Spatial-temporal-spectral source imaging of bioelectrical activity

Abstract

Disclosed herein is a spatial-temporal-spectral imaging (STSI) framework that is capable of imaging the location, extent and dynamics of biomarkers with varied spatial, temporal and spectral profiles. The framework decomposes complex, high-dimensional electrophysiological data (e.g., EEG / MEG) into low-dimensional representations, delineates the spectral characteristics of different biomarkers in electrophysiological measurements, and images the location, dynamics, and extent of sources generating these biomarkers.

Description

Attorney Docket: 8350.2025-118WOSPATIAL-TEMPORAL-SPECTRAL SOURCE IMAGING OF BIOELECTRICAL ACTIVITYRelated Applications

[0001] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 725,299, filed November 26, 2024, the contents of which are incorporated herein in their entirety.Government Interest

[0002] This invention was made with United States government support under contracts NS096761 and NS127849 awarded by the National Institutes of health. The U.S. government has certain rights in the invention.Background

[0003] Bioelectric activity is generated by excitable cells within a biological system. The bioelectric sources due to cell activation generate electrical potential or magnetic field over the body surface. Electroencephalography (EEG) is the electrical potential over the scalp generated by brain electrical sources. Magnetoencephalography (MEG) is the magnetic field over the scalp generated by brain electrical sources. Electrocardiography (ECG) is the electrical potential over the body surface generated by cardiac electrical sources.Magnetocardiography (MCG) is the magnetic field over the body surfaceAttorney Docket: 8350.2025-118WO generated by cardiac electrical sources. Similarly, electromyography (EMG) is the electrical potential generated by muscle electrical activity.

[0004] Localizing and imaging of bioelectrical sources from body surface recorded electrical or magnetic field measurements involves solving of an inverse imaging problem. The estimation and imaging of bioelectric sources from clinically available EEG / MEG / ECG / MCG / EMG provide important information about the function and dysfunction of organ systems such as the brain, the heart and muscles. Such information is important for clinical diagnosis, treatment and rehabilitation of various diseases.

[0005] Epilepsy is one of the most prevalent neurological disorders, affecting over 65 million people worldwide. Approximately 30% of epilepsy patients experience seizures that cannot be controlled with medications alone. For this subset, brain surgery can be a viable treatment option, if the epileptogenic zone (EZ) can be accurately localized and safely removed. The EZ represents the minimal brain region that must be resected or disconnected to achieve seizure freedom, making a successful pre-surgical evaluation critical.

[0006] The current state-of-the-art standard for pre-surgical evaluation is intracranial electroencephalography (iEEG), in which electrodes are surgically placed over the presumed epileptogenic regions to perform functional mapping, and record epileptogenic brain activity, such as interictal discharges and ictal events. However, due to its invasive nature, iEEG is associated with inherent risks and is limited in its ability to cover large cortical areas comprehensively.Attorney Docket: 8350.2025-118WO

[0007] Electrophysiological source imaging (ESI) is a non-invasive method that estimates brain activity at the source level, using electroencephalography (EEG) or magnetoencephalography (MEG) recordings. Various biomarkers present in EEG / MEG data, such as interictal spikes, high-frequency oscillations (HFOs), and seizures, can serve as inputs to ESI algorithms. These biomarkers play distinct roles in delineating epileptogenic sources. Research has shown that the spatial extent of a source is frequency-dependent: higher frequencies typically correspond to smaller sources and weaker EEG / MEG signals. As a result, low- frequency spikes are generally easier to detect compared to HFOs. Notably, previous studies have suggested that brain regions generating HFOs are more strongly associated with seizure freedom than the seizure onset zone (SOZ) or irritative zone (IZ). This underscores the clinical need for accurately estimating the sources of various biomarkers.

[0008] However, due to the distinct temporal and spectral characteristics of different epileptic biomarkers, researchers often employ separate preprocessing techniques tailored to each biomarker. For example, filtering is commonly used to separate biomarkers with different frequencies, but this approach depends heavily on subjective choices of filter parameters. Alternative tailored approaches include applying spatio-spectral decomposition on HFO segments to enhance signal clarity before applying an ESI algorithm such as the standardized low-resolution brain electromagnetic tomography (sLORETA) source imaging, or employing wavelet-based source imaging algorithms on HFOs alongsideAttorney Docket: 8350.2025-118WO conventional dipole source imaging for spikes. Such customized pipelines limit the generalizability of ESI methods.

