High-frequency oscillation fusion micro-potential epilepsy focus precise positioning system

Abstract

The application belongs to the technical field of epilepsy focus positioning, and relates to a high-frequency oscillation fusion micro-potential epilepsy focus precise positioning system. High-frequency oscillation characteristics and micro-potential fine waveform details of HFOs are independently extracted through double-channel parallel processing. Then, only HFOs and micro-potential combinations coexisting in a preset time window and spatial range are reserved. The correlation mode of the coupled events is extracted, and the weight is distributed according to the characteristic intensity, so that high-confidence events dominate positioning. The fuzzy positioning of HFOs is restricted by the smaller diffusion range of micro-potential, the coupled events are mapped to three-dimensional coordinates of electrode contacts, and finally, the probabilistic three-dimensional modeling is used to convert the weighted events into a brain area epileptogenic probability distribution map through spatial interpolation, and the threshold probability connected region is used as the core epileptogenic focus. The application uses the high-resolution characteristics of micro-potential to correct the positioning ambiguity of HFOs, and enhances specificity through collaborative event screening, so as to realize precise three-dimensional positioning of the epilepsy focus on the basis of retaining high sensitivity of HFOs.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of epileptogenic focus positioning, more specifically, it relates to a high-frequency oscillation fusion micro-potential epileptogenic focus precise positioning system. BACKGROUND

[0002] The success of surgical treatment of drug-refractory epilepsy is highly dependent on the precise positioning of the epileptogenic focus; resection or disconnection of brain tissue containing the epileptogenic core is the key to radical treatment; intracranial electroencephalogram, with its high temporal resolution and high signal-to-noise ratio, has become the gold standard method for positioning the epileptogenic focus; traditionally, clinicians mainly rely on visual recognition of seizure period discharge patterns and interictal epileptiform discharges for positioning.

[0003] In recent years, high-frequency oscillations (HFOs, usually referring to >80Hz brain electrical activity, especially Ripple 80-250Hz and Fast Ripple >250Hz) have been widely recognized as more sensitive and specific epileptogenic biomarkers than IEDs due to their high correlation with epileptogenic brain tissue and specific expression before / during seizures; in the interictal period, the area producing HFOs is considered to be closer to the epileptogenic core, however, the actual clinical application of HFOs faces significant challenges, as follows:

[0004] Not all detected HFOs are epileptogenic, physiological HFOs and passive HFOs generated by distant epileptogenic foci are widespread; existing automatic detection algorithms based on energy, frequency or morphology are difficult to reliably distinguish between epileptogenic and non-epileptogenic HFOs, resulting in high false positive rates and making it difficult to accurately locate the true epileptogenic core; secondly, the signal energy of HFOs is mainly concentrated in the high frequency band, its spatial diffusion range is relatively large, and it is easily affected by volume conduction effects, making it difficult to accurately locate the starting point of a small epileptogenic focus based on a single HFO event, and it is difficult to clearly define the small boundary of the epileptogenic focus.

[0005] Current mainstream research focuses on the optimization of a single biomarker, improving detection algorithms, exploring features to distinguish between epileptogenic / non-epileptogenic HFOs, and improving HFOs positioning accuracy, etc., however, the inherent positioning specificity deficiency and limited spatial resolution of HFOs make it difficult to fundamentally solve the problem through deep mining of its own characteristics, in addition, although the current micro-potential has excellent spatial resolution potential, its extremely low signal-to-noise ratio and lack of direct correlation with epileptogenic evidence make it difficult to be used as a reliable positioning marker alone; therefore, how to overcome the defects of HFOs and micro-potential while fully utilizing their complementary advantages to achieve more specific and higher spatial resolution epileptogenic focus positioning has become a technical problem that needs to be solved. SUMMARY

[0006] The application provides a high-frequency oscillation and micro-potential fused epilepsy focus precise positioning system, aims to effectively fuse two biomarkers of HFOs and micro-potential, and overcome the technical problems of insufficient positioning specificity and limited spatial resolution when HFOs are used alone.

[0007] The high-frequency oscillation and micro-potential fused epilepsy focus precise positioning system comprises a signal acquisition module, a double-channel preprocessing module, a feature extraction module, a coupled event processing module, a spatial positioning and weighting module, and an epilepsy focus core determination module.

[0008] The signal acquisition module is used to acquire intracranial electroencephalogram signals collected by electrode contacts.

[0009] The double-channel preprocessing module comprises a filtering unit, an artifact and noise removal unit, and a baseline correction unit, the filtering unit is used to perform double independent channel filtering on the intracranial electroencephalogram signals, the artifact and noise removal unit is used to remove artifacts and noise in the filtered signals, and the baseline correction unit is used to perform baseline correction on the signals after artifact removal and noise removal, finally outputting a first clean signal suitable for high-frequency oscillation analysis and a second clean signal suitable for micro-potential analysis.

[0010] The feature extraction module comprises a high-frequency oscillation feature extraction unit and a micro-potential feature extraction unit, the high-frequency oscillation feature extraction unit is used to perform event detection, cluster denoising and feature extraction on the first clean signal, and output a high-frequency oscillation event list; the micro-potential feature extraction unit is used to perform signal-to-noise ratio enhancement, event detection and feature extraction on the second clean signal, and output a micro-potential event list.

[0011] The coupled event processing module is used to align the high-frequency oscillation event list and the micro-potential event list on a time axis, screen event pairs meeting coupling conditions through spatio-temporal correlation analysis, extract joint features of coupled events and exclude pseudo-coupled events, and output an effective coupled event set.

