A method and system for assessing car-t treatment-related neurotoxicity based on behavioral characteristics

Abstract

The application discloses a method and system for evaluating CAR-T treatment related neurotoxicity based on behavior characteristics, relates to the technical field of medical behavior analysis, and provides the following scheme, which comprises the following steps: synchronously collecting angular velocity time series data of double upper limbs of a patient and electromyographic signals corresponding to muscle surfaces of the double upper limbs; performing frequency band optimization filtering on the angular velocity time series data of the double upper limbs respectively; generating angular velocity waveforms; taking a wave peak time point of the angular velocity waveform as a reference; extending forward for a preset time length; and calculating an energy integral curve of the electromyographic signals. The angular velocity time series data of the double upper limbs is processed through frequency band optimization to generate angular velocity waveforms, the energy integral curve of the electromyographic signals is extracted forward based on the wave peak time point of the angular velocity, candidate time points are obtained, and local cross-correlation analysis is enabled to locate and intercept points when the wave peak value is lower than a preset threshold, so that the problem of missing weak tremor signals is solved, effective detection and reliable interception of low-amplitude tremors are realized, and weak signal features are avoided from being lost.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical behavior analysis, more particularly, the present application relates to a method and system for evaluating CAR-T treatment related neurotoxicity based on behavior characteristics. BACKGROUND

[0002] CAR-T cell therapy has shown significant efficacy in the treatment of hematological tumors, but some patients will cause immune effector cell related neurotoxicity syndrome, and the incidence of severe neurotoxicity of grade 3 or above is as high as 15%-25%. At present, the clinical mainly relies on subjective neurofunction evaluation tools such as CARTOX-10 scale and modified Rankin score, supplemented by cerebrospinal fluid detection and neuroimaging examination. However, the scale evaluation has a lagging nature, and objective biomarkers are scarce, such as GFAP and NfL, which require invasive sampling and have insufficient specificity, resulting in a serious compression of the early intervention time window, especially for subclinical damage such as cerebellar-cortical pathway conduction delay, and there is a lack of quantifiable and non-invasive dynamic monitoring means.

[0003] The prior art usually adopts the way of counting the frequency or amplitude of limb tremor to evaluate the related neurotoxicity, but due to the lack of time synchronization accuracy, the weak tremor signal is missed, it is difficult to capture the phase difference dynamic characteristics and time-varying nonlinear characteristics of the bilateral limb tremor waveform, and the neurotoxicity caused by CAR-T treatment is difficult to realize non-invasive early specific evaluation, so a method and system for evaluating CAR-T treatment related neurotoxicity based on behavior characteristics are proposed to solve this problem. SUMMARY

[0004] To solve the above technical problems, a method and system for evaluating CAR-T treatment related neurotoxicity based on behavior characteristics are provided, which solve the problems raised in the background art.

[0005] In order to achieve the above purpose, the technical scheme of the present application is as follows:

[0006] In a first aspect, the present application provides a method for evaluating CAR-T treatment related neurotoxicity based on behavior characteristics, which comprises:

[0007] Synchronously collecting angular velocity time series data of the patient's bilateral upper limbs and the corresponding electromyographic signals on the surface of the muscle groups thereof;

[0008] Respectively performing frequency band optimization filtering on the angular velocity time series data of the bilateral upper limbs to generate angular velocity waveforms;

[0009] Taking the wave peak time point of the angular velocity waveform as a reference, extending a preset time length forward and calculating the energy integral curve of the electromyographic signal, and extracting the time point at which the energy integral curve first reaches a target integral threshold as a candidate time point;

[0010] if the peak value of the angular velocity waveform is less than a preset peak threshold, then in a neighborhood time window of the candidate time point, a time point of maximum correlation between the angular velocity waveform and the energy integral curve is determined as the intercept point through local cross-correlation analysis, otherwise the peak time point is taken as the intercept point;

[0011] Based on the intercept points generated independently by the two upper limbs, an intercept point pair is formed, the angular velocity time series data of the two upper limbs are aligned, and an instantaneous phase difference sequence is calculated through Hilbert transform;

[0012] The feature vector of the instantaneous phase difference sequence is extracted and input into a pre-trained neural toxicity grading model to output a neural toxicity risk index.

