A method for localizing the epileptic focus based on analysis of multiple epileptogenic pathological markers in a single lead.

Abstract

This invention discloses a single-lead SOZ localization method based on the analysis of multiple epileptogenic pathological biomarkers. First, the epileptogenic characteristics of a single lead are calculated from both signal distribution and signal energy perspectives. A pathological ripple normalization feature is proposed, namely, a normalized ripple rate of the region + 10% threshold, to minimize the interference of physiological ripples on localization. Then, a Shapley value and hypothesis testing feature selection algorithm are used to avoid interference from irrelevant or redundant features. Next, an attention mechanism and focus loss algorithm are applied to a shallow neural network classifier to overcome the limitation of imbalance between epileptogenic and non-epilepsogenic touchpoints, thereby better learning features significantly related to localization and achieving high-sensitivity touchpoint identification. Finally, the epileptogenic coefficients of the analyzed touchpoints are displayed in magnetic resonance imaging images using a whole-brain mapping method, providing clinicians with reliable and interpretable auxiliary localization results for more accurate preoperative assessment.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a method for SOZ localization based on the analysis of multiple epileptogenic pathological markers in a single lead. Background Technology

[0002] Antiepileptic drugs can effectively control seizures in most epilepsy patients, but more than 30% of patients remain incurable, a condition known as drug-resistant epilepsy (DRE). For these DRE patients, the best treatment is surgical removal or ablation of the brain region causing the seizures to reduce or prevent further seizures. The brain region that actually produces clinical seizures is called the seizure onset zone (SOZ). Therefore, accurate localization of the SOZ is crucial for successful surgery.

[0003] Stereotactic electroencephalography (SEEG) is the gold standard for SOZ localization. Unlike invasive electrocorticography (ECOG), it is recorded by deep electrodes implanted deep within the brain. SEEG can record electrical signals not only on the surface of the cerebral cortex but also in the amygdala and deep structures of the hippocampus. However, visually examining long-term interictal SEEG to supplement pathological information is extremely laborious, time-consuming, and subjective for clinicians.

[0004] For the SEEG during the interictal period, the identification of the epileptogenic trigger point of SOZ based on artificial intelligence algorithms is usually completed in three stages: (1) using typical signal analysis methods, machine learning, deep learning and other algorithms to detect potential epileptogenic pathological markers in the SEEG; (2) analyzing and extracting the epileptogenic characteristics of a single lead based on the epileptogenic pathological markers detected in the previous stage; (3) usually requiring feature selection, and then classifying the epileptogenic trigger point through machine learning algorithms.

[0005] A crucial first step in SOZ localization is the detection of pathological biomarkers. In recent years, researchers have discovered several pathological biomarkers with localization potential that are independent of epileptic seizures, such as spikes, high-frequency oscillations (HFOs), and interictal epileptiform discharges (SEEGs). Regarding spike detection, the inventors previously demonstrated that deep learning can detect subtle pathological changes in SEEGs, and subsequently designed a more adaptive and highly interpretable SEEG-Net model, achieving detection performance with higher sensitivity and strong generalization. For HFO detection, the inventors previously jointly analyzed the filtered signal of the original signal with time-frequency images, achieving signal detection with high accuracy and low false negative rate. Furthermore, for other important epileptiform discharge features, Akter et al. used information-theoretic features extracted from high-frequency subbands to detect epileptogenic foci in interictal EEG. Klimes et al. localized SOZs based on interictal data by calculating oscillatory events of whole-brain contacts, univariate spectral analysis, and other features. Therefore, improving the detection accuracy of epileptogenic pathological biomarkers is crucial and can lead to better SOZ localization results. However, most studies focus only on the signal level, specifically detecting whether SEEG signal fragments originate from SOZ. Such studies cannot accurately trace the epileptogenic trigger point and lesion area from the overall perspective of individual patients. Therefore, conducting a second-step study is imperative.

