Lightweight anesthesia wake-up early warning method and system based on heart-brain linkage timing characteristics

Abstract

This invention relates to the field of medical monitoring technology, specifically a lightweight anesthesia awakening early warning method and system based on the temporal characteristics of heart-brain linkage. The method includes the following steps: extracting the peak difference of electrocardiograms to construct a beat interval sequence; reconstructing a multidimensional phase space to calculate the probability logarithmic ratio and extracting polarity reversal density features; extracting decentralized third-order cumulants and autocorrelation to generate decay features; combining multidimensional features with a regression model to perform mapping and generate a trend inference vector; and determining the monitoring window to issue an awakening early warning command. In this invention, early weak variations are captured by mining the micro-temporal nonlinear evolution characteristics, and the continuous step size decay amplitude is calculated to eliminate redundant interference, reducing data load and ensuring a lightweight response. Simultaneously, the inference vector is used to dynamically delineate the risk monitoring window to accurately adapt to individual patient metabolic differences, fundamentally eliminating the lag and false alarm defects caused by fixed thresholds and comprehensively improving perioperative life safety.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, and in particular to a lightweight anesthesia awakening early warning method and system based on the temporal characteristics of heart-brain linkage. Background Technology

[0002] The field of medical monitoring technology involves core aspects such as continuous acquisition of vital signs, analysis of physiological signals, assessment of clinical status, and medical early warning and control. It mainly includes electrocardiogram (ECG) signal acquisition, electroencephalogram (EEG) signal acquisition, blood oxygen saturation monitoring, respiratory rate monitoring, pulse wave detection, and synchronous recording of multi-parameter physiological data. Through electrode patches, EEG leads, finger clip sensors, bedside monitoring equipment, and time-series data recording devices, it continuously acquires physiological changes of patients during the perioperative period, intensive care, and anesthesia recovery stages. It also analyzes patients' states of consciousness, circulatory status, and neural activity status by combining time series comparison, waveform feature extraction, threshold determination, and state grading.

[0003] Among them, the traditional lightweight anesthesia awakening early warning method based on the temporal characteristics of heart-brain linkage refers to the method of simultaneously collecting EEG waveforms, ECG waveforms, heart rate variability data and blood oxygen parameters during the patient's anesthesia recovery phase. It is based on the changes in EEG frequency bands, RR interval changes, brain-heart rhythm synchronization relationship and the characteristic sequence corresponding to the awakening phase within different time windows. The method uses frontal EEG electrodes to obtain delta wave, theta wave and alpha wave change data, uses precordial leads to collect ECG temporal signals, and continuously records the EEG power spectrum, heart rate fluctuation range and brain-heart coupling temporal sequence according to a fixed sampling period. Then, by setting a wake-up threshold, comparing stage features and judging time period trends, it identifies the premature awakening state that occurs during the anesthesia recovery process.

[0004] Traditional cardiac-brain linkage anesthesia awakening early warning methods rely on fixed sampling cycles to extract EEG spectra and compare them with heart rate intervals. This static extraction and fixed feature matching mode is difficult to capture the nonlinear dynamic evolution of physiological signals at the micro-temporal level. It is easy to ignore the early accumulation process of weak variation fluctuations. Multi-parameter synchronous evaluation calculations consume a lot of memory resources, thus hindering high-frequency real-time monitoring response. When facing the anesthetic metabolic process with huge individual differences, relying on stage threshold judgment is prone to delayed inference or false alarm risk, directly exacerbating the safety risks and clinical monitoring burden of perioperative patients. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as traditional cardio-brain linkage anesthesia recovery early warning methods that rely on fixed sampling cycles to extract EEG spectra and compare them with heart rate intervals, this static extraction and fixed feature matching mode struggles to capture the nonlinear dynamic evolution of physiological signals at the microscopic temporal level. It easily overlooks the early accumulation of subtle fluctuations, and the multi-parameter synchronous evaluation computation is extremely memory-intensive, hindering high-frequency real-time monitoring responses. Furthermore, relying on stage thresholds to determine the anesthetic metabolic process, which exhibits significant individual differences, can easily lead to delayed inferences or false alarms, directly exacerbating perioperative patient safety risks and the burden of clinical monitoring. Therefore, this invention provides a lightweight anesthesia recovery early warning method based on cardio-brain linkage temporal characteristics.

