZ) during HFO sequences: degree centrality, betweenness centrality, and eigenvector centrality. R: right, Amyg: amygdala, Hip: hippocampus, Mid: middle, T: temporal, Ant: anterior, Lat: lateral, Fr, frontal.

[0055] Evaluation

[0056] To evaluate the spatial accuracy of the HFO-based markers relative to the clinical EZ, three performance metrics are used. The activation patterns are analyzed by evaluating the spatial distribution of HFO-seq across all sampled locations of the iEEG electrodes. The extent of the active regions of HFO sequences, also referred to as the highly recruited HFO-core, is determined using a threshold derived from the 10% confidence interval of the mean spatial distribution, calculated through a bootstrap method with 1000 replications. This technique provided an overarching estimation of the EZ using HFO-seq data. Similarly, this mapping approach was also applied to analyze other well-established benchmarks, including all detected HFOs (aHFOs) and the onset of HFOs (oHFOs), to estimate their respective spatial distributions. The estimated distributions of these HFO markers are quantitatively evaluated for each subject by comparing them against the clinically presumed epileptogenic zone (EZ), which is identified based on key clinical findings, including the SOZ and the areas of surgical resection. The SOZ is localized through the co-registration of post-implantation CT images with pre-surgical MRI, as determined by clinical experts, and the resection area is modeled from post-surgical MRI, co-registered with pre-surgical MRI.

[0057] To conduct this evaluation effectively, several established performance metrics were utilized: (i) localization error (LE), (ii) normalized overlap ratio (NOR), and (iii) spatial dispersion (SD). The localization error (LE) was defined as the minimum distance between the identified HFO markers and thenearest point in the clinically defined EZ (the "ground truth"). The normalized overlap ratio (NOR) was calculated by determining the overlap between the estimated distribution and the clinical EZ, normalized by the area of either the ground truth or the estimation. This ratio was expressed in two forms, recall and precision, and the combined geometric mean was also computed to provide an integrated value ranging from 0 (no overlap) to 1 (perfect overlap). Spatial dispersion (SD) was quantified as the weighted quadratic mean of the minimum pairwise distances between the estimated distribution and the clinical EZ. This metric characterizes the spatial spread of the estimated distribution relative to the ground truth. The specific mathematical formulations for calculating the normalized overlap ratio and spatial dispersion are detailed below.

[0058] where SOvipis the overlap area between the estimation and the ground truth, and Ssand SGare the area of the estimation and the ground truth, respectively.

[0059] where jtis the spatial recurrence rate of estimation i, di Gis the minimum distance from the estimation i to the ground truth, and Nsis the number of distributed locations in the estimation.

[0060] The method may be embodied in a system comprising a processor and software executed by the processor to cause the system to perform the steps of the method of the framework. The high-dimensional electrophysiological may be collected from an EEG or iEEG device operatively coupled to the processor, or the data may be pre-recorded and stored in database, from where the software may retrieve the data. The visualization of the localization of the area of the biological system producing the biomarkers of interest may be displayed on a display coupled to the processor.

[0061] While the disclosure has been described in detail and with reference to specific embodiments thereof, it will be apparent to one skilled in the art that various changes and modification can be made therein without departing from the spirit and scope of the embodiments. Thus, it is intended that the present disclosure cover the modifications and variations of this disclosure provided they come within the scope of the appended claims and their equivalents.

[0062] Further, the features disclosed in the foregoing description, or the following claims, or the accompanying drawings, expressed in their specific forms or in terms of a means for performing the disclosed function, or a method or process for attaining the disclosed result, as appropriate, may, separately, orin any combination of such features, be utilized for realizing the invention in diverse forms thereof. In particular, one or more features in any of the embodiments described herein may be combined with one or more features from any other embodiments described herein.

[0063] Together, this method enables robust spatial mapping of HFO dynamics in relation to individual cortical anatomy and provides a quantitative framework for comparing candidate HFO biomarkers against clinically validated epileptogenic regions.