[0012] The spatial positioning and weighting module is used to map each effective coupled event to a corresponding electrode contact three-dimensional space coordinate, give a basic weight based on the event type, and dynamically adjust the weight combined with the event features, and output a coupled event space distribution map with weights.

[0013] The epilepsy focus core determination module is used to generate an epileptogenic probability distribution map on a three-dimensional brain structure model of a patient based on the coupled event space distribution map, and determine an epilepsy focus core area by setting a probability threshold.

[0014] In the present application, the high-frequency oscillation characteristics and fine waveform details of micro-potentials of HFOs are independently extracted through double-channel parallel processing, signal interference is avoided, and only HFOs and micro-potentials co-occurring in a preset time window and spatial range are reserved, isolated HFOs events unrelated to epilepsy are filtered out, specificity is significantly improved, the correlation mode of coupled events is extracted again, and the high-confidence events dominate positioning according to the feature intensity; the diffusion range of micro-potentials is smaller, the fuzzy positioning of HFOs is restricted, the coupled events are mapped to the three-dimensional coordinates of the electrode contacts, and the spatial resolution limit of single HFOs is broken through; finally, the probabilistic three-dimensional modeling is used to convert the weighted events into a brain region epileptogenic probability distribution map through spatial interpolation, and the threshold probability connected region is used as the core epileptogenic focus; therefore, the present application uses the high-resolution characteristics of micro-potentials to correct the fuzzy positioning of HFOs, and enhances the specificity through cooperative event screening, thereby realizing precise three-dimensional positioning of the epileptogenic focus on the basis of retaining the high sensitivity of HFOs.

[0015] Preferably, each unit of the double-channel preprocessing module is configured as follows:

[0016] The filter unit is configured with two independent filter channels, the passband of the first channel is 80Hz to 500Hz, and the stopband attenuation is not less than 40dB; the passband of the second channel is 100Hz to 1000Hz, and the stopband attenuation is not less than 50dB; both channels use bidirectional filtering to eliminate phase shift, and output high-frequency oscillation detection signal streams and micro-potential detection signal streams, respectively;

[0017] The artifact and noise removal unit first decomposes the two signal streams output by the filter unit into a plurality of independent signal components using independent component analysis, selects and removes artifact components with abnormal kurtosis and power spectrum not conforming to the characteristics of electroencephalogram, and reconstructs the remaining components into de-artifact signals; the de-artifact signals corresponding to the micro-potential detection signal stream are further decomposed into approximation coefficients and detail coefficients using 8-layer wavelet decomposition, the detail coefficients are subjected to soft threshold processing, and the approximation coefficients are reconstructed again to obtain intermediate signals with enhanced signal-to-noise ratio;

[0018] The baseline correction unit sets a sliding window for the de-artifact signals of the high-frequency oscillation detection signal stream and the intermediate signals of the micro-potential detection signal stream, respectively, wherein the window width of the first channel is greater than that of the second channel, the local average value of each time point is calculated as the corresponding baseline through the sliding window, and the original signal is subtracted from the baseline signal to output the first clean signal and the second clean signal, respectively.

[0019] Preferably, the filter unit uses an FIR filter to realize bidirectional filtering, and the signal delay is compensated through a zero-phase offset algorithm during the filtering process to ensure the time synchronization of the high-frequency oscillation detection signal stream and the micro-potential detection signal stream.

[0020] Preferably, in the baseline correction unit, the sliding window width of the first channel is 50ms to 100ms, the sliding window width of the second channel is 20ms to 50ms, and the window sliding step size is 1ms. The baseline value at each time point is calculated by a local weighted average algorithm.

[0021] Preferably, the high-frequency oscillation feature extraction unit calculates the time spectrum in different time periods and filters candidate intervals by combining the baseline signal energy threshold. Within the candidate intervals, oscillation events are detected by the Teager energy operator. After morphological examination to eliminate non-oscillatory artifacts, the time, frequency domain, energy, and morphological features of the events are extracted. Then, outliers are eliminated through cluster analysis to obtain a list of high-frequency oscillation events.

[0022] Preferably, the micropotential feature extraction unit sets an energy threshold by combining a short-time energy sequence with the background noise level, calculates the instantaneous slope sequence obtained by the first-order difference of the signal and sets a slope threshold, filters candidate events that simultaneously meet the requirements of energy threshold, slope threshold, minimum duration and peak amplitude, retains events with repetitive waveform characteristics and extracts relevant features to form a micropotential event list.

[0023] Preferably, the spatiotemporal correlation analysis of the coupled event processing module includes:

[0024] Define a micropotential correlation time window for high-frequency oscillation events, and screen time-correlated event pairs whose peak time falls within the time window and whose time difference meets the physiological coupling range; calculate the three-dimensional spatial distance between the electrode contacts corresponding to the time-correlated event pairs, and screen spatially correlated event pairs whose distance does not exceed a set threshold as high-frequency oscillation-micropotential coupling events.

[0025] Preferably, the coupling event processing module extracts four types of quantitative features of the coupling events: standardized latency, amplitude coupling strength, waveform morphology correlation, and spatial coupling strength, forming a four-dimensional feature vector. Then, it uses a density clustering algorithm to remove isolated clusters with sample sizes less than a set threshold to obtain an effective set of coupling events.

[0026] Preferably, the dynamic weight adjustment of the spatial positioning weighting module includes a time tightness factor, an amplitude correlation factor, a time frequency factor, and a high-frequency oscillation subtype factor.

[0027] The time tightness factor is negatively correlated with the time difference between the peak value of the micropotential and the onset of the high-frequency oscillation.