[0013] In a second aspect, the present application provides a system for evaluating CAR-T treatment-related neural toxicity based on behavioral characteristics, which is used to implement any of the above-mentioned methods for evaluating CAR-T treatment-related neural toxicity based on behavioral characteristics, comprising:

[0014] A signal synchronous acquisition module is configured to synchronously acquire angular velocity time series data of the two upper limbs of the patient and the corresponding electromyographic signals of the muscle groups on the surface thereof;

[0015] A signal preprocessing module is configured to perform frequency band optimization filtering on the angular velocity time series data of the two upper limbs respectively to generate angular velocity waveforms;

[0016] A feature time point extraction module is configured to take the peak time point of the angular velocity waveform as a reference, extend a preset time length forward, and calculate an energy integral curve of the electromyographic signal, and extract a time point at which the energy integral curve first reaches a target integral threshold as a candidate time point;

[0017] A spatio-temporal alignment decision module is configured to, if the peak value of the angular velocity waveform is less than a preset peak threshold, then in a neighborhood time window of the candidate time point, a time point of maximum correlation between the angular velocity waveform and the energy integral curve is determined as the intercept point through local cross-correlation analysis, otherwise the peak time point is taken as the intercept point;

[0018] A two-limb collaborative analysis module is configured to form an intercept point pair based on the intercept points generated independently by the two upper limbs, align the angular velocity time series data of the two upper limbs, and calculate an instantaneous phase difference sequence through Hilbert transform;

[0019] A neural toxicity risk evaluation module is configured to extract a feature vector of the instantaneous phase difference sequence, input the feature vector into a pre-trained neural toxicity grading model, and output a neural toxicity risk index.

[0020] In a third aspect, the present application provides a computer device, which comprises a memory and a processor, the memory stores code, and the processor is configured to acquire the code and execute the above-mentioned method for evaluating CAR-T treatment-related neural toxicity based on behavioral characteristics.

[0021] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method for evaluating CAR-T treatment related neurotoxicity based on behavior characteristics.

[0022] Compared with the prior art, the present application has the following beneficial effects:

[0023] The present application generates angular velocity waveform by frequency band optimization processing double upper limb angular velocity time series data, extracts energy integral curve of electromyographic signal based on angular velocity wave peak time point, obtains candidate time point, and enables local cross-correlation analysis to locate and intercept point when wave peak value is lower than preset threshold, thereby solving the problem of missing weak tremor signal detection, realizing effective detection and reliable interception of low amplitude tremor, and avoiding loss of weak signal characteristics.

[0024] The present application forms a point pair based on the double upper limb interception points, synchronizes the double limb angular velocity time series data, and generates an instantaneous phase difference sequence by using Hilbert transform, overcomes the defect that the dynamic characteristics of the phase of the tremor waveform of the bilateral limbs are difficult to capture, and realizes quantitative characterization of the nerve conduction delay, taking the dynamic phase difference sequence as an objective index of early nerve injury.

[0025] The present application extracts the instantaneous phase difference feature vector containing time-frequency energy skewness variance, sample entropy value and scale index, inputs a pre-trained neurotoxicity grading model, overcomes the defect that the time-varying nonlinear characteristics of the phase difference sequence are difficult to capture, completes multi-dimensional quantitative analysis of non-stationarity, complexity and long-range correlation, and identifies subtle synergistic abnormalities caused by neurotoxicity. BRIEF DESCRIPTION OF DRAWINGS

[0026] The disclosure of the present application will be described with reference to the accompanying drawings. It should be understood that the drawings are only for illustrative purposes, and are not intended to limit the scope of protection of the present application. Among them:

[0027] Figure 1 A flowchart of a method for evaluating CAR-T treatment related neurotoxicity based on behavior characteristics according to the present application is provided;

[0028] Figure 2 A structural block diagram of a system for evaluating CAR-T treatment related neurotoxicity based on behavior characteristics according to the present application is provided. DETAILED DESCRIPTION

[0029] It is easy to understand that, according to the technical solutions of the present application, those skilled in the art can propose various structure modes and implementation modes which can be replaced with each other without changing the essential spirit of the present application. Therefore, the following specific embodiments and drawings are only exemplary descriptions of the technical solutions of the present application, and should not be regarded as the whole or as the limitation or restriction of the technical solutions of the present application.

[0030] In the prior art, the evaluation of neurotoxicity induced by CAR-T cell therapy mainly relies on subjective scales and invasive detection methods, which have the defects of lag and insufficient specificity of biomarkers. The traditional monitoring method uses statistical double limb tremor frequency or amplitude for analysis, but is limited by low time synchronization accuracy and weak signal missing detection, making it difficult to accurately capture the dynamic phase difference characteristics of bilateral limb movement, resulting in the inability to timely identify the subclinical stage of nerve conduction delay damage. For example, when the cerebellum-cortical pathway of the patient has conduction abnormalities, although the coordination of the upper limbs has changed slightly, the prior art lacks non-invasive and high-sensitivity means to quantify such early nerve function damage.