[0006] The second crucial step in SOZ localization is the extraction of epileptogenic features from a single lead. Since even a single pathological marker is insufficient for precise localization of the epileptogenic focus, Klimes et al. used multiple pathological markers and a multi-feature fusion method, achieving superior localization results compared to single-marker methods, thus promoting the development of localization studies using combinations of multiple markers. Therefore, representing the epileptogenic features of a single lead based on the multi-dimensional features of multiple pathological markers is a future research trend. Furthermore, high-frequency physiological signals can interfere with epileptogenic focus localization, and currently, we cannot clearly distinguish between pathological and physiological high-frequency signals from a signal dimension perspective. Zweiphenning et al. investigated methods for correcting physiological ripples to reduce the impact of physiological high frequencies on localization and enhance the performance of epileptogenic trigger point identification. Compared to traditional methods that only calculate high-frequency signals, ripple standardization enhances clinical applicability. Therefore, quantifying the pathological ripple rate from a single-lead perspective to improve SOZ localization performance is an important problem we need to address.

[0007] In the third step, selecting effective features and analyzing the salient features during feature extraction is also quite challenging. Traditional clinical analysis generally uses thresholding methods. Zweiphenning et al. analyzed various thresholding methods, with the best achieving an accuracy of 77.3% and a sensitivity of 27%. However, thresholding methods still suffer from poor adaptability. In recent years, research combining AI algorithms has employed sparse LDA for feature selection and machine learning classifiers such as support vector machines for SOZ touchpoint recognition, achieving a sensitivity of 52.70%. However, these methods still suffer from insufficient adaptability and sensitivity issues. Therefore, designing suitable feature selection methods and SOZ recognition classifiers is the focus of this invention.

[0008] Based on the above limitations, this invention aims to solve the following three problems to improve the sensitivity and overall performance of SOZ localization: (1) How to combine and analyze the multi-dimensional features of multiple epileptogenic pathological markers and extract the epileptogenic features of the contact point from multiple perspectives and aspects? (2) High-frequency oscillation signals are typical and important epileptogenic pathological markers. How to reduce the influence of physiological high-frequency signals on localization? (3) How to effectively select features from a large number of related features, remove irrelevant features and redundant information, thereby improving the sensitivity of epileptogenic contact point identification? Summary of the Invention

[0009] This invention addresses the problems of incomplete feature analysis, low sensitivity, and poor individual generalization in existing interictal epileptogenic focus localization studies. It aims to develop an objective, highly sensitive single-lead localization method for SOZ based on long-term interictal SEEG data, which requires no manual visual inspection. This method will provide reliable assistance to clinicians and enable accurate preoperative assessment of the interictal period.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A method for SOZ localization based on analysis of multiple epileptogenic pathological markers in a single lead includes the following steps:

[0012] S1. Detect spike waves, spike-and-slow-wave complexes, and high-frequency oscillation signals as epileptogenic pathological markers; among which, high-frequency oscillation signals include Ripples, Fast Ripples, Ripple, and Fast Ripple complexes.

[0013] S2. Extract epileptic features of a single lead from both the signal distribution and signal energy of spike waves, spike-slow-wave complexes, and high-frequency oscillation signals.

[0014] S3. Use a feature selection algorithm based on Shapley value and hypothesis testing to select significant features;

[0015] S4. Apply attention mechanism and focus loss algorithm to shallow neural network classifier to realize lead identification and calculate the predicted score of SOZ contact.

[0016] S5. The predicted scores of SOZ touch points are displayed in magnetic resonance imaging images using a whole-brain mapping method.

[0017] Furthermore, in step S1, the high-frequency oscillation signal detection method can identify events with more than 6 oscillations within a specified time period based on the amplitude of the Hilbert envelope of the bandpass filtered signal.

[0018] Furthermore, step S2 employs pathological Ripples standardization with a region + 10% threshold to improve localization performance.

[0019] Furthermore, in step S2, the epileptic features of the single lead in terms of signal distribution include the occurrence rate of spikes, the occurrence rate of ripples, the occurrence rate of standardized pathological ripples, the occurrence rate of fast ripples, and the occurrence rate of ripple and fast ripple complex waves.

[0020] Furthermore, in step S2, the spike occurrence rate is determined by three different spike detectors to identify the spike signal: a typical signal feature extraction method, a one-dimensional convolutional neural network, and a SEEG network.

[0021] Furthermore, in step S2, the epileptic features of a single lead in terms of signal energy include spike energy, Ripple energy, Fast Ripple energy, and Ripple and Fast Ripple complex wave energy.