[0028] The amplitude correlation factor is obtained based on the correlation coefficient conversion between the high-frequency oscillation signal amplitude sequence and the micro-potential signal amplitude sequence.

[0029] The time frequency factor is the natural logarithm of the number of coupled events per unit time.

[0030] The high-frequency oscillation subtype factor sets different values according to spectral characteristics.

[0031] Preferably, the epilepsy focus core determination module creates a regular three-dimensional grid covering the whole brain, calculates the spatial distance of each grid point to all electrode contacts, aggregates and normalizes the contact weight after distance attenuation, generates the epileptogenic probability value and maps it into a probability cloud map, sets a dynamic threshold based on the mean and standard deviation of the whole brain probability, marks the connected region higher than the threshold and eliminates isolated noise points, and determines the epilepsy focus core region.

[0032] The beneficial effects of the present application include:

[0033] In the present application, the high-frequency oscillation characteristics of HFOs and the fine waveform details of micro-potentials are independently extracted through double-channel parallel processing to avoid signal interference. Secondly, only the combination of HFOs and micro-potentials that coexist within the preset time window and spatial range is retained, and isolated HFOs events that are not related to epilepsy are filtered out to significantly improve specificity. Then, the correlation mode of the coupled events is extracted, and the weight is assigned according to the feature strength, so that high-confidence events dominate the positioning. Again, the diffusion range of micro-potentials is smaller, which restricts the fuzzy positioning of HFOs, maps the coupled events to the three-dimensional coordinates of the electrode contacts, and breaks through the spatial resolution limit of single HFOs. Finally, the probabilistic three-dimensional modeling is used to convert the weighted events into a brain region epileptogenic probability distribution map through spatial interpolation, and the threshold is used to improve the probability connected region as the core epileptogenic focus. Therefore, the present application uses the high-resolution characteristics of micro-potentials to correct the fuzzy positioning of HFOs, and at the same time, through the screening of collaborative events, the specificity is enhanced, and on the basis of retaining the high sensitivity of HFOs, the precise three-dimensional positioning of the epilepsy focus is realized. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0035] Figure 1 The system block diagram provided by the embodiments of the present application.

[0036] Figure 2 The contrast schematic diagram before and after pre-processing provided by the embodiments of the present application.

[0037] Figure 3 The HFOs-micro-potential coupled event time correlation analysis schematic diagram provided by the embodiments of the present application.

[0038] Figure 4A clustering result schematic diagram provided by the embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the technical problems, technical solutions and beneficial effects to be solved in the present application clearer and more apparent, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0040] Referring to Figure 1 As shown in the figure, the high-frequency oscillation micro-potential epilepsy focus precise positioning system provided by the present application realizes efficient fusion and precise positioning of biomarkers through multi-module cooperation; the overall architecture is as shown in the figure, which comprises a signal acquisition module, a double-channel preprocessing module, a feature extraction module, a coupling event processing module, a spatial positioning weighting module and an epilepsy focus core determination module connected in sequence; the internal units of each module interact through a data bus, so as to ensure signal synchronization and data integrity. Figure 1

[0041] The signal acquisition module selects a multi-contact intracranial electrode, selects a stereoelectroencephalogram (SEEG) electrode or an electrocorticogram (ECoG) electrode, each electrode is equipped with 8-16 contacts, the contact spacing is flexibly adjusted within 3-10 mm according to the anatomical structure of the target brain region such as the temporal lobe and the frontal lobe; the electrode needs to cover the suspected epilepsy focus region and the normal brain tissue around it more than 2 cm, so as to realize comprehensive spatial sampling; at the same time, the electrode contact impedance is controlled below 5kΩ to prevent signal distortion caused by too high impedance.

[0042] In the present embodiment, during signal acquisition, the intracranial electroencephalogram signal is continuously acquired at a sampling rate of not less than 2000Hz, the dynamic range reaches 16 bits and above (±5mV), and video data is recorded synchronously for assisting in excluding motion artifacts; since the highest frequency of the micro-potential signal can reach 1000Hz, the sampling rate of 2000Hz meets the Nyquist sampling theorem and can completely retain the fine waveform of the micro-potential; the high dynamic range can accurately capture weak micro-potential signals, and the peak-to-peak value is more than 3 times the noise standard deviation; then the acquired original intracranial electroencephalogram signal (including synchronous time stamp) is transmitted to the double-channel preprocessing module, and the original data is stored at the same time, which is convenient for subsequent traceability analysis.

[0043] The double-channel preprocessing module is the key to realize the separation of HFOs and micro-potential signals, which is composed of a filtering unit, an artifact and noise removal unit and a baseline correction unit, and specifically as follows:

[0044] ​Filtering unit: two independent finite impulse response (FIR) filtering channels are configured, channel 1 is used for HFOs analysis, the passband is set to 80Hz-500Hz, and the stopband attenuation is not less than 40dB; channel 2 is used for micro-potential analysis, the passband is 100Hz-1000Hz, and the stopband attenuation is not less than 50dB; since the core frequency band of HFOs is 80Hz-500Hz (including 80-250Hz ripple wave and 250-500Hz fast ripple wave), and the micro-potential signal is mainly concentrated in 100Hz-1000Hz, independent frequency band design can effectively avoid mutual interference of the two signals.

[0045] A bidirectional filtering algorithm (zero phase offset compensation) is adopted, which is implemented in accordance with the FIR filter standard; by combining forward filtering and reverse filtering, signal phase offset is eliminated, and the high-frequency oscillation detection signal stream output by channel 1 and the micro-potential detection signal stream output by channel 2 are completely synchronized in the time axis, laying a foundation for subsequent time correlation analysis of coupling events.