Claims

Claims:

1. A method for localizing an epileptogenic zone in a brain of a subject, comprising: collecting a time series of electroencephalogram data from the subject; detecting high frequency oscillations within the data; identifying sequences of the high frequency oscillations; and mapping the sequences of high frequency oscillations to a model of the brain of the subject.

2. The method of claim 1 wherein the electroencephalogram data is collected intracranially.

3. The method of claim 1 further comprising: filtering the electroencephalogram data to remove high frequency oscillations activities of lesser clinical interest, leaving a plurality of candidate high frequency oscillations.

4. The method of claim 1 wherein the filtering comprises: exposing the electroencephalogram data to a high-pass filter.

5. The method of claim 3 wherein the filtering comprises:notch filtering the electroencephalogram data.

6. The method of claim 3 further comprising: band-pass filtering the electroencephalogram data; detecting an initial set of high frequency oscillations of interest based on4 using a moving-window energy thresholding approach wherein each channel is analyzed independently to identify segments that exceed a median amplitude by a predetermined amount over a timed baseline, producing a set of high-sensitivity candidate events.

7. The method of claim6 further comprising: extracting one or more features from each candidate high frequency oscillation, using one or more trained feature extractors.

8. The method of claim 7 further comprising: exposing the extracted features from each high frequency oscillation to a classifier that determines if extracted features are indicative of a high frequency oscillation of interest.

9. The method of claim 8 further comprising: clustering high frequency oscillations of interest based on temporal, spectral, and time-frequency characteristics to segregate detected eventsinto distinct groups based on morphological similarity and spectral profiles.

10. The method of claim 1 wherein identifying sequences of high frequency oscillations comprises iteratively performing the steps of: selecting a next available high frequency oscillation as a leading event; adding subsequent high frequency oscillations within a predetermined time period to the leading event to form a sequence of high frequency events; selecting sequences that pass one or more biologically-informed filters as sequences-of-interest; and repeating the iteration until the end of the electroencephalogram data has been reached.

11. The method of claim 10 wherein selecting sequences comprises: discarding sequences having less than a predetermined number of high frequency oscillations.

12. The method of claim 10 wherein selecting sequences comprises: discarding sequences wherein more than a predetermined percentage of the events cluster within a predetermined time window.

13. The method of claim 10 wherein selecting sequences comprises: performing trajectory-based clustering refinement to eliminate outlier sequences.

14. The method of claim 10 wherein selecting sequences comprises: discarding sequences having a propagation speed of greater than 10 meters per second between adjacent events to reflect plausible neural conduction velocities.

15. The method of claim 1 wherein mapping the sequences of high frequency oscillations to a model of the brain of the subject comprises: creating the model using segmentation of gray matter from data collected during magnetic resonance imaging of the brain of the subject.

16. The method of claim 15 wherein the model is a cortical surface model.

17. The method of claim 16 further comprising: co-registering locations of EEG electrodes to the MRI data; and projecting the electrode locations onto the cortical surface model.

18. The method of claim 17 further comprising:mapping the sequences of high frequency oscillations as spatial trajectories across electrode locations.

19. The method of claim 18 further comprising: aggregating individual high frequency oscillation events to form a spatiotemporal density map which highlights regions of consistently elevated high frequency oscillation activity corresponding to the epileptogenic zone.

20. The method of claim 19 further comprising: generating a distribution of spatial activity maps by resampling the identified sequences of high frequency oscillations to create a new cumulative spatial map; calculating a mean spatial density across iterations of the resampling for each electrode or cortical location; and applying a confidence interval to distinguish locations that are reliably and repeatedly involved in sequences of high frequency oscillations from those where activity may be due to noise or random fluctuations.

21. The method of claim 20 further comprising: visualizing the spatiotemporal density map on the cortical surface model.

2. A method for localizing an epileptogenic zone in a brain of a subject, comprising: collecting a time series of scalp electroencephalogram or magnetoencephalogram data from the subject; detecting high frequency oscillations within the collected data; performing inverse source imaging from the collected data to estimate current density solutions within the brain; identifying sequences of the high frequency oscillations from the source imaged signals; and mapping the sequences of high frequency oscillations to a model of the brain of the subject.

Images