[0022] Furthermore, in step S3, the Shapley value-based feature selection algorithm evaluates relevant features by calculating local correlation, global correlation, and an adaptive threshold. Specifically, the SHAP tree method is used to calculate local correlation, a two-stage method is used to adjust the threshold, and the proportion of features in the sequence with local correlation lower than the adaptive threshold is used to determine whether a feature is irrelevant. Irrelevant features are excluded through hypothesis testing.

[0023] Further, the whole-brain mapping method in step S5 is as follows: the epileptiformity value of the brain boundary voxel is set to the lowest quantization value; the epileptiformity values ​​of the SOZ contact voxel and the voxel closest to the SOZ contact are both set to the epileptiformity quantization value corresponding to the SOZ contact; the 3D Gaussian kernel algorithm is used to map all other voxel block values ​​in the whole brain to the epileptiformity values ​​of the SOZ contact and its surrounding voxels and the epileptiformity values ​​of the brain boundary voxels.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] This invention proposes a SOZ localization method based on the analysis of multiple epileptogenic pathological markers in a single lead. First, to improve the detection accuracy of multiple pathological markers (spikes, high-frequency oscillations), we simultaneously use three algorithms to detect spikes and an envelope-based high-frequency oscillation signal screening algorithm to detect ripples, fast ripples, and ripple-fast ripple complexes. Second, to fully and multidimensionally represent the epileptogenic characteristics of a single lead, we extract epileptogenic features based on both the signal distribution and signal energy of the aforementioned pathological markers. Specifically, we designed a standardized pathological ripple rate at the lead level to reduce the impact of physiological ripples on localization and improve localization performance. Next, to reduce interference from irrelevant features and redundant information, we use a feature selection algorithm based on Shapley values ​​and hypothesis testing (ShapHT+) to select salient features. Finally, to improve the severe imbalance between positive and negative samples and allow the model to focus on learning a small number of positive samples, we introduce an attention mechanism and a focus loss algorithm into the shallow neural network classifier to achieve lead recognition. Furthermore, based on interpolation, the predicted scores of the SOZ (Sexually Occult Focus) are visualized across the entire brain, enhancing the reliability and applicability of this invention in clinical applications. This invention selected ten patients with drug-resistant epilepsy to validate our method. Experiments confirmed that the method of this invention can improve the sensitivity and accuracy of epileptogenic focus localization and can help clinicians perform accurate and reliable preoperative assessments based on interictal SEEG (Sexually Occult Focus Emission). Attached Figure Description

[0026] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0027] Figure 1 The flowchart of the SOZ localization method based on single-lead analysis of multiple epileptogenic pathological markers provided in this embodiment of the invention is as follows: (A) The method includes the original SEEG recording, detection of multiple epileptogenic pathological markers, extraction of epileptogenic features in a single lead, feature selection, and touchpoint classification; (B) The epileptogenicity prediction score of the touchpoint is visualized on the patient's MRI. SOZ represents the epileptic focus seizure area; NSOZ represents the non-epileptic focus seizure area; FRs represent fast ripples; FT represents Fourier transform; DWT represents discrete wavelet transform.

[0028] Figure 2 The results are from the detection of multiple epileptogenic pathological markers.

[0029] Figure 3SOZ was described for an individual patient using bar charts and MRI visualization. Detailed Implementation

[0030] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.

[0031] This invention proposes a SOZ localization method based on single-contact (MEBM-SC) analysis of multiple epileptogenic biomarkers in a single lead, achieving intelligent SOZ localization with high sensitivity and high accuracy. Figure 1 The entire process of MEBM-SC is described.

[0032] Specifically, firstly, we calculate the epileptogenic features of a single lead from two aspects: signal distribution and signal energy. We propose a pathological ripple normalization feature, namely a normalized ripple rate of the region + 10% threshold, to minimize the interference of high-frequency physiological signals on localization and enhance the identification ability of epileptogenic contacts. Then, we employ a Shapley value and hypothesis testing (ShapHT+) feature selection algorithm to avoid interference from uncertain or irrelevant features. Next, we apply attention mechanisms (AM) and focus loss algorithms to overcome the limitations of the imbalance between epileptogenic and non-epilepsogenic contacts, thereby better learning features significantly related to localization and achieving high-sensitivity identification of contacts. Finally, we display the epileptogenic coefficients of the analyzed contacts in magnetic resonance imaging (MRI) images using a whole-brain mapping method, providing clinicians with reliable and interpretable auxiliary localization results for more accurate preoperative assessment.