[0006] To achieve the above objectives, this invention employs a lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage, comprising the following steps: S1: Collect ECG waveform data of patients in the anesthesia recovery period, segment the ECG waveform data according to the time window, detect the peak timestamp of the ECG waveform data within the time window, and perform a difference operation on adjacent timestamps to generate a beat interval sequence; S2: Perform phase space reconstruction operation on the jump interval sequence, calculate the probability log ratio of neighboring point pairs in the multidimensional state space to extract sample entropy values, count the number of times adjacent sign bits of sample entropy values ​​switch between positive and negative, and generate polarity flip density features. S3: Perform decentralization processing on the jump interval sequence and extract the third-order cumulant that represents the asymmetric distribution of the waveform. Evaluate the autocorrelation coefficient of the third-order cumulant under continuous delay steps. Perform attenuation amplitude convergence processing on the autocorrelation coefficient under adjacent delay steps to generate autocorrelation attenuation features. S4: Concatenate the polarity reversal density feature and the autocorrelation decay feature to construct a multi-dimensional joint feature space. Input the multi-dimensional joint feature space into the support vector regression model to perform kernel function mapping operation. Combine the preset mapping coefficients to perform bias compensation and generate a risk trend inference vector. S5: Analyze the risk trend inference vector to obtain the state inference value and time offset parameter. When the state inference value exceeds the preset light anesthesia judgment threshold, determine the high-risk monitoring window based on the current time and time offset parameter, and generate an anesthesia awakening warning instruction within the high-risk monitoring window.

[0007] As a further aspect of the present invention, the beat interval sequence includes heart rate interval, heart rate variability, and periodic volatility; the polarity reversal density feature includes sign change rate, sample entropy value, and complexity index; the autocorrelation decay feature includes autocorrelation decay coefficient, autocorrelation function value, and time delay correlation coefficient; the risk trend inference vector includes risk score, trend slope, and deviation correction value; and the anesthesia awakening warning instruction includes warning level identifier, risk duration, and monitoring response category.

[0008] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Collect continuous electrocardiogram (ECG) data from the patient's body surface, retrieve the potential amplitude and corresponding time sequence markers within the continuous ECG data, perform a truncation operation on the continuous ECG data with a fixed step size using window scale limiting parameters, aggregate all sampled feature points falling within the same time span, and obtain a segmented waveform feature matrix; S102: Extract the extreme values ​​of potential amplitude of each row item in the segmented waveform feature matrix, compare the extreme values ​​with the preset peak determination benchmark threshold, retain the items corresponding to extreme values ​​greater than the preset peak determination benchmark threshold and retrieve the corresponding time stamps, arrange the corresponding time stamps according to the incrementing rule, and obtain the peak point timestamp sequence. S103: Extract the record values ​​of adjacent nodes in the timestamp sequence of the peak point, perform difference calculation by subtracting the timestamp value of the preceding node from the timestamp value of the subsequent node, and write the time difference variables output by all difference calculations into the chained cache space one by one according to the node arrangement order to generate the jump interval sequence.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Call the jump interval sequence, obtain the embedding dimension parameter and delay time parameter, perform a truncation and shift operation on the jump interval sequence, construct the reconstructed phase space matrix, calculate the Chebyshev distance between row vectors in the reconstructed phase space matrix, aggregate vector pairs whose Chebyshev distance is less than a preset tolerance threshold, divide the number of vector pairs by the total number of pairings, and obtain the probability set of neighboring points. S202: Based on the step size parameter, embed the dimension parameter, re-execute vector recombination and distance comparison to obtain the expanded dimension probability set, determine the ratio of the values ​​in the neighboring point probability set to the values ​​in the expanded dimension probability set to obtain the probability ratio, perform natural logarithm operation and extract the absolute value, and establish the sample entropy state sequence. S203: Based on the sample entropy state sequence, extract the difference between the node value and the overall mean, establish a positive and negative sign label column and retrieve adjacent items in the time series, compare the polarity state, accumulate the reverse switching times to obtain the cumulative flip times, and calculate the ratio of the cumulative flip times to the total number of node records to generate the polarity flip density feature.

[0010] As a further aspect of the present invention, the preset tolerance threshold is determined based on the ascending order of the Chebyshev distances between row vectors in the reconstructed phase space matrix, and the preset tolerance threshold is selected between 1 / 10 and 1 / 2 of the total number of Chebyshev distances between row vectors in the reconstructed phase space matrix.

[0011] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: For the jump interval sequence, calculate the expected parameter, extract the difference between the node value and the expected parameter, map the node difference state to construct a centralized deviation matrix, perform a cube operation on the elements in the centralized deviation matrix to obtain the sum parameter, calculate the skewness parameter based on the ratio of the sum parameter to the total number of nodes, and obtain the third-order cumulant parameter. S302: Call the third-order cumulant parameter to obtain the step size set, extract the mapping feature column based on the numerical translation of the third-order cumulant parameter within the step size set, calculate the sum of the product of the third-order cumulant parameter and the corresponding nodes of the mapping feature column, perform standardization processing based on the ratio of the product sum to the node variance, and establish an autocorrelation coefficient sequence. S303: Retrieve adjacent nodes in the autocorrelation coefficient sequence arranged in ascending order of time, perform numerical subtraction operation between adjacent nodes to obtain discrete differences, extract the absolute value of discrete differences to obtain the attenuation amplitude, summarize the attenuation amplitudes corresponding to the step size to construct an attenuation feature library, and solve the mean value in the attenuation feature library to generate autocorrelation attenuation features.

