Method and system for evaluating CAR-T treatment related neurotoxicity based on behavior characteristics

Abstract

The invention discloses a method and a 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: the method comprises the following steps: synchronously collecting angular velocity time sequence data of two upper limbs of a patient and electromyographic signals on the surfaces of corresponding muscle groups; frequency band optimization filtering is conducted on the angular velocity time sequence data of the two upper limbs, angular velocity waveforms are generated, the wave crest time point of the angular velocity waveforms serves as the reference, forward extension is conducted for the preset duration, and an energy integral curve of the electromyographic signals is calculated. The method comprises the following steps: processing angular velocity time sequence data of double upper limbs through frequency band optimization to generate an angular velocity waveform, forward extracting an electromyographic signal energy integral curve based on an angular velocity peak time point to obtain a candidate time point, and starting local cross-correlation analysis to position an interception point when a peak value of the peak is lower than a preset threshold value; therefore, the problem of leak detection of weak tremor signals is solved, effective detection and reliable interception of low-amplitude tremor are realized, and weak signal feature loss is avoided.

Description

Technical Field

[0001] This invention relates to the field of medical behavior analysis technology, and more specifically, this application relates to a method and system for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics. Background Technology

[0002] CAR-T cell therapy has demonstrated significant efficacy in the treatment of hematological malignancies, but some patients experience immune effector cell-related neurotoxicity syndromes, with a severe neurotoxicity rate of ≥ grade 3 as high as 15%-25%. Currently, clinical practice mainly relies on subjective neurological function assessment tools such as the CARTOX-10 scale and modified Rankin Scale, supplemented by cerebrospinal fluid analysis and neuroimaging examinations. However, scale assessments are lagging, and objective biomarkers are scarce. For example, GFAP and NfL require invasive sampling and have insufficient specificity, resulting in a severely compressed early intervention window. In particular, for subclinical damage such as cerebellar-cortical pathway conduction delays, there is a lack of quantifiable, non-invasive dynamic monitoring methods.

[0003] Existing technologies typically assess neurotoxicity by statistically analyzing limb tremor frequency or amplitude. However, due to insufficient time synchronization accuracy and the failure to detect weak tremor signals, it is difficult to capture the dynamic phase difference characteristics and time-varying nonlinear properties of bilateral limb tremor waveforms. This makes it difficult to achieve non-invasive, early, and specific assessment of delayed nerve conduction damage induced by CAR-T therapy. Therefore, a method and system for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics is proposed to solve this problem. Summary of the Invention

[0004] To address the aforementioned technical problems, a method and system for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics are provided. This technical solution resolves the issues raised in the background section.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, this application provides a method for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics, the method comprising:

[0007] Simultaneously collect the angular velocity time-series data of the patient's two upper limbs and the electromyographic signals of the corresponding muscle groups;

[0008] Frequency band optimization filtering was performed on the angular velocity time series data of both upper limbs to generate angular velocity waveforms;

[0009] Using the peak time point of the angular velocity waveform as a reference, the time is extended forward by a preset duration and the energy integral curve of the electromyographic signal is calculated. The time point when it first reaches the target integral threshold is extracted as a candidate time point.

[0010] If the peak value of the angular velocity waveform is less than the preset peak threshold, then within the neighborhood time window of the candidate time point, the time point with the maximum correlation between the angular velocity waveform and the energy integral curve is determined by local cross-correlation analysis as the cutoff point; otherwise, the peak time point is used as the cutoff point.

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

[0012] The feature vectors of the instantaneous phase difference sequence are extracted and input into a pre-trained neurotoxicity grading model, which outputs a neurotoxicity risk index.

[0013] Secondly, this application provides a system for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics, used to implement the method for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics as described in any of the above claims, comprising:

[0014] The signal synchronization acquisition module is used to synchronously acquire the angular velocity time-series data of the patient's two upper limbs and the electromyographic signals of the corresponding muscle groups.