[0031] To solve the above problems, it is found that the change of bilateral upper limb movement coordination can reflect the state of nerve conduction function, but the current monitoring technology is difficult to accurately capture the weak phase dynamic characteristics in the tremor signal. Through analysis, the traditional method has the following key limitations: the time asynchrony of electromyographic signal and movement signal collection causes time alignment deviation; the reliability of effective feature extraction is insufficient under strong noise background; the dynamic phase difference quantification index is lacking in bilateral limb coordination analysis. In view of the above problems, it is necessary to establish a synchronous signal collection system, develop a dynamic positioning method based on waveform feature correlation, and use instantaneous phase difference to realize quantitative evaluation of nerve conduction function.

[0032] Referring to Figure 1 A method for evaluating CAR-T treatment-related neurotoxicity based on behavioral characteristics, comprising:

[0033] Synchronously collecting angular velocity time series data of the patient's bilateral upper limbs and electromyographic signals on the corresponding muscle surface;

[0034] It should be noted that synchronous collection is achieved by sharing a high-precision clock source (sampling rate ≥ 1 kHz), and the electromyographic signal collection module has a built-in hardware trigger. When the rising slope of the electromyographic signal is detected to exceed a preset threshold, the double-channel angular velocity sensor is automatically triggered to start data recording, generating time-stamped angular velocity time series data and electromyographic signals;

[0035] The angular velocity time series data of the bilateral upper limbs are respectively subjected to frequency band optimization filtering to generate angular velocity waveforms;

[0036] Take the peak time point of the angular velocity waveform as the reference, extend forward for a preset time length and calculate the energy integral curve of the electromyographic signal, extract the time point when it first reaches the target integral threshold as the candidate time point;

[0037] If the peak value of the angular velocity waveform is less than the preset peak threshold, then in the neighborhood time window of the candidate time point, the maximum correlation time point of the angular velocity waveform and the energy integral curve is determined through local cross-correlation analysis as the intercept point, otherwise the peak time point is taken as the intercept point;

[0038] Based on the intercept points independently generated by the two upper limbs, an intercept point pair is formed, the angular velocity time series data of the two upper limbs are aligned, and the instantaneous phase difference sequence is calculated through Hilbert transform;

[0039] Exemplarily, the preset time length can be 190 ms, which can cover the maximum neuromuscular conduction delay of most CAR-T neurotoxicity patients and ensure the formation of a cooperative detection range with the neighborhood time window; the preset peak threshold can be 0.3 rad / s;

[0040] Extract the feature vector of the instantaneous phase difference sequence, input it into the pre-trained neurotoxicity grading model, and output the neurotoxicity risk index;

[0041] It should be noted that the neurotoxicity grading model adopts a hybrid architecture of gradient boosting decision tree (GBDT) and time series convolutional network (TCN), and the specific architecture is as follows:

[0042] TCN path: used to receive the feature vector sequence of continuous 3 detections, extract the dynamic evolution features of tremor phase difference through the time domain convolution layer (convolution kernel = 5), and output a 32-dimensional time domain feature vector;

[0043] GBDT path: used to analyze the nonlinear relationship of the three sub-items of the feature vector in single detection, generate 200 decision trees with a depth of 7, and finally concatenate the leaf node hit probabilities of all decision trees into a 10-dimensional static feature vector. The parameter combination of 200 and 7 is determined through grid search and cross-validation, which can achieve an AUC value of 89% on 350 training data, while ensuring that the risk index confidence interval width corresponding to each neurotoxicity grade is less than 5 units;

[0044] Fusion layer: used to perform weighted linear fusion of the time domain feature vector and the static feature vector through a fully connected neural network, and normalize it to a neurotoxicity risk index through a Sigmoid function, wherein the fully connected layer is represented as a linear combination of weight matrix 、 and bias The specific values are determined through historical patient data optimization during model pre-training;

[0045] The weighted linear fusion formula is:

[0046] ;

[0047] wherein, is a neurotoxicity risk index, is a time domain feature vector, is a static feature vector;

[0048] Exemplarily, the neurotoxicity risk index is a [0, 100] continuous value, and the mapping rule is: <30 is level 0, 30-55 is level 1, 55-75 is level 2, 75-90 is level 3, and ≥90 is level 4;

[0049] Through the above technical solution, the application realizes high-precision dynamic monitoring of the coordination of the two upper limbs, can quantize early nerve conduction delay damage in a non-invasive manner, effectively solves the problems of lagging assessment and inconvenient invasive detection of traditional scales, and provides an objective quantitative index for early warning of CAR-T treatment related neurotoxicity.