[0012] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: For the polarity reversal density feature and the autocorrelation decay feature, extract the dimension and record length, align the nodes within the dimension according to the record length, establish a multi-dimensional spliced ​​data architecture, map the features to the adjacent channels of the multi-dimensional spliced ​​data architecture in order, and fuse the node attributes to establish a multi-dimensional joint feature space. S402: Call the multidimensional joint feature space, obtain the preset support vector set and kernel function scale parameter, calculate the Euclidean distance between the node in the feature space and each vector in the support vector set, construct the proportional relationship between the squared distance value and the kernel function scale parameter to obtain the intermediate quotient value, perform the natural exponent negative operation to extract nonlinear mapping features, and aggregate to establish a regression mapping matrix; S403: Based on the regression mapping matrix, read the preset mapping coefficient column and intercept, perform the inner product operation of the regression mapping matrix and the preset mapping coefficient column to obtain the linear combination value, and add it with the intercept to obtain the preliminary trend value. Perform smoothing filtering on the preliminary trend value to extract trend nodes and generate a risk trend inference vector.

[0013] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Read the numerical sequence inside the risk trend inference vector, extract the state inference value of the first node representing the risk level, retrieve the time offset parameter of the last node representing the time span, aggregate the state inference value and the time offset parameter, fuse the node position index, and construct the parsing parameter pair. S502: Call the parsing parameter pair, extract the state inference value and time offset parameter, calculate the state inference value minus the preset light anesthesia judgment threshold to obtain the difference parameter, filter the difference parameter as the positive value corresponding to the inference value, add the current time and time offset parameter to obtain the cutoff time node, integrate the current time and the cutoff time node, and establish a high-risk monitoring window. S503: Based on the high-risk monitoring window, collect the vital signs data sequence within the window, perform a matching operation between the vital signs data sequence and the awakening feature image library to obtain the waveform overlap rate, extract the waveform segments corresponding to waveform overlap rates greater than the reference ratio, encapsulate the waveform segments into control code, configure the emergency flag bit inside the control code, and generate an anesthesia awakening early warning command.

[0014] As a further aspect of the present invention, the preset light anesthesia determination threshold is the median of the range of risk level nodes in the numerical sequence within the risk trend inference vector.

[0015] A lightweight anesthesia recovery early warning system based on the temporal characteristics of heart-brain linkage includes: The signal waveform analysis module collects electrocardiogram (ECG) waveform data from patients in the anesthesia recovery period, segments the ECG waveform data according to a time window, detects the peak timestamps of the ECG waveform data within the time window, performs a difference operation on adjacent timestamps, and generates a beat interval sequence. The polarity feature extraction module performs phase space reconstruction operation on the jump interval sequence, calculates the probability log ratio of neighboring point pairs in the multidimensional state space to extract sample entropy values, counts the number of times adjacent sign bits of the sample entropy values ​​switch between positive and negative, and generates polarity flip density features. The autocorrelation evaluation module performs decentralization processing on the jump interval sequence and extracts the third-order cumulant that characterizes the asymmetric distribution of the waveform. It evaluates the autocorrelation coefficient of the third-order cumulant under continuous delay steps and performs attenuation amplitude convergence processing on the autocorrelation coefficients under adjacent delay steps to generate autocorrelation attenuation features. The risk trend inference module concatenates the polarity reversal density feature and the autocorrelation decay feature to construct a multi-dimensional joint feature space. The multi-dimensional joint feature space is then input into the support vector regression model to perform kernel function mapping operations. Bias compensation is performed by combining preset mapping coefficients to generate a risk trend inference vector. The awakening status early warning module parses the risk trend inference vector to obtain the status inference value and time offset parameter. When the status inference value exceeds the preset light anesthesia judgment threshold, it determines the high-risk monitoring window based on the current time and time offset parameter, and generates an anesthesia awakening early warning command within the high-risk monitoring window.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the peak difference of electrocardiogram is extracted to construct a beat interval sequence, and the polarity reversal density is derived by calculating the probability logarithmic ratio in the reconstructed phase space. This allows for in-depth exploration of the nonlinear evolution characteristics within the microscopic temporal sequence of physiological signals to capture early weak variations. Decentralization is used to extract third-order cumulative quantities and calculate the attenuation amplitude corresponding to continuous delay step size. Redundant interference dimensions are significantly eliminated and the computational load is reduced to ensure lightweight and high-speed response. Multidimensional joint features are mapped by kernel function and bias compensation to obtain state inference values ​​and time shift parameters. By dynamically delineating high-risk monitoring windows and issuing early warning commands, the invention accurately adapts to individual metabolic differences, fundamentally eliminating the lag and false alarm defects caused by fixed thresholds and comprehensively improving perioperative safety. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0021] Please see Figure 1 This invention provides a lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage, comprising the following steps: S1: Collect ECG waveform data of patients in the anesthesia recovery period, segment the ECG waveform data according to the time window, detect the peak timestamp of the ECG waveform data within the time window, and perform a difference operation on adjacent timestamps to generate a beat interval sequence; S2: Perform phase space reconstruction operation on the jump interval sequence, calculate the probability log ratio of neighboring point pairs in the multidimensional state space to extract sample entropy values, count the number of times adjacent sign bits of sample entropy values ​​switch between positive and negative, and generate polarity reversal density features. S3: Perform decentralization processing on the jump interval sequence and extract the third-order cumulant that represents the asymmetric distribution of the waveform. Evaluate the autocorrelation coefficient of the third-order cumulant under continuous delay steps. Perform attenuation amplitude convergence processing on the autocorrelation coefficient under adjacent delay steps to generate autocorrelation attenuation features. S4: Concatenate the polarity reversal density feature and the autocorrelation decay feature to construct a multi-dimensional joint feature space. Input the multi-dimensional joint feature space into the support vector regression model to perform kernel function mapping operation. Combine the preset mapping coefficients to perform bias compensation and generate a risk trend inference vector. S5: Analyze the risk trend inference vector to obtain the state inference value and time offset parameter. When the state inference value exceeds the preset light anesthesia judgment threshold, determine the high-risk monitoring window based on the current time and time offset parameter, and generate an anesthesia awakening warning instruction within the high-risk monitoring window.