[0015] The signal preprocessing module is used to perform frequency band optimization filtering on the angular velocity time series data of both upper limbs to generate angular velocity waveforms;

[0016] The feature time point extraction module is used to extend forward by a preset time based on the peak time point of the angular velocity waveform and calculate the energy integral curve of the electromyographic signal, and extract the time point when it first reaches the target integral threshold as the candidate time point.

[0017] The spatiotemporal alignment decision module is used to determine the time point with the greatest correlation between the angular velocity waveform and the energy integral curve within the neighborhood time window of the candidate time point if the peak value of the angular velocity waveform is less than the preset peak threshold. Otherwise, the peak time point is used as the cutoff point.

[0018] The dual-limb collaborative analysis module is used to form intercept point pairs 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 the instantaneous phase difference sequence through Hilbert transform;

[0019] The neurotoxicity risk assessment module is used to 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.

[0020] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described feedback adjustment method based on haptic interaction.

[0021] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described feedback adjustment method based on haptic interaction.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] This application generates angular velocity waveforms by optimizing the timing data of angular velocity of both upper limbs through frequency band optimization. Based on the peak time point of the angular velocity wave, it extracts the energy integral curve of the electromyographic signal to obtain candidate time points. When the peak value is lower than a preset threshold, it enables local cross-correlation analysis to locate the interception point, thereby solving the problem of missing weak tremor signals and achieving effective detection and reliable interception of low-amplitude tremors, avoiding the loss of weak signal features.

[0024] This application forms point pairs based on the amputation points of both upper limbs, synchronizes the temporal data of the angular velocity of both limbs, and uses Hilbert transform to generate an instantaneous phase difference sequence. This overcomes the defect that it is difficult to capture the dynamic phase characteristics of bilateral limb tremor waveforms, realizes the quantitative characterization of nerve conduction delay, and uses the dynamic phase difference sequence as an objective indicator of early nerve injury.

[0025] This application overcomes the difficulty in capturing the time-varying nonlinear characteristics of phase difference sequences by extracting instantaneous phase difference feature vectors containing time-frequency energy skewness variance, sample entropy value, and scaling exponent, and inputting them into a pre-trained neurotoxicity grading model. It completes multidimensional quantitative analysis of nonstationarity, complexity, and long-range correlation, and identifies subtle synergistic anomalies caused by neurotoxicity. Attached Figure Description

[0026] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein:

[0027] Figure 1 This is a flowchart of a method for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics, as proposed in this invention.

[0028] Figure 2 This is a structural block diagram of a system for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics, as proposed in this invention. Detailed Implementation

[0029] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0030] In existing technologies, the assessment of neurotoxicity induced by CAR-T cell therapy mainly relies on subjective scales and invasive testing methods, which suffer from drawbacks such as lag and insufficient biomarker specificity. Traditional monitoring methods analyze the statistical frequency or amplitude of bilateral limb tremors, but are limited by low time synchronization accuracy and the failure to detect weak signals, making it difficult to accurately capture the dynamic phase difference characteristics of bilateral limb movements. This leads to the failure to promptly identify subclinical neural conduction delay damage. For example, when a patient experiences conduction abnormalities in the cerebellar-cortical pathway, although subtle changes have occurred in the coordination of bilateral upper limb motor functions, current technologies lack non-invasive, highly sensitive methods to quantify such early neurological functional impairment.

[0031] To address these issues, research has found that changes in the synergy of bilateral upper limb motor movements can reflect the state of nerve conduction function; however, current monitoring technologies struggle to accurately capture the subtle dynamic phase features in tremor signals. Analysis reveals the following key limitations of traditional methods: the asynchronous acquisition of electromyographic and motor signals causes timing alignment errors; the reliability of effective feature extraction is insufficient in noisy environments; and bilateral synergy analysis lacks a quantitative index of dynamic phase difference. To resolve these problems, it is necessary to establish a synchronous signal acquisition system, develop a dynamic localization method for intercept points based on waveform feature correlation, and utilize instantaneous phase difference to achieve a quantitative assessment of nerve conduction function.