[0050] In an alternative embodiment, the frequency band optimization filtering employs an adaptive band-pass filter, and the passband range is dynamically adjusted based on the tremor database of the patient, specifically including:

[0051] extracting a power spectral density main peak distribution interval of the tremor waveform data from the tremor database; wherein the power spectral density main peak distribution interval is the frequency range in which the tremor waveform has the most concentrated energy in the frequency domain, and can be obtained by extracting the frequency band range corresponding to the maximum energy peak value after calculating the power spectrum through fast Fourier transform;

[0052] It should be noted that the CAR-T related tremor frequency has individual differences, such as 4-6 Hz for elderly patients and up to 7-8 Hz for young adults. By dynamically adjusting the passband range, the main frequency components of tremor of different patients are ensured to be completely retained. The tremor database contains double-limb angular velocity data of historical CAR-T treated patients, at least more than 100 cases and covering all levels of neurotoxicity;

[0053] processing a preset number of tremor waveform data through a clustering analysis method to determine a low-frequency boundary value and a high-frequency boundary value of the power spectral density main peak distribution interval; wherein the clustering analysis method can employ a K-means algorithm, and the 5% quantile of the frequency point distribution of the largest cluster after clustering is taken as the low-frequency boundary and the 95% quantile is taken as the high-frequency boundary, avoiding accidental interference of a single data sample;

[0054] calculating the difference between the high-frequency boundary value and the low-frequency boundary value, and taking one quarter of the difference as an offset; wherein the setting of the offset can cover more than 90% of the tremor harmonic components of the patients, ensuring that the main frequency and harmonic components of low-amplitude tremor are not lost;

[0055] Compared with the prior art, the conventional fixed passband filter cannot adapt to the individual differences of different patient tremor frequencies, leading to inaccurate signal filtering. By dynamically adjusting the passband range, the filtering interval can be adaptively expanded according to the group characteristics in the historical data, which not only retains the main tremor components but also avoids excessive filtering of the biological electrical signals in the related frequency band, thereby improving the accuracy of subsequent phase difference analysis.

[0056] The low-frequency boundary value minus the offset is taken as the lower limit, and the high-frequency boundary value plus the offset is taken as the upper limit to construct the passband range.

[0057] Through the above technical solution, the present application solves the signal distortion problem caused by the fixed passband of the filter in the prior art. By dynamically adjusting the frequency band based on group tremor waveform data, the electromyographic signals and angular velocity time series of different patients are ensured to be completely retained within the key frequency band, thereby providing a high-quality preprocessing data basis for subsequent time point extraction and phase difference calculation.

[0058] In an optional implementation, a signal validity discrimination based on dynamic noise evaluation is performed on the energy integral curve before extracting the candidate time point at which it first reaches the target integral threshold, specifically including:

[0059] It should be noted that the electromyographic signals of CAR-T patients are often mixed with electrocardiogram interference and motion artifacts, such as sudden noise caused by coughing, which causes the traditional fixed threshold method to easily misjudge noise as valid signals during low-amplitude tremor periods. By dynamically identifying the noise level using a sliding standard deviation, invalid signal segments are avoided from polluting subsequent phase difference analysis.

[0060] A time window of a preset length is used to calculate a sliding standard deviation sequence of the energy integral curve;

[0061] If the standard deviation of the time window is less than a preset noise threshold, the electromyographic signal segment and the corresponding candidate time point of the time window are removed, otherwise the target integral threshold is set to a fixed multiple of the standard deviation. Wherein, the standard deviation less than the preset noise threshold indicates that the window signal fluctuation is too small, which may belong to an invalid resting period or electrode detachment state, and is directly removed to avoid false positives.