[0022] The beat interval sequence includes the heart rate interval, heart rate variability, and periodic volatility; the polarity reversal density features include the sign change rate, sample entropy value, and complexity index; the autocorrelation decay features include the autocorrelation decay coefficient, autocorrelation function value, and time delay correlation coefficient; the risk trend inference vector includes the risk score, trend slope, and deviation correction value; and the anesthesia awakening warning instructions include the warning level identifier, risk duration, and monitoring response category.

[0023] Please see Figure 2 The specific steps of S1 are as follows: S101: Collect continuous electrocardiogram (ECG) data from the patient's body surface, retrieve the potential amplitude and corresponding time sequence markers within the continuous ECG data, perform a truncation operation on the continuous ECG data with a fixed step size using window scale limiting parameters, aggregate all sampled feature points falling within the same time span, and obtain a segmented waveform feature matrix; The system receives continuous ECG data streams transmitted in real-time from the patient's body surface via a chest patch-type silver chloride electrode sensor. The acquisition frequency is uniformly configured at 500 Hz, and the acquired data format is a two-dimensional floating-point sequence containing timestamps and potential amplitudes. For the input continuous ECG data, after pre-removing invalid data frames with empty potential values, a sequence traversal operation is performed to extract the potential amplitude value sequence, and the timestamps corresponding to each amplitude are extracted simultaneously as time sequence markers. Preset window size parameters are invoked, specifically including two custom values: the truncation window width and the sliding step size. Based on the acquired truncation window width of 2000 milliseconds and sliding step size of 500 milliseconds, the ECG data is truncated at a fixed step size. The initial window is set with a start of 0 milliseconds and an end of 2000 milliseconds, which is set as the first time span.

[0024] Table 1: ECG Window Capture Parameter Configuration Table Monitoring start time 0 milliseconds 500 milliseconds 1000 milliseconds Monitoring termination time 2000 milliseconds 2500 milliseconds 3000 milliseconds Table 1 visually illustrates the continuous window definition relationship. By comparing all sample point markers within the data, sampling feature points falling within the current range are selected. The potential amplitudes of these feature points are then vertically aggregated in chronological order to generate a local waveform data column. After completion, the start and end points are incremented by 500 milliseconds to enter the next range for aggregation, repeating this process until the total duration is covered. Each extracted waveform data column is used as a row vector and concatenated level by level to obtain a segmented waveform feature matrix. When processing data with a total duration of 5000 milliseconds, the first extraction captures 1000 points within 0 to 2000 milliseconds to form the first row. Sliding by 500 milliseconds, the second extraction captures 1000 points within 500 to 2500 milliseconds to form the second row. This process is repeated seven times to generate a segmented waveform feature matrix containing 7 rows and 1000 columns.