[0032] Reference Figure 1 As shown, a method for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics includes:

[0033] Simultaneously collect the angular velocity time-series data of the patient's two upper limbs and the electromyographic signals of the corresponding muscle groups;

[0034] It should be noted that synchronous acquisition is achieved by sharing a high-precision clock source (sampling rate ≥ 1kHz), and the electromyography signal acquisition module has a built-in hardware trigger. When the slope of the rising edge of the electromyography signal is detected to exceed the preset threshold, the dual-channel angular velocity sensor is automatically triggered to synchronously start data recording, generating time-stamp-aligned angular velocity timing data and electromyography signal.

[0035] Frequency band optimization filtering was performed on the angular velocity time series data of both upper limbs to generate angular velocity waveforms;

[0036] Using the peak time point of the angular velocity waveform as a reference, the time is extended forward by a preset duration and the energy integral curve of the electromyographic signal is calculated. The time point when it first reaches the target integral threshold is extracted as a candidate time point.

[0037] If the peak value of the angular velocity waveform is less than the preset peak threshold, then within the neighborhood time window of the candidate time point, the time point with the maximum correlation between the angular velocity waveform and the energy integral curve is determined by local cross-correlation analysis as the cutoff point; otherwise, the peak time point is used as the cutoff point.

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

[0039] For example, the preset duration can be 190ms, which can cover the maximum neuromuscular conduction delay of most CAR-T neurotoxic patients and ensure a synergistic detection range with the neighboring time window; the preset peak threshold can be 0.3rad / s.

[0040] The feature vectors of the instantaneous phase difference sequence are extracted and input into the pre-trained neurotoxicity grading model to 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 temporal convolutional network (TCN), specifically as follows:

[0042] TCN pathway: used to receive feature vector sequences from three consecutive detections, extract the dynamic evolution features of the tremor phase difference through a temporal convolutional layer (convolutional kernel=5), and output a 32-dimensional temporal feature vector;

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

[0044] Fusion Layer: Used to perform weighted linear fusion of temporal feature vectors and static feature vectors through a fully connected neural network, and normalize it to a neurotoxicity risk index using the Sigmoid function. The fully connected layer is represented by the weight matrix. , and bias The specific values ​​of the linear combination are determined through optimization using historical patient data during the model pre-training phase.

[0045] The weighted linear fusion formula is:

[0046] ;

[0047] In the formula, This is a neurotoxicity risk index. For time-domain feature vectors, These are static feature vectors;

[0048] For example, the neurotoxicity risk index is a continuous value of [0, 100], 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, this application achieves high-precision dynamic monitoring of the motor coordination of both upper limbs, and can quantify early nerve conduction delay damage in a non-invasive manner. This method effectively solves the problems of lagging assessment and inconvenience of invasive detection in traditional scales, and provides objective quantitative indicators for early warning of CAR-T therapy-related neurotoxicity.

[0050] In one alternative implementation, the frequency band optimization filtering employs an adaptive bandpass filter whose passband range is dynamically adjusted based on the patient's tremor database, specifically including:

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

[0052] It should be noted that there are individual differences in the frequency of CAR-T-related tremor. For example, elderly patients are mostly 4-6 Hz, while young adults can reach 7-8 Hz. By dynamically adjusting the passband range, we can ensure that the dominant frequency component of the tremor is completely preserved in different patients. The tremor database contains angular velocity data of the upper limbs of patients who have received historical CAR-T treatment, with at least more than 100 cases and covering all levels of neurotoxicity.

[0053] Cluster analysis is used to process a preset number of tremor waveform data to determine the low-frequency and high-frequency boundary values ​​of the main peak distribution range of the power spectral density. The cluster analysis method can use the K-means algorithm, with the main peak frequency of the power spectrum as the feature vector, and the 5th percentile of the frequency distribution of the largest cluster after clustering is taken as the low-frequency boundary and the 95th percentile as the high-frequency boundary, to avoid the random interference of a single data sample.

[0054] Calculate the difference between the high-frequency boundary value and the low-frequency boundary value, and take one-quarter of it as the offset; the offset setting can cover the tremor harmonic components of more than 90% of patients, ensuring that the dominant frequency and harmonic components of low-amplitude tremor are not lost.