[0062] For example, the tremor caused by CAR-T treatment usually occurs in the frequency range of 4-8 Hz with a period of 125-250 ms. Taking 200 ms as the preset length of the time window, 1-2 complete tremor periods are covered to ensure signal feature stability. The preset noise threshold is calibrated by the resting state electromyographic signals of healthy subjects, and the mean value plus 2 times the standard deviation is taken, which can be specifically taken as The fixed multiple can be taken as 1.5, which is commonly used in traditional signal detection Noise is excluded, but the CAR-T related tremor signal has a low signal-to-noise ratio, is a compromise between sensitivity and specificity;

[0063] Through the technical scheme, the application effectively solves the problem of false feature extraction caused by noise interference in the process of collecting myoelectric signals. Through the dual mechanism of dynamic noise evaluation and threshold adjustment, the accuracy of the candidate time point is ensured. For example, when the patient has a short muscle tremor, the false signal generated by the non-autonomous tremor can be accurately identified and excluded.

[0064] In an optional embodiment, the local cross-correlation analysis is implemented by a dynamic time warping algorithm, specifically including:

[0065] In the neighborhood time window, a cost matrix of the angular velocity waveform and the energy integral curve is constructed, and each element value of the cost matrix is the Euclidean distance of the angular velocity waveform and the energy integral curve at the corresponding time point; wherein the angular velocity waveform and the energy integral are normalized to eliminate dimensional differences; for example, the neighborhood time window can be ±150ms, covering the maximum conduction delay of CAR-T tremor;

[0066] The minimum cumulative cost path in the cost matrix is searched, and the time coordinate of the end point of the warping path is determined as the intercept point; wherein the bandwidth of the warping path search can be 20% of the neighborhood time window, the path slope is limited to ≤2 to prevent excessive distortion of the match, and the minimum cumulative cost path avoids local extreme traps to ensure that the intercept point reflects the true physiological event;

[0067] It should be noted that the weak tremor of the CAR-T patient, such as amplitude <0.2 rad / s, has a nonlinear time deformation of the myoelectric signal and the motion signal, such as muscle fatigue causing myoelectric delay. The traditional cross-correlation requires strict synchronization of the signals, and cannot handle the time axis stretching and deformation. The dynamic time warping algorithm tolerates the time axis deformation and accurately locates the delay point of the neuromuscular conduction under weak tremor;

[0068] Specifically, in the neighborhood time window of the candidate time point, the angular velocity waveform and the energy integral curve are intercepted as equal-length time sequence fragments. The Euclidean distance of the two signals at the corresponding time point is calculated point by point to form the element value of the cost matrix. The dynamic time warping algorithm takes the upper left corner of the matrix as the starting point and the lower right corner as the end point, searches for the path with the minimum cumulative cost along the diagonal direction, and allows the signal fragments to be nonlinearly stretched or compressed on the time axis, thereby eliminating the timing deviation caused by muscle contraction delay or signal acquisition error.

[0069] Through the technical scheme, the application effectively solves the problem of insufficient synchronization accuracy caused by time deformation under weak tremor signals. Through the dynamic time warping algorithm, the timing noise interference in the signal acquisition process is eliminated, providing a reliable data basis for accurate calculation of the subsequent instantaneous phase difference sequence.

[0070] In an alternative embodiment, the feature vector of the instantaneous phase difference sequence specifically comprises:

[0071] The continuous wavelet transform is performed on the instantaneous phase difference sequence to extract a time-varying energy distribution matrix thereof in a preset frequency band; the preset frequency band can be specifically 4-8 Hz, covering the main frequency of CAR-T tremor;

[0072] Based on the time-varying energy distribution matrix, a time-varying skewness function is generated by calculating the ratio of the frequency energy weighted sum at each time point to the total energy, and the variance value on the time axis is taken as the time-frequency energy skewness variance;

[0073] Specifically, the time-varying skewness function is:

[0074] ;

[0075] In the formula, is the time-varying skewness function, is the time-frequency energy distribution matrix, is the frequency coordinate, i.e., the frequency value converted from the scale parameter of the wavelet transform, and Ω is the preset frequency band, is the sampling time point of the phase difference sequence;

[0076] The sample entropy value of the instantaneous phase difference sequence is calculated, and the scale index of the instantaneous phase difference sequence is calculated by detrended fluctuation analysis;

[0077] It should be noted that the time-frequency energy skewness variance is used to reflect the stability of bilateral nerve conduction, and the higher the skewness variance, the worse the coordination between the left and right brains; the sample entropy value is used to quantify the complexity of the phase difference sequence, and an increase in the entropy value indicates a disorder of nerve discharge; and the scale index is used to evaluate the long-range correlation of nerve conduction;