[0025] S102: Extract the extreme values ​​of potential amplitude of each row in the segmented waveform feature matrix, compare the extreme values ​​with the preset peak judgment benchmark threshold, retain the corresponding items whose extreme values ​​are greater than the preset peak judgment benchmark threshold and retrieve the corresponding time stamp, arrange the corresponding time stamp according to the incrementing rule, and obtain the peak point timestamp sequence. Read the segmented waveform feature matrix output from the previous stage, and perform a first-order difference operation on the potential amplitude of each row in the matrix. For each row of the matrix, calculate the difference between the potential amplitudes of two adjacent sampling points, locate the node where the difference value changes from positive to negative, determine it as an extreme value within the local waveform, and extract the corresponding maximum potential amplitude. For all extracted maxima, retrieve a preset peak determination benchmark threshold, which is 0.5 mV, based on objective statistical values ​​of the lower limit of the range of jump potential amplitudes in resting ECG waveforms of healthy adults. Compare each extracted maxima with the preset peak determination benchmark threshold. Items with extreme values ​​strictly greater than 0.5 mV are retained; items with extreme values ​​less than or equal to 0.5 mV are directly removed from the current row of data. For the retained maximum items, based on their original row and column index information within the matrix, retrieve their associated specific time stamps to obtain the specific timestamp of the occurrence of the maximum value. After obtaining the timestamp data, all retained timestamps are rearranged according to an ascending order of their occurrence, generating a peak timestamp sequence composed purely of time values. For example, when processing the first row of 1000 sample points, differential positioning identifies three extreme values ​​with amplitudes of 0.3 mV, 1.2 mV, and 0.4 mV. These are compared to a baseline threshold of 0.5 mV; the 0.3 mV and 0.4 mV values, being less than the threshold, are discarded, retaining the 1.2 mV node. A reverse lookup reveals that the corresponding time for this 1.2 mV is 850 milliseconds. After a full matrix comparison, five valid timestamps are collected: 3350 milliseconds, 850 milliseconds, 4200 milliseconds, 1650 milliseconds, and 2400 milliseconds. In ascending order, the peak timestamp sequence is obtained as 850 milliseconds, 1650 milliseconds, 2400 milliseconds, 3350 milliseconds, and 4200 milliseconds.

[0026] S103: Extract the record values ​​of adjacent nodes in the timestamp sequence of the peak point, subtract the timestamp value of the preceding node from the timestamp value of the subsequent node, perform the difference calculation, and write all the time difference variables output by the difference calculation into the chain buffer space one by one according to the node arrangement order to generate the jump interval sequence. The system retrieves the peak timestamp sequence after ascending sorting and performs subtraction operations between adjacent node values ​​according to the sequence node order. Starting from the first node, it searches for its adjacent items in the sequence one by one. The second timestamp is extracted as the subsequent node, and the first timestamp as the preceding node. The difference between the timestamp values ​​of the subsequent and preceding nodes is calculated to obtain the time span difference, generating the first time difference variable. Subsequently, the node cursor moves one position forward, using the third timestamp as the subsequent node and the second timestamp as the preceding node, and performs the difference calculation again to obtain the second time difference variable. This process is repeated for the entire peak timestamp sequence, performing the difference calculation for all adjacent subsequent and preceding nodes, and outputting all corresponding time difference variables. For storing the time difference variables, a dynamically expandable unidirectional linked cache space is pre-established in the system memory. The time difference variables calculated from all the subtraction operations are written into the linked cache space one by one, according to the original node arrangement order at the time of calculation, to generate the jump interval sequence. For example, the previous peak timestamp sequence is retrieved, containing 5 nodes: 850 milliseconds, 1650 milliseconds, 2400 milliseconds, 3350 milliseconds, and 4200 milliseconds. For the first pair of adjacent nodes, the 1650 milliseconds of the subsequent node is subtracted from the 850 milliseconds of the preceding node, resulting in a time difference variable of 800 milliseconds. The cursor is then moved to process the second pair of nodes, subtracting the 1650 milliseconds of the preceding node from the 2400 milliseconds of the subsequent node, resulting in a time difference variable of 750 milliseconds. For the third pair of nodes, 3350 milliseconds is subtracted from 2400 milliseconds to obtain 950 milliseconds. For the fourth pair of nodes, 4200 milliseconds is subtracted from 3350 milliseconds to obtain 850 milliseconds. After the subtraction operation between adjacent nodes, a total of 4 time difference variables are generated. Store 800 milliseconds, 750 milliseconds, 950 milliseconds, and 850 milliseconds in a linked cache to construct a jump interval sequence.