[0046] Artifact and noise removal unit: independent component analysis (FastICA algorithm) is performed on the two filtered signals respectively, and the signal matrix X (dimension C×T, C represents the number of channels, and T represents the time point) is decomposed into an independent component matrix S=WX (W is a demixing matrix), and the kurtosis and power spectrum of each independent component are calculated: ;

[0047] In the formula: represents the i-th independent component; represents the value of the i-th independent component at time point t; represents the mean value of the i-th independent component; T represents the number of signal time points; represents the kurtosis value of the i-th independent component;

[0048] In this embodiment, the power spectrum is used to describe the distribution of signal power with frequency, and in this system, by calculating the power spectrum of each independent component, the energy distribution characteristics of the signal at different frequencies can be analyzed, thereby providing key frequency information basis for accurate positioning of the epileptic focus;

[0049] When the kurtosis value is calculated, the components with kurtosis value greater than 8 or significant peak value at 50 / 60Hz frequency band are marked as physiological artifacts (such as electro-oculogram, electromyogram) or line noise, and after removing these components, the signal is reconstructed to obtain the artifact-removed signal.

[0050] For the de-artifact signal of channel 2, further 8-layer wavelet denoising processing is carried out, the signal is decomposed into 8 layers by selecting db4 wavelet basis, the detail coefficients are processed by soft threshold value (the threshold value is dynamically set according to noise estimation), and then the approximate coefficients are reconstructed, so as to obtain the intermediate signal with enhanced signal-to-noise ratio; through wavelet denoising, the present embodiment can effectively suppress residual noise while retaining high-frequency transient characteristics, and avoid that the amplitude of weak micro-potential signal is masked by noise.

[0051] Baseline correction unit: channel 1 (high-frequency oscillation detection signal stream) adopts a sliding window with a width of 50ms-100ms, channel 2 (intermediate signal) adopts a sliding window with a width of 20ms-50ms, and the window sliding step is 1ms, and the baseline value at each time point is calculated by local weighted average algorithm: ;

[0052] In the formula: represents the value of the signal at time point k in the window; represents the moving average window length; represents the half window size; represents the baseline signal at time point n;

[0053] In the present embodiment, different window widths are set to avoid excessive fluctuation of the baseline of the relatively flat HFOs signal; while the micro-potential signal changes rapidly, a narrower window can accurately track the baseline drift and prevent signal distortion.

[0054] The baseline correction formula is: , wherein is the original signal, is the local average baseline, and the final output is the first clean signal suitable for HFOs analysis and the second clean signal suitable for micro-potential analysis; the signal comparison before and after preprocessing is shown in Figure 2 , in which the abscissa represents time and the ordinate represents signal amplitude; after preprocessing, the low-frequency noise and spike artifacts of the original signal of channel A (corresponding to the high-frequency oscillation detection signal stream) are effectively removed, and the oscillation characteristics of HFOs are clearly presented; after preprocessing, the background noise of the original signal of channel B (corresponding to the micro-potential detection signal stream) is greatly reduced, and the transient peak characteristics of the micro-potential are accurately identified; it can be seen that after preprocessing, the artifacts and noise of the signal are obviously reduced, and the recognition degree of effective signal is significantly improved.

[0055] Feature extraction module: used for extracting effective event features from two clean signals, including high-frequency oscillation feature extraction unit and micro-potential feature extraction unit:

[0056] The high-frequency oscillation feature extraction unit divides the first clean signal into segments with a length of 40 ms, with a 50% overlap between adjacent segments, calculates the time-frequency spectrum of each segment, and extracts the total energy value in the frequency band of 80-500 Hz; an energy threshold is set based on the average energy and fluctuation range of the baseline signal, and the segments with a total energy value exceeding the threshold are selected as candidate HFOs intervals, wherein the energy threshold is obtained by a preset formula as follows: ;

[0057] In the formula: represents the average energy of the 80-500 Hz frequency band in the baseline period; represents the energy standard deviation in the baseline period; N represents an empirical coefficient; represents the energy threshold;

[0058] The Teager energy operator (TEO) is used to calculate the signal envelope in the candidate HFOs interval, and an amplitude threshold and a duration threshold of 10-100 ms are set: ;

[0059] ;

[0060] In the formula: represents the amplitude of the discrete signal at point n; represents the Teager energy value of the discrete signal at sample point n; represents the amplitude of the discrete signal at the previous sample point of sample n; represents the amplitude of the discrete signal at the next sample point of sample n; represents the amplitude threshold of time domain refinement detection; represents the mean value of the TEO signal envelope in the candidate interval; represents the standard deviation of the TEO signal envelope in the candidate interval; represents a dynamic coefficient;

[0061] The continuous segment that satisfies and has a duration D within [10 ms, 100 ms] is marked as an HFOs candidate event, i.e., a preliminary HFOs event segment;

[0062] For each preliminary HFOs event obtained, morphological examination is performed to exclude non-oscillatory artifacts; first, the number of times the signal crosses zero (i.e., the number of times the signs of adjacent sample points change) in the event segment is counted, and it is required that the zero-crossing rate of the event segment reaches at least 4 times per 10 ms (equivalent to 80 Hz oscillation):

[0063] ;

[0064] In the formula: T represents the event length; represents an indicator function; signal amplitude at the i-th sample point within the HFOs candidate event; signal amplitude at the i+1-th sample point within the HFOs candidate event

[0065] For kurtosis, calculate the kurtosis of the event segment signal (a measure of the shape of the signal distribution, reflecting the deviation from the normal distribution), requiring the kurtosis to be greater than 3:

[0066] In the formula: kurtosis value of the candidate time; mean value of the signal within the candidate event;

[0067] Based on the above non-oscillatory artifact rejection, the verified HFOs event segment is obtained;

[0068] For each verified HFOs event segment, the following features are extracted:

[0069] Time features: event start time (the first point in the event segment that exceeds the amplitude threshold), end time (the last point in the event segment that exceeds the amplitude threshold), peak time (the time point corresponding to the maximum value of the TEO envelope), duration (end time minus start time).