[0055] Compared with existing technologies, traditional fixed passband filters cannot adapt to the individual differences in tremor frequency among different patients, resulting in inaccurate signal filtering. By dynamically adjusting the passband range, the filtering interval can be adaptively expanded according to the group characteristics in historical data, which not only retains the main tremor components, but also avoids excessive filtering of bioelectrical signals in related frequency bands, thereby improving the accuracy of subsequent phase difference analysis.

[0056] The passband range is constructed by subtracting the offset from the low-frequency boundary value as the lower limit and adding the offset to the high-frequency boundary value as the upper limit.

[0057] Through the above technical solution, this 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 the group tremor waveform data, it ensures that the electromyographic signals and angular velocity timing of different patients are completely preserved in the key frequency band, providing a high-quality preprocessing data foundation for subsequent time point interception and phase difference calculation.

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

[0059] It should be noted that the electromyography signals of CAR-T patients are often mixed with electrocardiogram interference and motion artifacts, such as sudden noise caused by coughing. This makes it easy for the traditional fixed threshold method to misjudge noise as a valid signal during low-amplitude tremor periods. The noise level can be dynamically identified by sliding standard deviation to avoid invalid signal segments from contaminating subsequent phase difference analysis.

[0060] The sliding standard deviation sequence of the energy integral curve is calculated using a time window of preset length;

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

[0062] For example, CAR-T therapy-induced tremors typically exhibit a frequency range of 4-8 Hz and a period of 125-250 ms. A preset time window length of 200 ms is used to cover 1-2 complete tremor cycles, ensuring signal characteristic stability. The preset noise threshold is calibrated using resting-state electromyography signals from healthy subjects, taking the mean plus twice the standard deviation; specifically, it can be set to... A fixed multiplier of 1.5 is commonly used in traditional signal detection. Even after excluding noise, CAR-T-related flutter signals have a low signal-to-noise ratio. It is a compromise between balancing sensitivity and specificity;

[0063] Through the above technical solution, this application effectively solves the problem of feature mis-extraction caused by noise interference during electromyography signal acquisition. Through the dual mechanism of dynamic noise assessment and threshold adjustment, the accuracy of candidate time points is ensured. For example, when a patient experiences brief muscle tremors, it can accurately identify and exclude spurious signals generated by involuntary tremors.

[0064] In one alternative implementation, local cross-correlation analysis is performed using a dynamic time warping algorithm, specifically including:

[0065] Within the neighborhood time window, a cost matrix is ​​constructed for the angular velocity waveform and the energy integral curve. Each element of the cost matrix is ​​the Euclidean distance between the angular velocity waveform and the energy integral curve at the corresponding time point. The angular velocity waveform and the energy integral are normalized to eliminate dimensional differences. For example, the neighborhood time window can be ±150ms to cover the maximum conduction delay of CAR-T tremor.

[0066] The search finds the regularized path with the minimum cumulative cost in the cost matrix and determines the time coordinate of the end point of the regularized path as the cut-off point. The bandwidth of the regularized path search can be 20% of the neighborhood time window, and the path slope is limited to ≤2 to prevent excessive distorted matching. The path with the minimum cumulative cost avoids local extremum traps and ensures that the cut-off point reflects the real physiological event.

[0067] It should be noted that in CAR-T patients with weak tremors, such as when the amplitude is <0.2 rad / s, there is a non-linear time distortion between electromyographic signals and motor signals. For example, muscle fatigue can cause delays in electromyographic signals. Traditional cross-correlation requires strict synchronization of signals and cannot handle time axis stretching and deformation. By using dynamic time warping algorithm to tolerate time axis deformation, the delay point of neuromuscular conduction under weak tremors can be accurately located.