[0009] To address these limitations, a unified and adaptive ESI framework capable of imaging the source location, extent, and temporal dynamics of all epileptic biomarkers - including low-frequency and high-frequency events regardless of their frequency contents or ranges - would significantly facilitate clinical applications of epilepsy source localization, enhancing pre-surgical planning for drug-resistant epilepsy (DRE) patients.Summary of the Invention

[0010] Disclosed herein is a spatial-temporal-spectral imaging (STSI) framework. This framework decomposes complex, high-dimensional electrophysiological data (i.e., EEG / MEG) into low-dimensional representations, delineates the spectral characteristics of different biomarkers in EEG / MEG measurements, and images the location, extent, and temporal dynamics of sources generating these biomarkers with data-driven LI norm-based optimization. The framework is capable of imaging the location, extent and temporal dynamics of biomarkers with varied spatial, temporal and spectral profiles.Attorney Docket: 8350.2025-118WOBrief 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 the concept and framework.

[0013] FIG. 2 illustrates the process of tensor decomposition using canonical polyadic decomposition.

[0014] FIG. 3 is an illustration of the L operator, which transforms the source space tensor into the sensor space tensor.

[0015] FIG. 4 is an illustration of the norm of difference between original and reconstructed EEG / MEG tensor slice at each frequency being less than the noise.

[0016] FIG. 5 is an illustration of the iterative solving process whereby the solution obtained with the unconstrained problem is projected to a hyper-ellipsoid representing the constraint.

[0017] FIG. 6A is an illustration showing the alternating projection method.

[0018] FIG. 6B is an illustration showing the average projection method.

[0019] FIG. 7(A-E) is an illustration showing tensor components of an example EEG signal.

[0020] FIG. 8(A-D) shows an exemplary application of the STSI framework for subjects having overlapping HFOs and interictal spikes.Attorney Docket: 8350.2025-118WODetailed Description

[0021] Disclosed herein is a unified framework (and aspects, methods, systems, implementations, and associated apparatuses thereof) to perform source imaging using a spatial-temporal-spectral imaging (STSI) approach analysing various electrophysiological biomarkers.

[0022] In some embodiments, the Spatial-Temporal-Spectral Imaging (STSI) framework may include an advanced methodology for source imaging of electrical activity from electromagnetic signals originated from a biological system. In some embodiments, the STSI technology may efficiently decompose and handle biomarkers of various frequency ranges. In some embodiments, the biological system is the brain of a subject and the electrical activity is collected using electrical or magnetic sensors (e.g., EEG or MEG). The invention is described herein using brain activity collected by EEG, but, as would be realized by one of skill the art, the invention may be applied to any biological system and may use electrical or magnetic measurements.

[0023] In a first step of the method, STSI begins by decomposing higher order electrophysiological signals (collected by electric or magnetic sensors) into three- dimensional tensors having spatial, temporal, and spectral dimensions, using tensor decomposition methods. In one embodiment, the method operates on high-dimensional electrophysiological data that has been previously recorded and stored, rather than data acquired in real-time. In other embodiments, the framework may operate in real time.Attorney Docket: 8350.2025-118WO

[0024] Tensor decomposition, a generalization of matrix decomposition, enables the extraction of low-dimensional representations from complex, multi-dimensional data, facilitating analysis across diverse dimensions. In neuroscience, tensor component analysis (TCA) has been instrumental for analysing both invasive and non-invasive recordings and uncovers key structures in neural activity.

[0025] Given the wide frequency range spanned by different epileptic biomarkers, TCA isolates biomarkers across frequency domains, thereby aiding in the localization of epileptogenic sources. By decomposing higher-order EEG data structure into interpretable tensor components, TCA bridges the gap between highdimensional EEG data and actionable models of epileptic network dynamics.

[0026] The first step also identifies components representing biomarkers with distinct spatial, temporal and spectral profiles observed in a biological system. Examples include, but are not limited to, low-frequency activity including evoked potentials and event-related potentials, interictal spikes, and event-related synchronization / desynchronization in a brain, or electrical activity associated with normal or abnormal heart conduction, and high-frequency oscillations associated with physiology or pathology in a brain, heart or muscle.

[0027] The co-occurrence of HFOs with interictal spikes is well-known phenomena.These events, commonly referred to as "HFO riding spikes" or "spike ripples" (if the HFOs fall within the ripple band of 80 - 250 Hz), have shown utility in predicting the EZ. For instance, HFOs co-occurring with spikes can be used to enhance localization of the EZ, and spikes associated with HFO events exhibitAttorney Docket: 8350.2025-118WO improved performance in source localization. Spike ripples, a combination of epileptiform spikes and ripples, serve as a more precise biomarker for identifying epileptogenic tissue than other interictal biomarkers such as spikes, spikegamma, wideband HFOs, ripples, or fast ripples.