[0070] Frequency domain features: calculate the power spectrum within the event segment, the center frequency is the weighted average frequency of the power spectrum, and the bandwidth is the mean square deviation of the frequency distribution (reflecting the concentration of frequency).

[0071] Energy features: total energy (sum of squares) of the original signal within the event segment, and maximum value of the TEO envelope (instantaneous energy peak).

[0072] Morphological features: number of oscillation periods (calculated by zero-crossing rate, equal to half the number of zero-crossings multiplied by duration), peak-to-peak amplitude (difference between maximum and minimum values of the signal within the event segment), rising slope (amplitude change divided by time difference from the start of the event to the maximum point of the signal).

[0073] Then combine multiple features (center frequency, bandwidth, duration, peak-to-peak amplitude, kurtosis) of all events into a feature vector, use K-means clustering method, assuming that the class with the most number represents the true HFOs, and other classes are noise, then further exclude outliers in the cluster, i.e. calculate the Mahalanobis distance of each point to the class center, points exceeding the statistical threshold are considered outliers; after excluding outliers, the filtered HFOs events and their features are obtained, and then the HFOs event list is output.

[0074] ​Micro-potential feature extraction unit: Discrete wavelet transform is performed on the second clean signal (db4 wavelet basis is adopted, and decomposition is performed for 5 layers), high-frequency detail coefficients cD3-cD5 (corresponding to 100-500 Hz frequency band) are reserved, low-frequency coefficients and noise dominant layer coefficients are set to zero, and then inverse transform is performed to reconstruct an enhanced signal.

[0075] Event detection: A short-time energy sequence of the enhanced signal is calculated (the half window length W is dynamically set according to actual conditions), and an energy threshold is set according to the median absolute deviation (MAD) of the background noise; a first-order difference of the signal is calculated to obtain an instantaneous slope sequence, and a 95% quantile is taken as a slope threshold. Candidate events that meet the conditions of “energy exceeding the threshold, duration greater than 1 ms, slope exceeding the threshold sample point ratio more than 50%, and peak-to-peak amplitude greater than 3 times the noise standard deviation” are screened.

[0076] Feature extraction and screening: Time positioning features (start / peak / end time), amplitude features (peak amplitude, peak-to-peak amplitude), waveform features (rise / fall slope, asymmetry), and spectral features (center frequency, bandwidth) of the candidate events are extracted, false detection events are excluded through waveform similarity screening (correlation coefficient greater than 0.7) and spatial consistency screening (adjacent electrode correlation coefficient greater than 0.6), and finally a micro-potential event list is output.

[0077] Specifically, in the embodiment, the preprocessed channel B signal is first denoised by wavelet transform, and high-frequency transient components are reserved. The wavelet transform adopts discrete wavelet transform, the wavelet basis is db4, the decomposition layer number is 5, only high-frequency detail coefficients cD3-cD5 (corresponding to 100-500 Hz) are reserved, low-frequency coefficients cA5 and noise dominant layers cD1-cD2 are set to zero, and then an enhanced signal is reconstructed through inverse DWT ;

[0078] For the enhanced signal, a short-time energy is calculated: ;

[0079] In the formula: En(n) represents the short-time energy of the time point n; xn represents the enhanced signal value; W represents the half window length;

[0080] A dynamic energy threshold is determined based on the background noise level of the short-time energy sequence: ;

[0081] In the formula: σ(E) represents the standard deviation estimation of the short-time energy sequence E; MAD(E) represents the median absolute deviation of the energy sequence E; Th represents the dynamic energy threshold; μ(E) represents the median of the energy sequence E; denotes the energy standard deviation; denotes the empirical coefficient;

[0082] The first order difference of the enhanced signal is calculated to obtain an approximate instantaneous slope sequence: ;

[0083] In the formula: denotes the signal slope at time point n; denotes the time point index; denotes the amplitude of the enhanced signal at time point n; denotes the amplitude of the enhanced signal at time point n; denotes the sampling interval;

[0084] The 95th percentile of the entire signal slope sequence is taken as the slope threshold;

[0085] On the short-time energy curve, a section exceeding the dynamic energy threshold is found, and the duration of the section is greater than the minimum duration (such as 1 ms), and at the same time, the sampling rate of the instantaneous slope exceeding the slope threshold in the section is checked (such as requiring more than 50%), and finally, it is checked whether the peak-to-peak amplitude of the event (i.e., the difference between the maximum and minimum values in the event window) is greater than 3 times the noise standard deviation, and the above conditions are met at the same time, which is marked as a candidate micro-potential event.

[0086] For each detected candidate micro-potential event, the following features are extracted:

[0087] Time positioning features: including the start time, peak time and end time of the event (unit: seconds).

[0088] Duration: end time minus start time (usually less than 10 milliseconds).