[0068] Specifically, within the neighborhood time window of the candidate time point, the angular velocity waveform and energy integral curve are truncated into time series segments of equal length. The cost matrix element values ​​are formed by calculating the Euclidean distance between the two signals at corresponding time points point by point. The dynamic time warping algorithm searches for the path with the minimum cumulative cost along the diagonal direction, starting from the upper left corner of the matrix and ending at the lower right corner. This path allows the signal segment 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 above technical solution, this application effectively solves the problem of insufficient synchronization accuracy caused by time deformation under weak flutter signals. By using a dynamic time warping algorithm to eliminate temporal noise interference in the signal acquisition process, it provides a reliable data foundation for the accurate calculation of the subsequent instantaneous phase difference sequence.

[0070] In one optional implementation, extracting the feature vector of the instantaneous phase difference sequence specifically includes:

[0071] A continuous wavelet transform is performed on the instantaneous phase difference sequence to extract its time-varying energy distribution matrix in a preset frequency band; wherein, the preset frequency band can be specifically taken as 4-8Hz, covering the CAR-T flutter main frequency;

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

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

[0074] ;

[0075] In the formula, The time-varying skewness function, The time-frequency energy distribution matrix, The frequency coordinates are the frequency values ​​transformed by the scale parameter of the wavelet transform, where Ω represents the preset frequency band. These are the sampling time points of the phase difference sequence;

[0076] Calculate the sample entropy value of the instantaneous phase difference sequence, and calculate the scaling exponent of the instantaneous phase difference sequence through detrended fluctuation analysis;

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

[0078] The time-frequency energy skewness variance, sample entropy value, and scaling exponent are combined into a feature vector;

[0079] Compared with existing technologies, traditional methods usually only count the mean of the amplitude or frequency of bilateral tremors, which cannot capture the dynamic time-frequency characteristics and nonlinear dynamic properties of phase difference. This scheme reflects the dynamic skewness of energy distribution through time-frequency energy skewness variance. Combined with sample entropy and scaling exponent, it can more comprehensively quantify the non-stationarity, complexity and long-range correlation of phase difference sequences, thereby revealing subtle abnormalities in neural conduction delay.

[0080] Through the above technical solutions, this application can characterize the phase synergy abnormality of bilateral upper limb movements from a multidimensional feature perspective, effectively distinguish between physiological tremor and pathological phase desynchronization caused by neurotoxicity, improve the sensitivity and specificity of subclinical neurological injury detection, and provide quantitative basis for early warning of CAR-T therapy-related neurotoxicity.

[0081] In one optional implementation, after the feature parameters are calculated, the re-detection of the cutoff point specifically includes:

[0082] When the feature vector satisfies any of the following re-detection trigger conditions, the cut-off point re-detection process is executed:

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

[0084] The scaling index deviates from the preset flutter fluctuation scaling range;

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

[0086] For example, taking the cerebellar-cortical pathway as an example, the preset entropy threshold can be 1.2, the preset tremor fluctuation scale range can be set to [0.5, 1.0], and the healthy mean can be 0.15. For other neural pathways, adaptive adjustments can be made based on the data. The n value can be 3 and can be dynamically adjusted according to the patient's treatment stage. For example, the induction period can be set to 2.5 to improve sensitivity, and the remission period can be set to 3.5 to reduce false positives.

[0087] Specifically, when the sample entropy value exceeds the preset entropy threshold, it indicates an abnormally high randomness in the neural conduction signal, which may be caused by signal acquisition interference or pathological tremor. When the scaling index deviates from the preset tremor fluctuation scaling range, it indicates that the self-similarity characteristics of the tremor waveform do not conform to the normal physiological tremor pattern. When the time-frequency energy skewness variance exceeds n times the baseline, it indicates a significant imbalance in the energy distribution of the coordinated movement of both upper limbs. At this time, the system automatically triggers the re-examination process, verifies the validity of the intercept point by recalculating the amplitude squared coherence function spectrum. If frequency offset or insufficient coherence is found, the target integral threshold, preset peak threshold, and neighborhood time window are dynamically adjusted to relocate the intercept point and update the feature vector.