[0028] It should be noted that the framework does not inherently look for a certain type of frequency profiles or biomarkers. Instead, after transforming the EEG datastream into interpretable components that represent potential biomarkers (such as HFO and interictal spikes), a user chooses which components are to be imaged before feeding them into the source imaging part of the framework. The process works as follows: the EEG datastream is first transformed into the spectral domain, forming the tensor, and decomposed into multiple tensor components with distinct spatial, temporal and spectral profiles, as described above. If the epileptic biomarker (HFO, seizure) exists in the original EEG datastream, it will be expressed by certain component(s) with focused spatial topology, characteristic temporal evolution (e.g. oscillation with envelope for HFO, etc), and the spectral peak within the frequency range of that type of epileptic biomarker.

[0029] Given the complexity of real-world data, in some embodiments, the framework does not define fixed criteria for component classification (i.e., a spike, HFO or other irrelevant components (noise, etc)). Instead, it relies on a set of data- driven rules based on spatial topography, temporal evolution (i.e., "temporal dynamics"), and spectral frequency characteristics to guide the user through theAttorney Docket: 8350.2025-118WO detection and selection process. In one embodiment, the user reviews each component's spatial topography, temporal waveform, and spectral distribution, which will be further explained with reference to FIG. 7, and selects those corresponding to epileptic biomarkers. This review process may involve visual inspection or data-driven selection based on predefined criteria.

[0030] The method includes a second step directed to source imaging and estimation, which includes a source imaging algorithm that minimizes source sparsity and edge sparsity to estimate the location, extent and temporal dynamics of the underlying electric sources generating these biomarkers. This process is automated. In some embodiments, electrophysiological signal components from another step (including, but not limited to, the first step of the method previously described) serve as input for an optimization problem, which may minimize a weighted combination of source sparsity and source edge sparsity, with a constraint stating that the reconstructed electrophysiological signal closely matches the original signal (with a defined tolerance estimated from recording samples). This approach leverages a biological assumption that bioelectric signal sources are localized within specific regions of limited spatial extent. In some embodiments, through an iterative solving procedure, the source locations, extents, and temporal dynamics may be estimated, providing a comprehensive map of the bioelectrical activity.

[0031] The iterative optimization process starts with an initial estimate of the spatial distribution of the source activity over the cortex. This algorithm uses theAttorney Docket: 8350.2025-118WO minimum norm solution, a classic source imaging algorithm as initialization. There are two types of iterations, an outer iteration and an inner iteration. Within each outer iteration step is an inner iteration in which the unconstrained solution is iteratively obtained and projected to the hyper-ellipsoid that represents the constraint, which ensures that the solution satisfies the constraint. When the stopping criteria for this inner iteration are satisfied (e.g., the difference between two iterations is small enough, or a certain number of iterations has passed), the outer iteration begins, where a weighting coefficient is applied to each dipole location, so that locations with less amplitude are penalized more than locations with larger amplitude. Then a new inner iteration begins. Through this iterative refinement, the estimated source distribution gradually narrows, forming a clear boundary of estimation around the source distribution. This approach, known as extended source estimation, helps delineate the true source distribution more distinctly from the surrounding background activity.

[0032] The final source localization is represented as the distribution of current source intensity across cortical locations. The cortex is modelled as a triangulated mesh grid, where each triangular element is assigned a color based on a color bar scale: higher values (stronger source activity) may be represented by, for example, red, while lower values may be represented by, for example, orange, and locations with zero estimated activation represented by, for example, grey. This visualization makes it easier to interpret the results.Attorney Docket: 8350.2025-118WO

[0033] The proposed spatial-temporal-spectral imaging (STSI) framework contains several steps as explained above: 1) a pipeline to identify and delineate components representing biomarkers with different spectral profiles (such as low-frequency interictal spikes and high-frequency HFOs), and 2) a source imaging algorithm that minimizes the source sparsity and edge sparsity to estimate the location, extent, and dynamics of the underlying sources that produce these biomarkers, in an objective and automatic manner.

[0034] The activity at each brain region provides a time-frequency representation (TFR). Because of the quasi-static approximation of the Maxwell equation, such a timevarying spectrum could be linearly translated to a topographic distribution on the sensor space of the EEG / MEG oscillatory activity, at each time segment. In short, electrical / magnetic activity recorded by EEG / MEG can be modelled as a third-order tensor of space, time, and frequency. By using tensor decomposition techniques, such a tensor is decomposed into distinct components, which represent the EEG / MEG activity in a specific area (space), during a specific time window (time), and within a specific frequency band (frequency). The method puts this tensor-decomposed structure into a LI norm optimization problem that minimizes the edge sparsity and source sparsity, to obtain the source imaging solution.