[0089] Amplitude features: peak amplitude (with positive and negative polarity, unit: microvolts) and peak-to-peak amplitude (the difference between the positive and negative peaks, unit: microvolts).

[0090] Waveform features: rising slope (average slope from the start of the event to the peak point), falling slope (average slope from the peak point to the end of the event), and waveform asymmetry (the difference between the absolute values of the rising slope and the falling slope divided by the sum of the two).

[0091] Spectral features: spectral analysis (such as FFT) is performed within the time window of the event occurrence to extract the center frequency, bandwidth and energy proportion in the 80-250Hz frequency band.

[0092] Further, in the present embodiment, in order to exclude false detections, the following screening rules are enhanced:

[0093] Amplitude filtering: only keep events with peak-to-peak amplitude greater than 3 times the standard deviation of noise;

[0094] Waveform similarity filtering: for adjacent events on the same electrode, calculate the correlation coefficient of waveform shape, only keep events with correlation coefficient greater than 0.7;

[0095] Spatial consistency filtering: for events detected synchronously by adjacent electrodes (such as distance less than 5mm), calculate the correlation coefficient of waveforms, if greater than 0.6, keep the entire event cluster;

[0096] Based on the above filtering, the filtered micro-potential events and corresponding features are obtained.

[0097] The coupling event processing module realizes the effective fusion of HFOs and micro-potential events through spatio-temporal correlation analysis, as follows:

[0098] Based on the synchronization time stamp during signal acquisition, the high-frequency oscillation event list and the micro-potential event list are mapped to the same time axis to ensure the consistency of event time stamps.

[0099] As shown in Figure 3 , according to the physiological mechanism of seizure focus neuron discharge (synaptic transmission delay 1-10ms, cluster synchronization time 10-50ms), set the correlation time window ( starting time of HFOs event) for each HFOs event; filter the event pairs whose peak times fall within the window, calculate the time difference , keep (time correlation event pair in the effective physiological coupling time range) which is more likely to be from the coordinated discharge of the same neuron cluster rather than random co-occurrence.

[0100] It should be noted that Figure 3 the 10ms-20ms marked position falls exactly within the interval of 500-1500ms on the horizontal coordinate, which is not a corresponding relationship, but only to clearly highlight the core parameters on the macro time axis.

[0101] Extract the three-dimensional coordinates of the electrode contacts corresponding to the time correlation event pairs, calculate the Euclidean distance between the two points, set the spatial distance threshold (matching the distance between electrode contacts), and filter out event pairs as HFOs-micro-potential coupling events; event coupling of adjacent electrode contacts (belonging to the same neuron cluster) is more epileptogenic, and long-distance coupling may be a passive response caused by signal propagation.

[0102] Based on the above logic, it needs to meet the conditions of spatial correlation analysis and time correlation analysis; wherein The spatial distance threshold is represented, and the value is determined according to the distance between the intracranial electrode contacts, such as 5mm. As described previously, the distance between the intracranial electrode contacts is 3-10mm, and the contacts within 5mm are considered to be locally adjacent regions, representing the activity of the same neuron cluster. It should be noted that the spatial distance threshold can be dynamically adjusted according to the actual distance between the electrode contacts, and the adjustment range is 3-10mm, consistent with the electrode arrangement parameters.

[0103] Extract the four-dimensional quantitative features of each coupling event:

[0104] Standardize the latency to reflect the time coupling tightness;

[0105] Amplitude coupling strength , wherein represents the amplitude coupling strength; represents the peak amplitude of the micro-potential; represents the peak amplitude of the HFOs;

[0106] The waveform morphology correlation is the waveform segment (20ms window) of 10ms before and after the micro-potential peak time, and the waveform signals of HFOs and micro-potential in the window are extracted respectively. The closer the correlation coefficient is to 1, the better the waveform synchronization is, and the stronger the neuron activity coordination is.

[0107] The spatial coupling strength is to standardize the spatial distance to the spatial threshold range, reflecting the locality of spatial coupling: ;

[0108] In the formula: represents the spatial coupling strength; represents the spatial distance threshold;

[0109] The four-dimensional feature vector is input into the DBSCAN density clustering algorithm, and the effective cluster sample size threshold is set to 5. The isolated clusters (i.e. pseudo-coupling events) are removed, and the effective coupling event clusters are retained. The output effective coupling event set is shown in Figure 4 The clustering results are shown in

[0110] The spatial positioning weighting module traverses the effective coupling event set, and according to the electrode contact identifier recorded by the event, maps each coupling event to the corresponding three-dimensional spatial coordinates (determined by the preoperative imaging examination to determine the electrode contact position), establishes an event-coordinate association table, and the specific process is as follows:

[0111] The weighted strategy of basic weight + dynamic factor is adopted to make high confidence events dominate positioning, and the specific formula is as follows: ;

[0112] In the formula: represents the basic weight; for the basic weight, the value is assigned according to clinical experience, such as the basic weight of the coupling event is 1; the basic weight of the isolated HFOs event is 0.3; the basic weight of the isolated micro-potential event is 0.1; the epileptogenicity confidence of the coupling event is the highest, and the specificity of the isolated event is relatively insufficient.