[0033] 1. Detection of multiple epileptogenic pathological markers

[0034] Spikes, spike-and-slow-wave complexes, and high-frequency oscillations (including Ripple (80-250Hz), Fast Ripple (>250Hz), and ripples co-occurring with Ripple and Fast Ripple (>80Hz)) are the main epileptogenic pathological markers used in SOZ localization in this invention.

[0035] First, to detect signal patterns such as spikes and spike-slow-wave complexes, this section applies various algorithms across multiple domains. For spikes, typical signal processing methods are used, such as nonlinear features extracted in the high-frequency domain. For spike-slow-wave complexes, machine learning and deep learning algorithms are employed, such as the basic model of a one-dimensional convolutional neural network and the novel model SEEG-Net. Then, to detect signal patterns such as high-frequency oscillations, this invention improves upon previous high-frequency oscillation signal detectors. Based on the amplitude of the Hilbert envelope of the bandpass filtered signal, it identifies events with more than six oscillations within a specified time period. Detailed algorithms are shown in Table 1.

[0036] Table 1. Logic Table of Initial Screening Algorithm

[0037]

[0038] 2. Extraction of epileptic features at the touchpoint

[0039] The epileptogenic feature extraction of this invention is based on multiple features of various epileptogenic pathological markers, namely, multi-domain features in terms of pathological signal distribution and pathological signal intensity, totaling 126 items, as shown in Table 2.

[0040] Table 2. Distribution and intensity of pathological signals around the contact point: 126 epileptiform features of the contact point.

[0041]

[0042]

[0043] 2.1 Pathological signal distribution characteristics of contact points

[0044] (1) Spike occurrence rate

[0045] This invention employs three different spike detectors to identify spike signals. Specifically, these include: (1) a signal detection method based on typical signal feature extraction, which has high accuracy and strong interpretability based on high-frequency bands; (2) a one-dimensional convolutional neural network; and (3) SEEG-Net, which can improve problems in clinical scenarios such as sample imbalance, individual patient domain bias, and poor interpretability, and can achieve highly sensitive detection of pathological SEEG. For a certain contact point i of patient P, its spike rate S rate It can be represented as:

[0046]

[0047] Where j is the index of the three spike detectors.

[0048] (2) Occurrence rate of standardized pathological Ripples (R)

[0049] For a certain contact point i of patient P, the Ripple occurrence rate can be expressed as R. rate :

[0050]

[0051] in, It is the total number of times the Ripple signal appears in contact i, and T is the duration of a single data segment, i.e., thirty minutes.

[0052] Pathological standardization of ribs significantly enhances the ability to identify SOZs based on ribs signals, with a region +10% threshold showing even better performance. Specifically, after ribple detection of the original signal, ribs with detection results less than 10% of the total HFO (Hyperfocal Optimal Root) are discarded after brain region correction. This eliminates ribple values ​​that are only slightly higher than physiological ribple levels. Therefore, the standardized pathological ribple occurrence rate is expressed as R. normal_r :

[0053]

[0054] In formula (3) R region The constants defined in previous studies represent the occurrence rate of physiological high-frequency oscillation signals in different brain regions. Specific values ​​for each brain region involved in the dataset of this invention are shown in Table 3. The setting of quantile(data,percentile) is used to set contacts less than 10% of the total HFO to 0. N c This represents the total number of patient contacts.

[0055] Table 3. Different Rs occurrence rates in different brain regions

[0056]

[0057] (3) Occurrence rate of Fast Ripples (FRs)

[0058] Unlike standardized pathological ribs occurrence rates, pathological standardization of ribs does not improve their performance in SOZ localization, therefore, it is unnecessary to perform pathological standardization of ribs. Specifically, for a certain contact i of patient P, the occurrence rate of ribs can be expressed as ribs. rate :

[0059]

[0060] in, This represents the number of times FR appears on contact i.

[0061] (4) Occurrence rate of R&FRs composite waves

[0062] For a certain contact point i of patient P, the occurrence rate of R&FRs can be expressed as FR&R rate :

[0063]

[0064] in, This represents the number of times the R&FRs composite wave occurs at contact i.