[0035] The technical details of the multiple steps of the Spatial-Temporal-Spectral Imaging (STSI) framework will now be described.Attorney Docket: 8350.2025-118WO

[0036] FIG. 1 is a schematic illustration of the concept and framework. Biomarkers with diverse morphological features, for example, HFOs (high frequency oscillations) riding on spikes, spike overlapping with HFOs (where HFOs and spikes co-occur), and / or isolated HFOs and seizures mixed with noise, are input into the spatial- temporal-spectral framework. This framework first converts multi-channel temporal signals into a three-dimensional tensor comprising space, time, and frequency dimensions. Next, components that correspond to relevant biomarkers are identified, and an iterative source imaging algorithm estimates the location, extent, and temporal dynamic of the underlying sources.

[0037] The first step is the tensor decomposition of EEG / MEG data. The third-order tensor time-frequency representation (TFR) of EEG / MEG is decomposed into its corresponding components:

[0038] Where represents the third-order tensor from EEG / MEG measurements, whose entries are the continuous wavelet transform (CWT) coefficients parameter of a specific channel, time, and frequency band. Mathematically, it could be decomposed into the summation of outer products of spatialtemporal (tCh,n) and spectral (fCh.,n) components. The subscript 'ch' indicates that these values are obtained from the sensor data. Various algorithms exist for tensor decomposition. In one embodiment, canonical polyadic decompositionAttorney Docket: 8350.2025-118WO(CPD) is employed, because it decomposes a tensor into the sum of rank-one tensors, which matches the problem formulation.

[0039] FIG. 2 illustrates the concept of CPD. Another widely used method is Tucker decomposition which considers the interaction of components within a decomposition.

[0040] The spectral information of the source current dipole at each location, time, and frequency band makes up a three-dimensional tensor, and it could be decomposed into the outer product of the spatial component j / , temporal componentand spectral component ft. The inverse algorithm consists of solving the following optimization problem, which fits the scalp EEG / MEG measurement cf> while minimizing the source sparsity and edge sparsity:

[0041] In Eq. 3, (p is a third-order tensor of EEG / MEG whose entries are the amplitude or real / imaginary parts of the CWT. The L operator is the transform function that maps the source space third-order tensor to the sensor space, and the (?) operator denotes the outer product. If viewing by each frequency slice, thenAttorney Docket: 8350.2025-118WO such operator is simply a matrix multiplication of source level activity(Nsrcby Ntand the leadfield matrix, as illustrated in FIG. 3. Such operation is linear and thus does not affect the convexity of the optimization problem. Ncis the number of selected components, which was determined by visually inspecting the topography of sch n.

[0042] At a high level, the L operator is a set of matrix multiplications that transform the source space tensor into the sensor space tensor, and its design is rooted in the forward modeling of EEG. The details of the L operator are as follows.

[0043] Leadfield matrix: To understand the L operator, an explanation of the leadfield matrix is first necessary. In EEG forward modeling, each cortical dipole's activity produces a measurable EEG signal at the scalp. Due to the quasi-static approximation of Maxwell's equations, this relationship is linear and can be expressed by fixed weights. The collection of these weights for every dipoleelectrode pair forms the leadfield matrix. The spatial-temporal evolution of EEG signal could be expressed by the matrix multiplication of source activity and the leadfield matrix. This is a linear mapping.

[0044] Extension to the tensor domain: When considering the entire brain, cortical dipoles produce activity that evolves over time and across frequencies. This multidimensional data is naturally represented as a third-order tensor, with dimensions corresponding to space (cortical locations), time, and frequency, as previously discussed. Similarly, the EEG measurements across sensors, time, and frequency also form a third-order tensor. Mathematically, a method is needed toAttorney Docket: 8350.2025-118WO represent the mapping from cortical dipole activity tensor to the EEG tensor, which is where the £ operator comes into play.

[0045] The £ operator generalizes the concept of matrix multiplication to tensors. In this context, the £ operator applies the leadfield matrix along the spatial mode of the source space tensor. In other words, at a specific time and a specific frequency, the activity of EEG could be expressed by the leadfield matrix times the source activity. Using mathematical terminology, the operator performs a mode-1 product (if the spatial dimension is treated as the first mode), thereby mapping the source space tensor to the sensor space tensor.