[0113] Among them, the time density factor , wherein is the decay constant, is the time difference between the peak value of the micro-potential and the start of the HFOs, the smaller the time difference, the closer the factor to 1, reflecting the better time cooperativity; the amplitude correlation factor , wherein corr is the Pearson correlation coefficient, and the value range after conversion is 0.5-1, reflecting the amplitude correlation strength, represents the peak amplitude of the micro-potential in a single coupling event, represents the peak amplitude of the HFOs in the same coupling event; the time frequency factor , represents the total number of coupling events detected by the corresponding contact; T represents the total effective recording time length; represents the HFOs subtype factor, which distinguishes the epileptogenicity according to the spectral characteristics of HFOs, and the specific assignment rules are as follows:

[0114] If the HFOs are Fast Ripple (FR, 250-500 Hz), then is 1.2;

[0115] If the HFOs are Ripple (80-250 Hz), then is 1;

[0116] If the HFOs are Gamma (40-80 Hz) or others, then is 0.8.

[0117] The epileptic focus core determination module is based on the preoperative MRI / CT scan data of the patient, reconstructs a three-dimensional brain structure model, and creates a regular three-dimensional grid array covering the whole brain, with a grid spacing of 1mm to ensure positioning accuracy, as follows:

[0118] Based on the brain structure model of the patient, a regular three-dimensional grid array covering the whole brain is created ;

[0119] For each point in the grid , calculate its spatial distance to all electrode contacts, amplify the weight of the closest contact, attenuate the weight of the farthest contact, aggregate all contact's attenuated weight contribution to this grid point, and then generate the corresponding point's epileptogenic probability value through normalization processing, the specific expression is as follows:

[0120] ;

[0121] ;

[0122] In the formula: represents the Euclidean distance between point p and electrode contact ; represents the spatial kernel function; represents the total weight of the contact ; represents the original interpolation probability value of point p; represents the final epileptogenic probability of point p; represents the slope factor; represents the offset;

[0123] Map the epileptogenic probability value of the grid point to the corresponding spatial position of the brain structure model to form a continuous probability cloud map;

[0124] Based on the continuous probability cloud map, calculate the mean and standard deviation of the whole brain probability, and set a dynamic threshold based on the calculated mean and standard deviation:

[0125] ;

[0126] In the formula: represents the dynamic threshold; represents the mean of the whole brain probability; represents the standard deviation of the whole brain probability; represents the standard deviation multiple;

[0127] Label the area above the dynamic threshold in the probability cloud map as the candidate epileptogenic core area, and label the area below the threshold as the safe brain area;

[0128] Identify all connected candidate epileptogenic core areas in three-dimensional space, and remove isolated noise points with a volume less than the threshold. In the remaining connected regions, select the region with the highest internal probability peak as the core of the epileptic focus. If there are multiple peaks that are spatially separated, mark them as multiple core areas, and output the three-dimensional coordinates of the core area of the epileptic focus based on this.

[0129] In the present application, through coupling event screening, non-epileptogenic related isolated HFOs can be effectively excluded, and the specificity of epileptogenic focus localization is improved; through four-dimensional feature vector (standardized latency, amplitude coupling strength, waveform shape correlation, spatial coupling strength) + DBSCAN density clustering, isolated clusters (pseudo-coupling events) with sample size less than the threshold are excluded; through the high-resolution characteristics of micro-potential, the localization range of HFOs is constrained to a smaller area, the spatial resolution is improved to a small range near the electrode contact, and the sensitivity is retained through coupling screening optimization, while solving the specificity problem, avoiding missing the epileptogenic focus due to the pursuit of specificity; through spatial interpolation to generate a whole brain epileptogenic probability distribution map, combined with a dynamic threshold to determine the core area, continuous and accurate positioning in three-dimensional space can be realized, which adapts to the demand of clinical surgery for three-dimensional brain structure.

[0130] The above is only a preferred embodiment of the present application, and is not intended to limit the present application. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A high-frequency oscillation fusion micropotential-based system for precise localization of epileptic foci, characterized in that, It includes a signal acquisition module, a dual-channel preprocessing module, a feature extraction module, a coupled event processing module, a spatial positioning weighting module, and an epileptic focus core determination module; The signal acquisition module is used to acquire intracranial electroencephalogram (EEG) signals collected by the electrode contacts; The dual-channel preprocessing module includes a filtering unit, an artifact and noise removal unit, and a baseline correction unit. The filtering unit is used to perform dual independent channel filtering on the intracranial EEG signal. The artifact and noise removal unit is used to remove artifacts and noise from the filtered signal. The baseline correction unit is used to perform baseline correction on the artifact-removed and denoised signal, and finally outputs a first clean signal adapted to high-frequency oscillation analysis and a second clean signal adapted to micropotential analysis. The feature extraction module includes a high-frequency oscillation feature extraction unit and a micro-potential feature extraction unit. The high-frequency oscillation feature extraction unit is used to perform event detection, clustering denoising, and feature extraction on the first clean signal, and output a high-frequency oscillation event list. The micro-potential feature extraction unit is used to perform signal-to-noise ratio enhancement, event detection, and feature extraction on the second clean signal, and output a micro-potential event list. The coupling event processing module is used to align the high-frequency oscillation event list with the micro-potential event list along the time axis, filter event pairs that meet the coupling conditions through spatiotemporal correlation analysis, extract the joint features of the coupling events and exclude pseudo-coupling events, and output a set of effective coupling events. The spatial positioning weighting module is used to map each effective coupling event to the corresponding three-dimensional spatial coordinates of the electrode contact, assign basic weights based on the event type, and dynamically adjust the weights in combination with event characteristics to output a weighted spatial distribution map of coupling events. The epileptogenic focus core determination module is used to generate an epileptogenic probability distribution map on the patient's three-dimensional brain structure model based on the coupled event spatial distribution map and using a spatial interpolation algorithm, and to determine the epileptogenic focus core region by setting a probability threshold.