[0027] Please see Figure 3 The specific steps of S2 are as follows: S201: Call the jump interval sequence, obtain the embedding dimension parameter and delay time parameter, perform truncation and shift operation on the jump interval sequence, construct the reconstructed phase space matrix, calculate the Chebyshev distance between row vectors in the reconstructed phase space matrix, aggregate the vector pairs corresponding to Chebyshev distances less than the preset tolerance threshold, divide the number of vector pairs by the total number of pairings, and obtain the probability set of neighboring points. Read the above jump interval sequence and obtain the embedding dimension parameter value of 2 and the delay time parameter value of 1. Based on these two parameters, perform a truncation and shift operation on the jump interval sequence, and transpose the one-dimensional sequence into a multi-dimensional data group through a sliding window to combine them into a reconstructed phase space matrix. Calculate the Chebyshev distance between each row vector in the reconstructed phase space matrix. Specifically, calculate the absolute value of the coordinate difference between any two row vectors in their corresponding dimensions, and extract the largest value as the Chebyshev distance between the two vectors. Traverse all non-repeating row vector pairs in the reconstructed phase space matrix to obtain all Chebyshev distance values. Sort all Chebyshev distance values ​​in ascending order from smallest to largest. Determine a preset tolerance threshold based on the sorting result, and select the data value corresponding to the position set at 3 / 10 of the total number of Chebyshev distances, which is within the specified range of 1 / 10 to 1 / 2. After extracting the preset tolerance threshold, all previously calculated Chebyshev distances are iterated again. Row vector pairs with Chebyshev distances less than the preset tolerance threshold are aggregated, and the total number of vector pairs meeting the constraints is counted. The number of obtained vector pairs is divided by the total number of possible pairings within the matrix to obtain the nearest neighbor probability set. For example, with a jump interval sequence of 800, 750, 950, and 850, when the dimension is 2 and the delay is 1, the generated reconstruction matrix contains three row vectors: the first row contains 800 and 750, the second row contains 750 and 950, and the third row contains 950 and 850. The absolute values ​​of the differences between the first and second rows are calculated as 50 and 200, respectively, with the maximum value of 200. The maximum difference between the first and third rows is 150. The maximum difference between the second and third rows is 200. The distance set is 200, 150, and 200, arranged in ascending order as 150, 200, and 200. The first value, approximately 150, at 3 / 10 of the three distance values ​​is selected as the preset tolerance threshold. There is one vector pair with a distance less than or equal to 150, resulting in a total of three possible pairings. Dividing 1 by 3 yields a neighbor probability set of 0.33.

[0028] S202: Based on the step size parameter, embed the dimension parameter, re-execute vector recombination and distance comparison to obtain the expanded dimension probability set, determine the ratio of the values ​​in the neighboring point probability set to the values ​​in the expanded dimension probability set, perform natural logarithm operation and extract the absolute value, and establish the sample entropy state sequence. Extract the current embedding dimension parameter value and perform an incremental operation based on the set fixed step size parameter. Set the step size parameter value to 1, add the step size parameter to the original embedding dimension parameter, and obtain the expanded embedding dimension after dimensionality increase. Apply the expanded embedding dimension and the unchanged delay time parameter, and re-perform vector recombination operation on the original jump interval sequence to generate the expanded high-dimensional reconstructed phase space matrix. For each high-dimensional row vector in this high-dimensional reconstructed phase space matrix, calculate the Chebyshev distance between all non-repeating high-dimensional row vector pairs according to the principle of maximizing the absolute value of the coordinate difference of each dimension. Use the preset tolerance threshold determined in the previous stage and kept unchanged to perform an alignment operation with all newly calculated high-dimensional Chebyshev distances. Summarize the total number of corresponding vector pairs whose high-dimensional Chebyshev distance is less than the preset tolerance threshold, and divide the number of expanded vector pairs by the total number of pairs recalculated in the high-dimensional state to obtain the expanded dimension probability set. The previously calculated nearest-neighbor probability set is used as the numerator, and the expanded-dimensional probability set is used as the denominator to determine the proportional relationship between the two and obtain the probability ratio. The natural logarithm is performed on this probability ratio, and the absolute value of the output is extracted, the negative sign is removed, and this absolute value is used to establish the sample entropy state sequence. For example, the original jump interval sequence is 800, 750, 950, 850, the preset tolerance threshold is 150, and the nearest-neighbor probability set is 0.33. When the step size parameter is 1, the embedding dimension increases to 3. When the dimension is 3, only two row vectors are generated by recombination: the first row is 800, 750, 950, and the second row is 750, 950, 850. Chebyshev distance is calculated, and the absolute values ​​of the differences in each dimension are 50, 200, and 100, with a maximum value of 200. Since 200 is not less than the threshold of 150, the condition that the number of vector pairs is 0 is satisfied. To prevent the denominator from being zero, a bias constant of 0.01 is added to the number of expanded-dimensional vector pairs. The total number of high-dimensional pairings is 1, and the calculated probability set for the expanded dimension is 0.01. Dividing the probability set of neighboring points (0.33) by the probability set of the expanded dimension (0.01) yields a probability ratio of 33. Performing a natural logarithm operation on 33 and extracting the absolute value gives a value of 3.49. The final output is a sample entropy state sequence that includes 3.49.