[0046] In summary, the £ operator essentially maps the temporal-spectral evolution from each of the dipole-EEG activity pairs. It leverages the fact that the forward model is inherently linear, allowing the same weights (from the leadfield matrix) to be applied across the additional time and frequency dimensions.

[0047] However, the transformation into tensor operator adds complexity in terms of solving the optimization problem, as tensors are not as convenient as matrixes / scalars in operations such as projecting to hyper-ellipsoid. To solve this issue, the tensor representation of constraint can be rewritten into several matrix representations of constraint, as shown in FIG. 4. Specifically, the norm of the difference between original and reconstructed EEG / MEG tensor slice at each frequency should be less than the noise.

[0048] This optimization problem is solved with tnand fnestimated using the EEG / MEG measurement components tch nand fch n. It should be noted that tnand tch n,Attorney Docket: 8350.2025-118WO and „and fch nare not equal in a one-to-one manner. Nevertheless, because the brain dipole source strength and the EEG / MEG signals it produces generally follow a linear relationship under the assumption of a quasi-static approximation of Maxwell's equations, tnand fncould be expressed as the multiplication of a transformation matrix and tch nand fChn. In other words, tnand tch n, and fnand fch ncontains the same essential elements to represent the underlying brain activity.

[0049] To solve the optimization problem, an iterative process was performed wherein the j is first estimated from the unconstrained objective function described in equation (2), then the solution is projected to the hyper-ellipsoid representing the constraint. This process is shown in FIG. 5. When multiple frequencies are selected, each frequency corresponds to a constraint, and the projection needs to be in the direction of the intersection of hyper-ellipsoids representing each constraint. In one embodiment, the projection is done using the averaged projection, where the correction vector corresponding to each constraint is averaged to produce the overall projection vector. An illustration of averaged projection is shown in FIG. 6B.

[0050] In an alternate embodiment, alternative projection, shown in FIG. 6A, is used, where the solution is iteratively projected onto each hyper-ellipsoid in a sequential manner (e.g., if there are two constraints, then the projection is done first to the hyper-ellipsoid representing constraint 1, then to the hyper-ellipsoid representing constraint 2, and then to the hyper-ellipsoid representingAttorney Docket: 8350.2025-118WO constraint 1, etc.) The mathematical problem of using alternating projection or averaged projection to find a point in the intersection of a finite number of closed convex sets, where the constraints are hyper-ellipsoids (i.e. closed sets) has been shown to converge. In other words, in STSI, the convergence of finding the solution j that both minimizes the objective function and satisfies the constraints is found by projecting the unconstrained solution to the intersect of these constraints using either alternative projection or average projection. The number of iterations affects the source estimation solution. Specifically, in each iteration, the number of steps corresponds to the attempts made to project the solution onto the hyper-ellipsoid. At the end of each iteration, solutions with low amplitude are penalized.

[0051] The input to the optimization problem is the tensor components, which could be the complex number CWT coefficients, its amplitude, or the real / imaginary part of the CWT. For the last scenario, the real and imaginary part solution should be summed up to obtain the final solution. Specifically:Jn, complex Jn,re + ' Jn,im(4)

[0052] Where i is the imaginary unit, jn reand jn imare the solutions obtained from the real and imaginary parts of the CWT. Thus, the amplitude and phase of the corresponding oscillatory source could be determined by taking the modulus and angle of the complex number jniCOmpiex-Attorney Docket: 8350.2025-118WO

[0053] It is possible to perform tensor decomposition on the complex number transformed 3D EEG tensor. The presence of complex numbers, however, adds additional challenges to both the interpretation of components and the iterative solving procedure.

[0054] In one embodiment, the real part of the CWT coefficient is used as the tensor component for the source imaging problem. Mathematically, this approach is supported by the single integral formula theory, which posits that the signal can be reconstructed by summing the scaled CWT coefficients across all scales of the real part of the CWT, provided that an analytic wavelet, such as the Morlet wavelet used herein, is employed. In other words, the CWT allows a signal to be decomposed into components at various scales. To reconstruct the original signal, one sums these components in a manner that effectively "integrates" over the scales:

[0055] Where (t) is the real-valued signal such as EEG, Re{} denotes the real part,is the wavelet, the inner product < f(t), H'i(t) > denotes the wavelet coefficient at time t and scale a, capturing the signal's content at that resolution and moment, and C™ris the wavelet admissible constant. This means that the real part of the CWT coefficients contains all the necessary information to reconstruct the original signal to perform source imaging. This providesAttorney Docket: 8350.2025-118WO convenience, as we do not need the imaginary part of the CWT coefficients to reconstruct a real-valued signal such as EEG.