[0078] The time-frequency energy skewness variance, the sample entropy value and the scale index are combined as a feature vector;

[0079] Compared with the prior art, the conventional method generally only calculates the mean value of the amplitude or frequency of the double limb tremor, and cannot capture the dynamic time-frequency characteristics and nonlinear dynamic characteristics of the phase difference. The present scheme reflects the dynamic skewness of the energy distribution through the time-frequency energy skewness variance, and combines the sample entropy and the scale index to more comprehensively quantify the non-stationarity, complexity and long-range correlation of the phase difference sequence, thereby revealing subtle abnormalities of nerve conduction delay;

[0080] Through the above technical solutions, the present application can represent the phase coordination abnormality of the double upper limb movement from a multi-dimensional feature perspective, effectively distinguish between physiological tremor and pathological phase desynchronization caused by neurotoxicity, improve the sensitivity and specificity of subclinical nerve injury detection, and provide a quantitative basis for early warning of CAR-T treatment related neurotoxicity.

[0081] In an optional embodiment, after the feature vector is calculated, the re-checking of the intercept point specifically includes:

[0082] When the feature vector satisfies any of the following re-checking trigger conditions, the re-checking process of the intercept point is performed:

[0083] The sample entropy value exceeds a preset entropy threshold value;

[0084] The scale index deviates from a preset tremor fluctuation scale interval;

[0085] The time-frequency energy skewness variance is higher than n times of a preset baseline value;

[0086] Exemplarily, taking the cerebellum-cortex pathway as an example, the preset entropy threshold value can be 1.2, the preset tremor fluctuation scale interval can be [0.5, 1.0], the healthy mean value can be 0.15, and for other neural pathways, adaptive adjustment can be made according to data; the value of n can be 3, and can be dynamically adjusted according to the treatment stage of the patient, such as setting 2.5 in the induction period to improve sensitivity, and setting 3.5 in the remission period to reduce false positives;

[0087] Specifically, when the sample entropy value exceeds the preset entropy threshold value, it indicates that the randomness of the neural conduction signal is abnormally high, which can be caused by signal collection interference or pathological tremor, when the scale index deviates from the preset tremor fluctuation scale interval, it indicates that the self-similarity feature of the tremor waveform does not conform to the normal physiological tremor pattern, and when the time-frequency energy skewness variance exceeds the baseline n times, it indicates that the energy distribution of the bilateral upper limb coordinated movement is significantly unbalanced; At this time, the system automatically triggers the re-checking process, verifies the effectiveness of the intercept point by recalculating the amplitude squared coherence function spectrum, and if the frequency offset or coherence is insufficient, dynamically adjusts the target integral threshold, the preset peak threshold and the neighborhood time window, repositions the intercept point and updates the feature vector.

[0088] Through the above technical solutions, the application can effectively identify feature vector abnormalities caused by signal interference or pathological changes, improve the reliability of intercept point positioning through a dynamic parameter adjustment mechanism, and avoid false phase difference analysis caused by single calculation error.

[0089] In an optional embodiment, the re-checking process of the intercept point specifically includes:

[0090] Based on the electromyographic signal and the angular velocity waveform corresponding to the current intercept point positioning, the amplitude squared coherence function spectrum of both in a preset frequency band is calculated, and the highest coherence peak frequency and the corresponding coherence coefficient are extracted; wherein the preset frequency band is 4-8 Hz;

[0091] If the highest coherence peak frequency deviates from the preset frequency band or the coherence coefficient is less than the preset coefficient threshold, the target integral threshold, the preset peak threshold and the neighborhood time window are respectively shrunk by a preset ratio, and the instantaneous phase difference sequence and the feature vector are updated after the intercept point is re-determined; for example, the coherence coefficient threshold is 0.65, and the preset frequency band is 4Hz-8Hz;

[0092] For example, the parameter shrinkage rule is: the target integral threshold is shrunk by 20%, the myoelectric signal activation detection threshold is reduced, and weak signal missed detection is avoided; the preset peak threshold is shrunk by 15%, which is suitable for low amplitude tremor; the neighborhood time window is shrunk to ±100ms, the search range is reduced, and the alignment accuracy is improved;

[0093] Specifically, the amplitude squared coherence function is:

[0094] ;

[0095] In the formula, is the myoelectric signal power spectrum, is the angular velocity waveform power spectrum, is ;

[0096] Compared with the prior art, the prior method usually only relies on a single signal feature for intercept point determination, does not consider the frequency domain correlation characteristics of myoelectric signals and angular velocity waveforms, and is prone to misjudgment due to noise interference or signal asynchrony. The present scheme can effectively identify and correct the intercept point deviation caused by signal mismatch by introducing amplitude squared coherence function spectrum analysis and combining a dynamic parameter shrinkage mechanism, thereby improving the reliability of dual-limb collaborative analysis.