2. The high-frequency oscillation fusion micropotential precise localization system for epileptic foci according to claim 1, characterized in that, The configuration of each unit in the dual-channel preprocessing module is as follows: The filtering unit is configured with dual independent filtering channels. The first channel has a passband of 80Hz to 500Hz and a stopband attenuation of not less than 40dB. The second channel has a passband of 100Hz to 1000Hz and a stopband attenuation of not less than 50dB. Both channels use bidirectional filtering to eliminate phase shift and output high-frequency oscillation detection signal stream and micro-potential detection signal stream, respectively. The artifact removal and noise removal unit first decomposes the two signals output by the filtering unit into multiple independent signal components using independent component analysis. By calculating the kurtosis and power spectrum characteristics of each component, artifact components with abnormal kurtosis and power spectra that do not conform to the electroencephalographic characteristics are screened out and removed. The remaining components are reconstructed into an artifact-free signal. For the artifact-free signal corresponding to the micropotential detection signal stream, an 8-layer wavelet decomposition is further used to separate the approximation coefficients and detail coefficients. After soft thresholding of the detail coefficients, they are reconstructed with the approximation coefficients to obtain an intermediate signal with enhanced signal-to-noise ratio. The baseline correction unit sets sliding windows for the artifact-free signal of the high-frequency oscillation detection signal stream and the intermediate signal of the micro-potential detection signal stream, respectively. The window width of the first channel is larger than that of the second channel. The local average value of each time point is calculated by the sliding window as the corresponding baseline. After subtracting the baseline signal from the original signal, the first clean signal and the second clean signal are output respectively.

3. The high-frequency oscillation fusion micropotential precise localization system for epileptic foci according to claim 2, characterized in that, The filtering unit uses an FIR filter to achieve bidirectional filtering. During the filtering process, the signal delay is compensated by a zero-phase offset algorithm to ensure the time synchronization of the high-frequency oscillation detection signal stream and the micro-potential detection signal stream.

4. The high-frequency oscillation fusion micropotential precise localization system for epileptic foci according to claim 2, characterized in that, In the baseline correction unit, the sliding window width of the first channel is 50ms to 100ms, the sliding window width of the second channel is 20ms to 50ms, and the window sliding step size is 1ms. The baseline value at each time point is calculated by a local weighted average algorithm.

5. The high-frequency oscillation fusion micropotential precise localization system for epileptic foci according to claim 1, characterized in that, The high-frequency oscillation feature extraction unit calculates the time spectrum in different time periods and filters candidate intervals by combining the baseline signal energy threshold. Within the candidate intervals, oscillation events are detected by the Teager energy operator. After morphological examination to eliminate non-oscillatory artifacts, the time, frequency domain, energy and morphological features of the events are extracted. Then, outliers are eliminated through cluster analysis to obtain a list of high-frequency oscillation events.

6. The high-frequency oscillation fusion micropotential precise localization system for epileptic foci according to claim 1, characterized in that, The micropotential feature extraction unit sets an energy threshold by combining a short-time energy sequence with the background noise level, calculates the instantaneous slope sequence obtained by the first-order difference of the signal and sets a slope threshold, filters candidate events that simultaneously meet the requirements of energy threshold, slope threshold, minimum duration and peak-to-peak amplitude, retains events with repetitive waveform characteristics and extracts relevant features to form a micropotential event list.

7. The high-frequency oscillation fusion micropotential precise localization system for epileptic foci according to claim 1, characterized in that, The spatiotemporal correlation analysis of the coupled event processing module includes: Define a micropotential correlation time window for high-frequency oscillation events, and screen time-correlated event pairs whose peak time falls within the time window and whose time difference meets the physiological coupling range; calculate the three-dimensional spatial distance between the electrode contacts corresponding to the time-correlated event pairs, and screen spatially correlated event pairs whose distance does not exceed a set threshold as high-frequency oscillation-micropotential coupling events.

8. The high-frequency oscillation fusion micropotential precise localization system for epileptic foci according to claim 7, characterized in that, The coupling event processing module extracts four types of quantitative features of the coupling events: standardized latency, amplitude coupling strength, waveform morphology correlation, and spatial coupling strength, forming a four-dimensional feature vector. Then, a density clustering algorithm is used to remove isolated clusters with sample sizes less than a set threshold to obtain an effective set of coupling events.

9. The high-frequency oscillation fusion micropotential precise localization system for epileptic foci according to claim 1, characterized in that, The dynamic weight adjustment of the spatial positioning weighting module includes a time tightness factor, an amplitude correlation factor, a time frequency factor, and a high-frequency oscillation subtype factor. The time tightness factor is negatively correlated with the time difference between the peak value of the micropotential and the onset of the high-frequency oscillation. The amplitude correlation factor is obtained based on the correlation coefficient conversion between the high-frequency oscillation signal amplitude sequence and the micro-potential signal amplitude sequence. The time frequency factor is the natural logarithm of the number of coupled events per unit time. The high-frequency oscillation subtype factor is set to different values ​​according to the spectral characteristics.

10. The high-frequency oscillation fusion micropotential precise localization system for epileptic foci according to claim 1, characterized in that, The epileptogenic focus core determination module creates a regular three-dimensional grid covering the entire brain, calculates the spatial distance from each grid point to all electrode contacts, performs distance attenuation on the contact weights and then aggregates and normalizes them to generate epileptogenic probability values ​​and maps them to a probability cloud map. Based on the whole brain probability mean and standard deviation, a dynamic threshold is set, connected regions above the threshold are marked and isolated noise points are removed to determine the epileptogenic focus core region.

Images