[0056] An example of HFO riding spike EEG data and its tensor components obtained using the abovementioned method is illustrated in FIGS. 7(A-E). FIG. 7A shows the example HFO riding spike of 75 channel by 500 ms plotted together (called a butterfly plot). FIG. 7B shows the channel where HFO was observed, a slight bump could be seen in the raw signal within the rectangle box, which is the HFO signal. A clearer high frequency oscillation could be seen in the filtered signal to the right.

[0057] FIG. 7C illustrates two example spike components, with focused topography and frequency peak around the spike frequency (10Hz). It should be noted that because the spike event contains a relatively wide range of frequency contents, and the preprocessing pipeline transforms the raw data into the summation of oscillations, a spike activity will in general be represented by multiple components, each with the oscillatory pattern of a specific frequency. The framework takes into consideration that multiple components might be selected as the input. FIG. 7D illustrates the HFO component, with focused topography, HFO shaped oscillatory activity, and a frequency peak around the ripple range (80-250Hz). In one embodiment, the process delineates the frequency content of HFOs without requiring manual filtering. FIG. 7E illustrates an example noise component, with unfocused spatial topography (indicating that there is no focused source), and time course without any clear pattern.Attorney Docket: 8350.2025-118WO

[0058] A validation study of the STSI framework was performed using a cohort of drugresistant epilepsy patients. The STSI framework provided superior performance for source imaging of HFOs co-occurring with spikes, a biomarker of EZ.

[0059] The study focused on events where HFOs co-occur with interictal spikes in EEG. These events are referred to herein as pSpikes (HFO-riding spike) and pHFOs (spike-overlapping HFOs) for simplicity. This type of event is of particular interest, as prior research suggests that both pSpike and pHFO events serve as biomarkers that show an enhanced ability to delineate EZ. The STSI framework is well-suited for analysing co-occurring HFOs and spikes, as it can effectively separate their respective components and perform source imaging with extent estimation. FIG. 8 illustrates the STSI source imaging performance of 132 pHFOs and 132 pSpikes from a cohort of 30 patients, compared with benchmark algorithms sLORETA and LCMV. Across all evaluations, STSI demonstrated superior accuracy in estimating the source location and extent of biomarkers with distinct frequency profiles. FIG. 8A shows EEG topography and butterfly time series of spike and HFO events in a patient with focal epilepsy. FIG. 8B shows source imaging results for this example patient using STSI, sLORETA, LCMV methods.

[0060] For spike analysis, the group average metrics for STSI, sLORETA and LCMV are LE (mm): 9.1 ± 11, 15.72 ± 6.16, 21.41 ± 5.48; SD (mm): 11.06 ± 13.94, 31.23 ± 12.55, 44.3 ± 10.71. The difference was significant (linear mixed effect model, p<0.001). All following comparisons were using the same statistical comparisonAttorney Docket: 8350.2025-118WO method of linear mixed effect model. Precision: 0.46 ± 0.36, 0.13 ± 0.1, 0.07 ± 0.05; Sensitivity: 0.63 ± 0.3, 0.89 ± 0.2, 0.93 ± 0.17; Specificity: 0.97 ± 0.03, 0.75 ± 0.07, 0.5 ± 0.08; Fl: 0.43 ± 0.26, 0.21 ± 0.14, 0.12 ± 0.09. The difference was significant with p<0.001, except for sensitivity in sLORETA vs LCMV (p = 0.066). The STSI spike source imaging outperformed sLORETA and LCMV across all metrics, except for sensitivity (where sLORETA had a slight edge). Notably, STSI exhibited significantly superior specificity compared to sLORETA and achieved a significantly enhanced Fl score, reflecting its ability to balance both sensitivity and specificity. This indicates that STSI provides a more reliable and precise characterization of spike sources in comparison to the benchmark methods.

[0061] For HFOs analysis, the group average metrics are LE (mm): 7.58 ± 9.56, 24.74 ± 8.8, 22.55 ± 5.85; SD (mm): 9.31 ± 11.59, 52.69 ± 16.71, 46.46 ± 10.82. These comparisons showed significance with p<0.001. Precision: 0.51 ± 0.36, 0.09 ± 0.07, 0.06 ± 0.05; Sensitivity: 0.61 ± 0.31, 0.88 ± 0.2, 0.86 ± 0.2; Specificity: 0.97 ± 0.03, 0.65 ± 0.14, 0.53 ± 0.06; Fl: 0.45 ± 0.26, 0.16 ± 0.11, 0.11 ± 0.08. The difference was significant with p<0.001, except for sensitivity in sLORETA vs LCMV (p = 0.51). The details of each performance metric and statistical analysis can be found in FIG. 8C,8D. Similar to spike source imaging, STSI provides a more reliable and precise characterization of HFO sources in comparison to the benchmark methods.