[0065] 2.2 Pathological signal energy characteristics of the contact point

[0066] (1) Spike energy at the contact point

[0067] This invention characterizes and calculates the spike energy of a contact point based on manually extracted typical signal features. Since the spike segment has a fixed length of 5 seconds, duration differences are not considered here. For a contact point i of patient P, the spike energy of the contact point can be expressed based on the eigenvalue of the discrete wavelet transform, as follows:

[0068]

[0069] in, This represents the number of times the spike occurs at contact i.

[0070] (2) Ripples energy at the contact point

[0071] This invention extracts the Ripple energy features of the contact point based on multiple domains, namely (1) the feature median based on the time domain; (2) the feature median based on the frequency domain; and (3) the feature median based on the discrete wavelet transform. The feature median calculation method of this invention is the same as that of formula (7), and the specific feature extraction formula is shown in Appendix Table 1. It is worth noting that the length of the Ripple signal is variable, so we consider the time concentration of the energy feature.

[0072] (3) The energy of FRs and the energy of R&FRs composite wave at the contact point

[0073] Since the length of FRs signals is shorter than that of Ripples signals, this invention removes two features based on the discrete wavelet transform domain: Kraskov entropy and Renyi entropy. The extraction of other features remains the same. Furthermore, the calculation of the energy of the R&FRs composite wave is performed using the same method as the calculation of the FRs signal energy.

[0074] 3. Feature selection and touchpoint classification

[0075] 3.1 Feature Selection Based on ShapHT+

[0076] The key to feature selection lies in how to derive a subset of features relevant to the problem domain from the initial set, thereby improving feature interpretability and the performance of subsequent lead identification. The crucial aspect of identifying relevant features is: defining and calculating a metric to evaluate relevance; and constructing a set of all relevant features based on the metric.

[0077] This invention utilizes a ShapHT+ feature selection strategy to select the most relevant features and reduce uncertain, irrelevant features. This strategy calculates local relevance, global relevance, and an adaptive threshold to evaluate relevant features. Specifically, the SHAP tree method is used to calculate local importance.

[0078]

[0079] in, Let T(·) represent the dataset, where T(·) is the model's predicted output value, and p0 is the expected output value, calculated as follows: Indicates sample X (n) Below, feature set Chinese x i Contributions; Let T(·) represent the model expectation, which is based on the sample X. (n) Below and feature subset As a condition. The absolute value of the contribution is represented by sample X. (n) Middle feature x i Local importance is defined as

[0080] This invention designs an adaptive threshold for evaluating the relevance of features and uses a two-stage method to adjust the threshold. Feature x i Global correlation:

[0081] GI i =E(I i (10)

[0082] Among them, I i Sequences representing local importance, E(I i ) represents the dataset The expected local importance; This indicates a random characteristic. The global importance of a feature is determined by the threshold setting. This means that if a feature's relevance is lower than that of random features, then that feature is considered irrelevant. Therefore, whether a feature is irrelevant is determined by calculating the proportion of features in the sequence whose local relevance is lower than an adaptive threshold.

[0083] Furthermore, hypothesis testing is used to eliminate irrelevant features. This feature selection algorithm can classify the initial features into relevant and irrelevant features one by one. The resulting subset of relevant features is well correlated with the labels, and the redundancy between features is minimal.

[0084] 3.2 Touch point classification based on AM and Focal-loss

[0085] Figure 1 The detailed structure of this part is shown. To improve the effectiveness of classification in deep neural networks, this invention introduces an attention mechanism (AM), expressed as formulas (11)-(13):

[0086] u t =tanh(W w e t +b w (11)

[0087]

[0088]

[0089] Among them, v t W represents the output of the attention layer. w u w and b w This represents the training weights and biases. t h represents the input value. t The learned time information is represented by e t with h t Multiplication, AM from e t We select and extract the relatively important temporal and spatial information for classification tasks.

[0090] Furthermore, in real-world clinical scenarios, the ratio of SOZ to NSOZ contacts is highly imbalanced, which can affect the final localization results. Although the small sample size of SOZ contacts accounts for a small proportion of the total loss, it plays a crucial role. Therefore, we use a Focal-loss-based loss function, a method proposed by researchers to address and handle imbalanced data in dense object detection.

[0091] 4. Visualization of individual patient SOZ predictions

[0092] This invention maps the predicted scores of SOZ contacts obtained from MEBM-SC calculation and analysis onto individual patient MRI images, such as... Figure 1 As shown in (B) above. The specific approach is as follows:

[0093] Considering the low correlation between brain boundary voxels and SOZ, the epileptiformity value of brain boundary voxels is set to the lowest quantization value.