[0029] S203: Based on the sample entropy state sequence, extract the difference between the node value and the overall mean, establish a positive and negative sign label column and retrieve adjacent items in the time series, compare the polarity state, accumulate the number of reverse switching to obtain the cumulative number of flips, and calculate the ratio of the cumulative number of flips to the total number of node records to generate the polarity flip density feature. Read the generated sample entropy state sequence, count the values ​​of all nodes contained in the sequence, and calculate the arithmetic mean of all node values ​​as the overall mean. Perform a traversal operation on the sample entropy state sequence, extracting the value of each individual node. Subtract the overall mean from the extracted node value to obtain the difference parameter corresponding to each node. Determine the positive or negative attribute of all difference parameters. If the difference parameter is greater than zero, record it as positive at its corresponding position; if the difference parameter is less than or equal to zero, record it as negative. Establish a positive / negative sign marker column containing only positive and negative signs. Retrieve two adjacent items in the positive / negative sign marker column arranged chronologically and perform a polarity comparison operation. Determine if the adjacent preceding and following markers are consistent. If a sign changes from positive to negative or vice versa, a reverse switch is considered to have occurred. Complete the retrieval of the entire sign marker column and accumulate the number of reverse switches to obtain the cumulative number of flips. The total number of node records in the positive / negative sign marker column is counted. The ratio of the previously obtained cumulative flip count to the total number of node records is calculated to generate a polarity flip density feature. For example, a sample entropy state sequence containing 5 nodes is received, with node values ​​of 3.49, 1.25, 4.62, 2.18, and 5.11 respectively. The sum of these 5 values ​​is 16.65, which, divided by the total number of 5, yields an overall mean of 3.33. Subtracting the overall mean of 3.33 from each node value yields differences of 0.16, -2.08, 1.29, -1.15, and 1.78 respectively. The positive / negative attribute of each difference is determined, and a positive / negative sign marker column is established, with the result being positive, negative, positive, negative, and positive. Comparing adjacent items, the first switch occurs when the sign of the first item is positive and the sign of the second item is negative; the second switch occurs when the sign of the second item is negative and the sign of the third item is positive; the third switch occurs when the sign of the third item is positive and the sign of the fourth item is negative; and the fourth switch occurs when the sign of the fourth item is negative and the sign of the fifth item is positive. The total number of flips is 4. The total number of records in this sequence is 5. Dividing the number of flips (4) by the total number of records (5) yields a final polarity flip density feature of 0.80.

[0030] Please see Figure 4 The specific steps of S3 are as follows: S301: For the jump interval sequence, calculate the expected parameter, extract the difference between the node value and the expected parameter, map the node difference state to construct a centered deviation matrix, perform a cube operation on the elements in the centered deviation matrix to obtain the sum parameter, calculate the skewness parameter based on the ratio of the sum parameter to the total number of nodes, and obtain the third-order cumulant parameter. The beat interval sequence established based on the initial ECG data is re-established. The total number of numerical terms within the sequence is counted. The values ​​of all nodes within the sequence are summed, and this sum is divided by the total number of numerical terms to calculate the arithmetic mean of the entire sequence, which is used as the expected parameter. Each individual node value within the beat interval sequence is extracted sequentially, and the expected parameter is subtracted from the node value to obtain the deviation difference for each node. All nodes are mapped one-to-one to their corresponding deviation difference states according to their chronological order, thus constructing a one-dimensional centered deviation matrix. A cubic operation is performed on each numerical element within this centered deviation matrix, that is, each deviation difference is multiplied three times consecutively. All the new values ​​generated after the cubic operation are summed to obtain a global sum parameter. Based on the total number of nodes obtained previously, the sum parameter is divided by the total number of nodes to obtain a ratio value, which is output as a skewness parameter, resulting in a third-order cumulant parameter reflecting the asymmetric characteristics of the data distribution. For example, the four basic node values ​​extracted from the jump interval sequence are 800, 750, 950, and 850. Summing these four node values ​​yields a total of 3350. Dividing 3350 by the total number of nodes (4) calculates the expected parameter as 837.5. Performing subtraction operations, subtracting 837.5 from 800 gives -37.5; subtracting 837.5 from 750 gives -87.5; subtracting 837.5 from 950 gives 112.5; and subtracting 837.5 from 850 gives 12.5. This constructs a centered deviation matrix containing -37.5, -87.5, 112.5, and 12.5. The differences were raised to the power of each value: -37.5 to the power of its cube equals -52734.375, -87.5 to the power of its cube equals -669921.875, 112.5 to the power of its cube equals 1423828.125, and 12.5 to the power of its cube equals 1953.125. These cubed results were then summed to obtain a total parameter of 703125. Dividing this total parameter by the total number of nodes (4) yielded a skewness parameter of 175781.25, which was ultimately determined as the third-order cumulant parameter.