[0062] These results demonstrate that, in events where HFO co-occur with interictal spikes, STSI exhibited an enhanced ability in estimating both the location andAttorney Docket: 8350.2025-118WO extent of the underlying sources for both pSpikes and pHFOs. Specifically, STSI achieved overall lower localization error and spatial dispersion, signifying a more precise estimation of source locations. Additionally, STSI demonstrated higher precision and higher specificity, indicating its ability to provide more accurate and confined source estimations compared to sLORETA and LCMV with Otsu thresholding. The latter approaches typically yield widespread and less focal solutions. Furthermore, the Fl score, which balances sensitivity and specificity, was significantly higher for STSI (even with somewhat lower sensitivity of STSI compared to sLORETA). These results highlight that STSI offers a more robust and reliable estimation of source extent compared to the benchmark algorithms.

[0063] Some embodiments may include imaging bioelectric activity within organ systems from surface recorded electrophysiological signals. Some embodiments may include imaging bioelectric sources within organ systems from electrophysiological signals recorded within the organ systems. Some embodiments may include localizing and imaging brain electric sources from electrophysiological measurements including, but not limited to, electroencephalography, magnetoencephalography, or intracranial electroencephalography. Some embodiments may use signal separation of electrophysiological signals into various frequency components and may estimate source distributions corresponding to various frequency components in recordings using a single spatial-temporal-spectral imaging (STSI) framework. Some embodiments may enable source imaging of low frequency events,Attorney Docket: 8350.2025-118WO including, but not limited to, individual epileptic spikes or evoked potentials, seizure oscillations, and high frequency oscillations, all using a single STSI approach.

[0064] Some embodiments may be used in various biomedical research and clinical applications, including, but not limited to, epilepsy source localization and imaging, brain source imaging and localization, cardiac arrhythmia source localization and imaging, and muscular source imaging and localization. Some embodiments may be applied in areas including, but not limited to, epilepsy source localization, brain source imaging, seizure localization, cardiac source imaging, and cardiac arrhythmia source localization and imaging. Some embodiments may have applications including, but not limited to, imaging brain sources from EEG or MEG, cardiac sources from ECG or MCG, and muscular activity from EMG.

[0065] Some embodiments may use an array of body surface electrodes to record ECG signals over the body surface for cardiac source imaging and localization of origin of cardiac arrhythmia, cardiac activation sequence, or myocardial infarction and ischemia.

[0066] Some embodiments may use an array of electrodes in a catheter to record electrical potentials over an endocardial surface or within a blood cavity for cardiac source imaging of cardiac abnormalities, and localization of origin of arrhythmia.Attorney Docket: 8350.2025-118WO

[0067] The invention is explained in the context of brain electric source imaging using EEG / MEG, but the approach can be extended to other systems, such as organ systems including, but not limited to, cardiac electric source imaging from ECG or MCG.

[0068] Herein, epilepsy source imaging is discussed as a non-limiting example, which may illustrate embodiments for EEG based brain source imaging. While various embodiments may be described in detail in the example for epilepsy source imaging, including imaging of spikes, HFOs and seizures, the STSI framework may be extended to areas including, but not limited to, epilepsy source imaging from MEG measurements where magnetic sensors are used to record MEG signals, brain source imaging using EEG or MEG for non-epileptic brains including healthy brains or brains with other neurological disorders. Applications in some embodiments may be extended to cardiac source imaging from ECG or MCG recordings for localization of ventricular and atrial arrhythmias, cardiac infarctions, cardiac ischemia, and other heart diseases. Applications in some embodiments may be extended to muscular source imaging from EMG for localization and imaging muscular electrical activity. Those skilled in the art will understand the concepts of the disclosure and will recognize applications of these concepts not particularly addressed herein.

[0069] In some embodiments, the invention can be used to image brain electrical activity associated with physiological processes of brain activation, from EEG or MEG, to understand how normal brain processes work. In such embodiments,Attorney Docket: 8350.2025-118WO evoked potentials or event-related potentials, physiological HFOs and neural oscillations at various frequencies may be used to study brain functions.

[0070] In some embodiments, the invention can be used to image brain pathologies other than epilepsy, such as those associated with stroke, pain, mental disorders and other neurological disorders.

[0071] In some embodiments, the invention can be used to image and localize brain activity to guide neuromodulation interventions by localizing brain response to neuromodulation, and then adjust and refine the neuromodulation targets and parameters.