[0094] Considering the strong correlation between the voxels closest to each SOZ contact location, the epileptiformity values ​​of the SOZ contact voxels and the epileptiformity values ​​of the voxels closest to the SOZ contact location are both set to the epileptiformity quantification values ​​corresponding to the SOZ contact.

[0095] To further realize the epileptogenicity value mapping of whole-brain voxels, the 3D Gaussian kernel algorithm is used to map the values ​​of all other voxel blocks in the whole brain to the epileptogenicity values ​​of SOZ contacts and the voxels around them, as well as the epileptogenicity values ​​of brain boundary voxels.

[0096] 5. Experimental parameter settings

[0097] Pathological biomarker detection models, feature selection methods, and deep learning classifiers may exhibit different performances under different parameter values. All settings for the feature selection method and deep learning classifier studied in this invention are shown in Table 4. All experiments were conducted using the PyTorch framework on a DELL PowerEdge R740 rack server equipped with three NVIDIA GeForce RTX 2080Ti cards.

[0098] Table 4 Optimal Parameter Values ​​for the Experiment

[0099]

[0100]

[0101] Experiment 1

[0102] The epileptogenic trigger localization method in this embodiment of the invention is a binary classification task, and we validated the proposed framework MEBM-SC on a real-world clinical dataset. Table 5 details the results for each patient. The method performed best for patients without seizures, with a sensitivity of 89.27%, specificity of 90.37%, accuracy of 90.87%, PPV of 81.38%, and NPV of 96.58%. Figure 2 The results of various epileptogenic pathological markers were displayed intuitively. Figure 3 The SOZ of individual patients was described using bar charts and MRI visualization.

[0103] Table 5. Cross-validation results for each patient in the method of this invention.

[0104]

[0105] Experiment 2

[0106] The experimental results were evaluated using 10-fold cross-validation and then averaged. Compared with five other state-of-the-art localization studies, our proposed method achieved the best performance in both sensitivity and accuracy, as shown in Table 6.

[0107] Table 6. Comparison of cross-validated MEBM-SC with the latest localization methods.

[0108]

[0109]

[0110] Experiment 3 (Ablation Experiment: Epileptogenic Characteristics of the Contact Point)

[0111] To more comprehensively compare the contributions, effectiveness, and performance of various biomarkers across multiple features, this invention conducted an ablation study. First, the multi-domain digital features of spike and high-frequency oscillation signals were used as foundational features, and other features were added to gradually construct the complete feature set. Then, the features of spike and high-frequency oscillation signals were fused. Furthermore, a standardized pathological ripple rate was incorporated. Finally, an improved standardization method for the pathological ripple rate using a regional +10% threshold was developed. Specifically, the comparative ablation experiments are as follows:

[0112] Feature a (multi-domain, multi-feature spike): based on the spike signal distribution and spike signal energy at the contact point;

[0113] Feature b (multi-domain, multi-feature HFOs): HFO signal distribution and signal energy based on the contact point;

[0114] Feature c (multi-domain, multi-feature of spikes and HFOs): based on the signal distribution and signal energy of the above two epileptogenic pathological markers;

[0115] Feature d (multi-domain multi-feature of spike and pathologically normalized HFO): Based on feature c, pathological normalization of the Ripple is performed based on the regional atlas threshold.

[0116] Feature e / Features used in this paper (multi-domain, multi-feature HFO with spikes and pathologically normalized features): Based on feature c, pathological normalization of the Ripple is performed based on a region +10% threshold.

[0117] Table 7(A) shows the key features extracted by our method, namely feature e, which is effective for SOZ contact identification, and exhibits the highest performance when using a region + 10% threshold. Figure 3As shown, feature c has higher accuracy and sensitivity than features a and b, indicating that the combined feature of spike and HFO signals is more effective. Feature b has higher specificity than all methods, indicating that the false negative rate of high-frequency oscillation signals is the lowest. Feature e has higher sensitivity and accuracy than other features, indicating the effectiveness of Ripple pathological standardization, a conclusion consistent with Zweiphenning's study. In summary, this ablation experiment demonstrates the rationality and effectiveness of the multi-feature epileptogenic pathological markers extracted by the method of this invention.