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

[0089] In one optional implementation, the intercept point re-examination process specifically includes:

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

[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 shrunk by a preset ratio respectively, and the intercept point is re-determined and the instantaneous phase difference sequence and feature vector are updated; for example, the coherence coefficient threshold is 0.65 and the preset frequency band is 4Hz~8Hz;

[0092] For example, the parameter shrinkage rules are as follows: the target integral threshold is shrunk by 20% to reduce the detection threshold of electromyographic signal activation and avoid missing weak signals; the preset peak threshold is shrunk by 15% to adapt to low-amplitude tremors; the neighborhood time window is shrunk to ±100ms to narrow the search range and improve alignment accuracy.

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

[0094] ;

[0095] In the formula, The power spectrum of electromyographic signals. The power spectrum of the angular velocity waveform. for ;

[0096] Compared with existing technologies, existing methods usually rely on only a single signal feature to determine the intercept point, without considering the frequency domain correlation characteristics of electromyographic signals and angular velocity waveforms. This makes them prone to misjudgment due to noise interference or signal asynchrony. This solution introduces amplitude squared coherence function spectrum analysis and combines it with dynamic parameter contraction mechanism to effectively identify and correct intercept point deviations caused by signal mismatch, thereby improving the reliability of bilimb collaborative analysis.

[0097] See Figure 2 As shown, this scheme proposes a system for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics, to implement the aforementioned method for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics, including:

[0098] The signal synchronization acquisition module is used to synchronously acquire the angular velocity time-series data of the patient's two upper limbs and the electromyographic signals of the corresponding muscle groups.

[0099] The signal preprocessing module is used to perform frequency band optimization filtering on the angular velocity time series data of both upper limbs to generate angular velocity waveforms;

[0100] The feature time point extraction module is used to extend forward by a preset time based on the peak time point of the angular velocity waveform and calculate the energy integral curve of the electromyographic signal, and extract the time point when it first reaches the target integral threshold as the candidate time point.

[0101] The spatiotemporal alignment decision module is used to determine the time point with the greatest correlation between the angular velocity waveform and the energy integral curve within the neighborhood time window of the candidate time point if the peak value of the angular velocity waveform is less than the preset peak threshold. Otherwise, the peak time point is used as the cutoff point.

[0102] The dual-limb collaborative analysis module is used to form intercept point pairs 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 the instantaneous phase difference sequence through Hilbert transform;

[0103] The neurotoxicity risk assessment module is used to 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.

[0104] In another embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0105] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described above.

[0106] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including 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, and executes the computer instructions, causing the computer device to perform the steps described above.

[0107] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics, characterized in that, The method includes: Simultaneously collect the angular velocity time-series data of the patient's two upper limbs and the electromyographic signals of the corresponding muscle groups; Frequency band optimization filtering was performed on the angular velocity time series data of both upper limbs to generate angular velocity waveforms; Using the peak time point of the angular velocity waveform as a reference, the time is extended forward by a preset duration and the energy integral curve of the electromyographic signal is calculated. The time point when it first reaches the target integral threshold is extracted as a candidate time point. If the peak value of the angular velocity waveform is less than the preset peak threshold, then within the neighborhood time window of the candidate time point, the time point with the maximum correlation between the angular velocity waveform and the energy integral curve is determined by local cross-correlation analysis as the cutoff point; otherwise, the peak time point is used as the cutoff point. Based on the interception points generated independently by both upper limbs, interception point pairs are formed, the angular velocity time series data of both upper limbs are aligned, and the instantaneous phase difference sequence is calculated through Hilbert transform; The feature vectors of the instantaneous phase difference sequence are extracted and input into a pre-trained neurotoxicity grading model, which outputs a neurotoxicity risk index.

2. The method according to claim 1, characterized in that, The frequency band optimization filtering employs an adaptive bandpass filter, whose passband range is dynamically adjusted based on the patient's tremor database, specifically including: Extract the main peak distribution range of the power spectral density of tremor waveform data from the tremor database; By processing a predetermined number of tremor waveform data using cluster analysis, the low-frequency and high-frequency boundary values ​​of the main peak distribution range of the power spectral density are determined. Calculate the difference between the high-frequency boundary value and the low-frequency boundary value, and take one-quarter of it as the offset; The passband range is constructed by subtracting the offset from the low-frequency boundary value as the lower limit and adding the offset to the high-frequency boundary value as the upper limit.