[0097] Referring to FIG. 1, Figure 2 The present scheme proposes a system for evaluating CAR-T treatment-related neurotoxicity based on behavior characteristics, which is used to implement the above-mentioned method for evaluating CAR-T treatment-related neurotoxicity based on behavior characteristics, and includes:

[0098] A signal synchronous acquisition module is configured to synchronously acquire angular velocity time series data of a patient's double upper limbs and myoelectric signals on the corresponding muscle surfaces of the double upper limbs.

[0099] A signal preprocessing module is configured to perform frequency band optimization filtering on the angular velocity time series data of the double upper limbs respectively, and generate angular velocity waveforms.

[0100] A feature time point extraction module is configured to take the peak time point of the angular velocity waveform as a reference, extend a preset time length forward and calculate an energy integral curve of the myoelectric signals, and extract a time point at which the energy integral curve first reaches a target integral threshold as a candidate time point.

[0101] The time-space alignment decision module is configured to, if the peak value of the angular velocity waveform is less than a preset peak threshold, determine a time point of maximum correlation between the angular velocity waveform and the energy integral curve in a neighborhood time window of the candidate time point as the intercept point through local cross-correlation analysis, or otherwise determine the peak time point as the intercept point;

[0102] The dual-limb coordination analysis module is configured to form an intercept point pair based on the independently generated intercept points of the two upper limbs, align the angular velocity time sequence data of the two upper limbs, and calculate an instantaneous phase difference sequence through Hilbert transform;

[0103] The neurotoxicity risk assessment module is configured to extract a feature vector of the instantaneous phase difference sequence, input the feature vector into a pre-trained neurotoxicity grading model, and output a neurotoxicity risk index.

[0104] In addition, in an embodiment, a computer device is provided, including a memory and a processor, the memory stores a computer program, and the processor implements the steps in the above embodiments when executing the computer program.

[0105] In an embodiment, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the steps in the above embodiments.

[0106] In an embodiment, a computer program product or a computer program is provided, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium. The processor executes the computer instructions, so that the computer device executes the steps in the above embodiments.

[0107] The technical scope of the present application is not limited to the above description. Those skilled in the art can make various modifications and changes to the above embodiments without departing from the technical idea of the present application, and these modifications and changes should be within the protection scope of the present application.

Claims

1. A method for assessing CAR-T treatment-related neurotoxicity based on behavioral characteristics, characterized in that, The method comprises: synchronously collecting angular velocity time series data of the patient's both upper limbs and the corresponding muscle group surface electromyography signals thereof; band-optimized filtering the angular velocity time series data of the both upper limbs respectively to generate angular velocity waveforms; taking the wave peak time point of the angular velocity waveform as a reference, extending a preset time length forward and calculating the energy integral curve of the electromyography signal, extracting the time point at which the energy integral curve first reaches a target integral threshold as a candidate time point; if the wave peak value of the angular velocity waveform is less than a preset wave peak threshold, determining the maximum correlation time point of the angular velocity waveform and the energy integral curve in a neighborhood time window of the candidate time point as a cutting point through local cross-correlation analysis, otherwise taking the wave peak time point as the cutting point; forming a cutting point pair based on the cutting points independently generated for the both upper limbs, aligning the angular velocity time series data of the both upper limbs, and calculating an instantaneous phase difference sequence through Hilbert transform; extracting a feature vector of the instantaneous phase difference sequence, inputting the feature vector into a pre-trained neural toxicity grading model, and outputting a neural toxicity risk index.

2. The method of claim 1, wherein, The band-optimized filtering adopts an adaptive band-pass filter, a passband range of the adaptive band-pass filter is dynamically adjusted based on a tremor database of the patient, and the band-optimized filtering specifically comprises: extracting a power spectral density main peak distribution interval of tremor waveform data from the tremor database; determining a low-frequency boundary value and a high-frequency boundary value of the power spectral density main peak distribution interval through a clustering analysis method for a preset number of tremor waveform data; calculating a difference value between the high-frequency boundary value and the low-frequency boundary value, and taking one quarter of the difference value as an offset; constructing the passband range by taking the low-frequency boundary value minus the offset as a lower limit and the high-frequency boundary value plus the offset as an upper limit.