[0118] Experiment 4 (Ablation Experiment: Feature Selection Method for Contact-Induced Epileptogenic Characteristics)

[0119] In this ablation study, we discuss the effectiveness of feature selection methods and classifiers compared to other methods. The explanation is as follows:

[0120] Feature selection a (XGBoost): XGBoost is used as the feature selection method and classifier;

[0121] Feature selection b (RFE-XGBoost): SVM-RFE is used as the feature selection method and classifier;

[0122] Feature selection c(ShapHT++DNN): Employs a shallow deep learning model, with DNN used as both the feature selection method and the classifier;

[0123] Feature selection d(ShapHT++self-AM+DNN): Adds a self-attention mechanism to the feature selection c.

[0124] Feature selection e / The feature selection method used in this paper (ShapHT+AM+DNN): Based on feature selection c, an attention mechanism is added;

[0125] Table 7(B) shows the performance improvement in feature selection e, indicating that the adopted feature selection method and classifier are necessary and effective. As shown in feature selection c, the sensitivity using shapHT+ is significantly better than that of feature selections a and b. The difference between feature selections d and e lies in the self-attention mechanism. Although their sensitivity levels are similar, feature selection d has lower specificity due to its poor fitting performance to negative samples. In summary, this ablation experiment demonstrates the effectiveness of ShapHT+ and AM used in this invention.

[0126] Table 7(A) Ablation Study Results

[0127]

[0128]

[0129] Table 7(B) Comparison of the effectiveness of ablation studies

[0130]

[0131] In summary, the entire method of this invention is an interdisciplinary innovation that applies artificial intelligence algorithms to SOZ localization and improves the sensitivity of SOZ localization based on the analysis of multiple epileptogenic pathological markers in a single lead.

[0132] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for localizing the epileptic focus area based on the analysis of multiple epileptogenic pathological markers in a single lead, characterized in that, Includes the following steps: S1. Detect spikes, spike-and-slow-wave complexes, and high-frequency oscillations in the SEEG signal as epileptogenic pathological markers; wherein, the high-frequency oscillations include Ripple, Fast Ripple, and Ripple and Fast Ripple complexes; the high-frequency oscillation detection method in step S1 can identify events with more than 6 oscillations within a specified time period based on the amplitude of the Hilbert envelope of the bandpass filtered signal. S2. Epileptiform features of single leads are extracted from both signal distribution and signal energy aspects of spike waves, spike-and-slow-wave complexes, and high-frequency oscillation signals. The epileptiform features of single leads in terms of signal distribution include the occurrence rate of spike waves, the occurrence rate of Ripple waves, the occurrence rate of FastRipple waves, and the occurrence rate of Ripple-and-Fast Ripple complexes. The spike wave occurrence rate is identified using a one-dimensional convolutional neural network. The epileptiform features of single leads in terms of signal energy include spike wave energy, Ripple energy, FastRipple energy, and Ripple-and-Fast Ripple complex energy. S3. Use a feature selection algorithm based on Shapley value and hypothesis testing to select relevant features. The feature selection algorithm based on Shapley value evaluates relevant features by calculating local correlation, global correlation and adaptive threshold. Specifically, the SHAP tree method is used to calculate local correlation, a two-stage method is used to adjust the threshold, and the proportion of features with local correlation below the adaptive threshold in the sequence is used to determine whether a feature is irrelevant. Hypothesis testing is used to exclude irrelevant features. S4. Apply attention mechanism and focus loss algorithm to shallow neural network classifier to realize lead recognition and calculate the predicted score of the contact point in the epileptic focus seizure area. S5. The predicted scores of the epileptic focus seizure area contact points are displayed in the magnetic resonance imaging image using a whole-brain mapping method. The whole-brain mapping method is as follows: the epileptiformity value of the brain boundary voxel is set to the lowest quantization value; the epileptiformity values ​​of the epileptic focus seizure area contact point voxels and the epileptiformity values ​​of the voxels closest to the epileptic focus seizure area contact points are both set to the epileptiformity quantization values ​​corresponding to the epileptic focus seizure area contact points; the 3D Gaussian kernel algorithm is used to map all other voxel values ​​in the entire brain to the epileptiformity values ​​of the epileptic focus seizure area contact point voxels, the epileptiformity values ​​of the voxels closest to the epileptic focus seizure area contact points, and the epileptiformity values ​​of the brain boundary voxels.

Images