3. The method according to claim 1, characterized in that, Before the candidate time point when the extracted energy integration curve first reaches the target integration threshold, a signal validity judgment based on dynamic noise assessment is performed on the energy integration curve, specifically including: The sliding standard deviation sequence of the energy integral curve is calculated using a time window of preset length; If the standard deviation of the time window is less than the preset noise threshold, the electromyographic signal segment and the corresponding candidate time point of that time window are removed; otherwise, the target integration threshold is set to a fixed multiple of the standard deviation.

4. The method according to claim 1, characterized in that, The local cross-correlation analysis is implemented using a dynamic time warping algorithm, specifically including: Within the neighborhood time window, a cost matrix is ​​constructed for the angular velocity waveform and the energy integral curve. Each element of the cost matrix is ​​the Euclidean distance between the angular velocity waveform and the energy integral curve at the corresponding time point. Search for the regularized path with the minimum cumulative cost in the cost matrix, and determine the time coordinate of the end point of the regularized path as the intercept point.

5. The method according to claim 1, characterized in that, The extraction of the feature vector of the instantaneous phase difference sequence specifically includes: Perform continuous wavelet transform on the instantaneous phase difference sequence to extract its time-varying energy distribution matrix in a preset frequency band; Based on the time-varying energy distribution matrix, a time-varying skewness function is generated by calculating the ratio of the frequency domain energy weighted sum to the total energy at each time point, and its variance value on the time axis is recorded as the time-frequency energy skewness variance. Calculate the sample entropy value of the instantaneous phase difference sequence, and calculate the scaling exponent of the instantaneous phase difference sequence through detrended fluctuation analysis; The time-frequency energy skewness variance, sample entropy value, and scaling exponent are combined into a feature vector.

6. The method according to claim 5, characterized in that, After the feature parameters are calculated, the re-detection of the cut-off point specifically includes: When the feature vector satisfies any of the following re-detection trigger conditions, the cut-off point re-detection process is executed: The sample entropy value exceeds a preset entropy threshold; The scaling index deviates from the preset flutter fluctuation scaling range; The time-frequency energy skewness variance is higher than n times the preset baseline value.

7. The method according to claim 6, characterized in that, The intercept point re-inspection process specifically includes: Based on the electromyographic signal and angular velocity waveform corresponding to the current intercept point location, calculate the amplitude squared coherence function spectrum of the two in the preset frequency band, and extract the highest coherence peak frequency and the corresponding coherence coefficient. 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 shrunk by the preset ratio respectively. After re-determining the intercept point, the instantaneous phase difference sequence and feature vector are updated.

8. A system for assessing CAR-T therapy-related neurotoxicity based on behavioral characteristics, characterized in that, A method for implementing a behavioral characteristic-based assessment of CAR-T therapy-related neurotoxicity as described in any one of claims 1-7, comprising: The signal synchronization acquisition module is used to synchronously acquire the angular velocity time-series data of the patient's two upper limbs and the electromyographic signals of the corresponding muscle groups. The signal preprocessing module is used to perform frequency band optimization filtering on the angular velocity time series data of both upper limbs to generate angular velocity waveforms; The feature time point extraction module is used to extend forward by a preset time based on the peak time point of the angular velocity waveform and calculate the energy integral curve of the electromyographic signal, and extract the time point when it first reaches the target integral threshold as the candidate time point. The spatiotemporal alignment decision module is used to determine the time point with the greatest correlation between the angular velocity waveform and the energy integral curve within the neighborhood time window of the candidate time point if the peak value of the angular velocity waveform is less than the preset peak threshold. Otherwise, the peak time point is used as the cutoff point. The dual-limb collaborative analysis module is used to form intercept point pairs 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 the instantaneous phase difference sequence through Hilbert transform; The neurotoxicity risk assessment module is used to 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.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images