[0072] Embodiments implementing methods and systems for brain source imaging, cardiac source imaging or muscular source imaging may be derived where electrophysiological signals may be recorded using an array of electrodes on areas including, but not limited to, the scalp, the chest and body surface of muscles, or an array of magnetic sensors over and close to areas including, but not limited to, the scalp, chest and body surface.

[0073] The framework 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.

Claims

Attorney Docket: 8350.2025-118WOClaims1. A method for localizing and imaging a source of bioelectrical activity in a biological system comprising: obtaining high-dimensional electrophysiological data from the biological system; decomposing the high-dimensional electrophysiological data into one or more three-dimensional tensors; selecting components of the one or more tensors representing one or more biomarkers of interest having distinct spectral profiles; estimating the location, extent, and temporal dynamics of the source within the biological system that generated the selected components; and imaging the source of the selected components within the biological system.

2. The method of claim 1 wherein the one or more three-dimensional tensors each comprise: a spatial component representing activity in a specific area of the biological system; a temporal component representing activity during a specific time window; and a spectral component representing activity within a specific frequency band.Attorney Docket: 8350.2025-118WO3. The method of claim 1 wherein the high-dimensional electrophysiological data represents brain activity of the subject.

4. The method of claim 3 wherein the high-dimensional electrophysiological data is collected via an EEG or MEG.

5. The method of claim 1 wherein the components of the tensors representing the biomarkers of interest are selected by a user.

6. The method of claim 5 further comprising: applying a set of data-driven rules based on spatial topography, temporal evolution, and spectral frequency characteristics to guide the user through the process of selecting the components.

7. The method of claim 1 wherein estimating the location, extent, and temporal dynamics of the selected components further comprises: performing an iterative inverse estimation using selected components of the sensor space tensor as the input; and reconstructing the source corresponding to the estimation; wherein the decomposed high-dimensional electrophysiological data defines a sensor space tensor.Attorney Docket: 8350.2025-118WO8. The method of claim 7 wherein estimating the location, extent, and temporal dynamics of the selected components further comprises: iteratively calculating an unconstrained solution and projecting the unconstrained solution to a hyper-ellipsoid representing a constraint until a stopping criterion is met; applying weighting coefficients to favor current source locations having a larger amplitude; and repeating the process until a clear boundary of the localization emerges.

9. The method of claim 7 wherein the stopping criterion is either a predetermined number of iterations or when a difference between two consecutive iterations is smaller than a predetermined threshold.

10. The method of claim 7 wherein the imaging of the source is represented as a distribution of current dipole sources across locations within the biological system.

11. The method of claim 7 wherein the imaging step further comprises: modelling the biological system as a triangulated mesh grid; and assigning a color to each triangulation based on the dipole source intensity; andAttorney Docket: 8350.2025-118WO displaying the colored mesh grid.

12. The method of claim 1 wherein the biomarkers of interest are high frequency oscillations overlapping interictal spikes.

13. The method of claim 12 wherein the biomarkers are used to localize an epileptogenic zone in the brain of a subject.

14. A system for localizing and imaging a source of bioelectrical activity in a biological system comprising: a device for collecting bioelectromagnetic measurement; a processor; a display; high-dimensional electrophysiological data collected from the biological system using the device; and one or more computer-readable media storing instructions that, when executed by the processor, cause the system to: select components of the one or more tensors representing one or more biomarkers of interest having distinct spectral profiles; estimate the location, extent, and temporal dynamics of the source within the biological system that generated the selected components; andAttorney Docket: 8350.2025-118WO image the source of the selected components within the biological system.

15. The system of claim 14 wherein the high-dimensional electrophysiological data is EEG or MEG data, further comprising: a plurality of sensors; the instructions further causing the system to: record the EEG or MEG data using the plurality of sensors.

16. The system of claim 14 wherein the plurality of sensors are electrodes used to record EEG data.

17. The system of claim 14 wherein the plurality of sensors are magnetic sensors used to record MEG data.

18. The system of claim 14 wherein the biological system is an epileptic human.

19. The system of claim 14 wherein the plurality of sensors includes a plurality of body surface electrodes for imaging and localizing cardiac electrical activity.Attorney Docket: 8350.2025-118WO20. The system of claim 19 wherein the high-dimensional electrophysiological data comprises body surface ECG used for localizing and imaging an origin of cardiac arrhythmia in a patient with cardiovascular diseases.

21. The system of claim 19 wherein the high-dimensional electrophysiological data comprises electrical potentials from a cardiac catheter used for localizing and imaging an origin of cardiac arrhythmia in a patient with cardiovascular diseases.

Images