3. The method of claim 1, wherein, Before the candidate time point at which the energy integral curve first reaches the target integral threshold, performing signal validity discrimination based on dynamic noise evaluation on the energy integral curve, and the signal validity discrimination specifically comprises: calculating a sliding standard deviation sequence of the energy integral curve by using a time window of a preset length; if the standard deviation of the time window is less than a preset noise threshold, discarding the electromyography signal segment and the corresponding candidate time point of the time window, otherwise setting the target integral threshold as a fixed multiple of the standard deviation.

4. The method of claim 1, wherein, The local cross-correlation analysis is implemented through a dynamic time warping algorithm, and the local cross-correlation analysis specifically comprises: in the neighborhood time window, constructing a cost matrix of the angular velocity waveform and the energy integral curve, each element value of the cost matrix being an Euclidean distance of the angular velocity waveform and the energy integral curve at a corresponding time point; searching for a warping path with a minimum accumulated cost in the cost matrix, and determining a time coordinate of an end point of the warping path as the cutting point.

5. The method of claim 1, wherein, The extraction of the feature vector of the instantaneous phase difference sequence specifically comprises: performing continuous wavelet transform on the instantaneous phase difference sequence to extract a time-varying energy distribution matrix thereof in a preset frequency band; based on the time-varying energy distribution matrix, generating a time-varying skewness function by calculating a ratio of a frequency energy weighted sum to total energy at each time point, and taking a variance value on a time axis of the time-varying skewness function as a time-frequency energy skewness variance; calculating a sample entropy value of the instantaneous phase difference sequence, and calculating a scaling exponent of the instantaneous phase difference sequence through detrended fluctuation analysis; combining the time-frequency energy skewness variance, the sample entropy value and the scaling exponent into the feature vector.

6. The method of claim 5, wherein, After the feature vector is calculated, the re-examination of the interception point is performed, and specifically includes: When the feature vector satisfies any of the following re-examination triggering conditions, the re-examination process of the interception point is performed: The sample entropy value exceeds a preset entropy threshold value; The scale index deviates from a preset scale interval of fluctuation of the tremor wave; The time-frequency energy skewness variance is higher than n times of a preset baseline value.

7. The method of claim 6, wherein, The re-examination process of the interception point specifically includes: Based on the electromyographic signals and angular velocity waveforms corresponding to the current interception point positioning, the amplitude squared coherence function spectrum of both in a preset frequency band is calculated, and the highest coherence peak frequency and the corresponding coherence coefficient are extracted; If the highest coherence peak frequency deviates from the preset frequency band or the coherence coefficient is less than a preset coefficient threshold value, the target integral threshold value, the preset peak threshold value and the neighborhood time window are respectively contracted by a preset proportion, the interception point is re-determined, and the instantaneous phase difference sequence and the feature vector are updated. 8.A system for assessing CAR-T treatment related neurotoxicity based on behavioral features, comprising: A method for evaluating CAR-T treatment-related neurotoxicity based on behavior characteristics, as claimed in any one of claims 1-7, comprises: A signal synchronous acquisition module is configured to synchronously acquire angular velocity time series data of both upper limbs of a patient and electromyographic signals corresponding to muscle surfaces of the upper limbs; A signal preprocessing module is configured to perform frequency band optimization filtering on the angular velocity time series data of both upper limbs to generate angular velocity waveforms; A feature time point extraction module is configured to expand a preset time length forward from a peak time point of the angular velocity waveforms and calculate an energy integral curve of the electromyographic signals, and extract a time point at which the energy integral curve first reaches a target integral threshold value as a candidate time point; A time-space alignment decision module is configured to, if a peak value of the angular velocity waveforms is less than a preset peak threshold value, determine a maximum correlation time point of the angular velocity waveforms and the energy integral curve in a neighborhood time window of the candidate time point as an interception point through local cross-correlation analysis, or determine the peak time point as the interception point; A dual-limb coordination analysis module is configured to form an interception point pair based on the independently generated interception points of both upper limbs, align the angular velocity time series data of both upper limbs, and calculate an instantaneous phase difference sequence through Hilbert transform; A neurotoxicity risk evaluation module is configured to extract a feature vector of the instantaneous phase difference sequence, input the feature vector into a pre-trained neurotoxicity grading model, and output a neurotoxicity risk index. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to implement the method of any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the method of any one of claims 1-7.

Images