[0031] S302: Call the third-order cumulant parameter to obtain the step size set, extract the mapping feature column by shifting the third-order cumulant parameter based on the numerical values ​​in the step size set, calculate the sum of the product of the third-order cumulant parameter and the corresponding nodes of the mapping feature column, perform standardization based on the ratio of the sum of the product to the node variance, and establish the autocorrelation coefficient sequence. The process involves using the acquired third-order cumulant parameters and the original jump interval sequence node bias data to define a step size set consisting of a series of positive integers, used to define the scale range of the lag time period. A single value within the step size set is selected as the lag step size. Based on this lag step size, the original centered bias sequence is shifted, sliding all data points backward by a specified number of positions, truncating any nodes exceeding the specified range, thereby extracting the corresponding mapping feature column. For the retained original centered bias sequence and the newly generated mapping feature column, nodes at corresponding positions are extracted one by one, and the corresponding node values ​​are multiplied. The product results from all positions are then summed to obtain the product. The sum of squares of the nodes in the original centered bias sequence is calculated and divided by the total number of nodes to obtain the node variance. The calculated product sum is then divided by the product of the node variance and the total number of effective nodes, performing standardization to eliminate dimensional differences. This process is repeated iteratively, substituting all values ​​within the step size set into the above steps to generate correlation data under different lag step size conditions. Finally, an autocorrelation coefficient sequence is established, arranged in ascending order of step size. For example, retrieve the centered deviation matrix generated in the previous stage, which contains four deviation nodes: -37.5, -87.5, 112.5, and 12.5. Calculate the squares of each node as 1406.25, 7656.25, 12656.25, and 156.25, respectively, and sum them to obtain a sum of squares of 21875. Divide the sum of squares by 4 to obtain the variance of 5468.75. Set one step size value within the step size set to 1. Shift the original sequence one position backward. The number of overlapping nodes between the original sequence and the shifted sequence is 3. Extract the last three terms of the original sequence, namely -87.5, 112.5, and 12.5, and multiply them by the first three terms of the mapped feature column, namely -37.5, -87.5, and 112.5. Multiplying -87.5 by -37.5 gives 3281.25, multiplying 112.5 by -87.5 gives -9843.75, and multiplying 12.5 by 112.5 gives 1406.25. Accumulating these multiplications yields a product sum of -5156.25. Dividing this product sum by the product of the variance 5468.75 and the number of effective nodes (3) yields a final standardized value of -0.31. After iterative calculations across multiple step sizes, the standardized value, including -0.31, is output, constructing an autocorrelation coefficient sequence.

[0032] S303: Retrieve adjacent nodes in the autocorrelation coefficient sequence arranged in ascending order of time, perform numerical subtraction of adjacent nodes to obtain discrete difference, extract the absolute value of discrete difference to obtain decay amplitude, summarize the decay amplitude corresponding to step size to construct decay feature library and solve the mean in decay feature library to generate autocorrelation decay feature. The generated autocorrelation coefficient sequence is read, and adjacent nodes within the sequence, arranged in strict ascending order of lag time steps, are traversed and compared. The starting node position in the autocorrelation coefficient sequence is located, and the current node value is extracted as the minuend. The value of the next adjacent node is extracted as the subtrahend, and subtraction is performed between the adjacent node values ​​to obtain the absolute fluctuation of the difference, i.e., the discrete difference. To quantify the decrease in correlation caused by fluctuation, the absolute value of the obtained discrete difference is directly extracted to eliminate positive and negative directional differences, obtaining the attenuation amplitude at the corresponding step size switch. All adjacent node pairs within the autocorrelation coefficient sequence are traversed, and the difference operation of this absolute value is repeatedly performed. All lag steps are mapped to the corresponding attenuation amplitude calculated each time, and all are loaded into the designated storage space to construct a complete attenuation feature library. The total number of attenuation amplitude data in the attenuation feature library is counted, and all attenuation amplitude values ​​in the feature library are summed. The accumulated sum is divided by the total number of data points to calculate the mean. The final output average value is the autocorrelation attenuation feature. For example, after the preceding steps, an autocorrelation coefficient sequence containing four correlation values ​​is obtained, arranged in ascending order of step size: 1.00, -0.31, 0.15, and -0.10. For the first pair of adjacent nodes, the first node (1.00) and the second node (-0.31) are extracted. A subtraction operation is performed: 1.00 minus -0.31 yields 1.31. The absolute value of 1.31 is extracted, representing the attenuation magnitude. For the second pair of adjacent nodes, -0.31 is extracted, minus 0.15 yields -0.46. The absolute value of 0.46 is extracted, representing the attenuation magnitude. For the third pair of adjacent nodes, 0.15 is extracted, minus -0.10 yields 0.25. The absolute value of 0.25 is extracted, representing the attenuation magnitude. The values ​​of 1.31, 0.46, and 0.25 are then combined to construct the attenuation feature library. The database contains a total of 3 data points. Summing 1.31, 0.46, and 0.25 yields a total value of 2.02. Dividing this total value of 2.02 by the total number of data points (3) gives a mean of 0.67. This value of 0.67 represents the autocorrelation decay characteristic.

[0033] Please see Figure 5 The specific steps of S4 are as follows: S401: For polarity reversal density features and autocorrelation decay features, extract the dimension and record length, align the nodes within the dimension according to the record length, establish a multi-dimensional spliced ​​data architecture, map the features to adjacent channels of the multi-dimensional spliced ​​data architecture in order, and integrate node attributes to establish a multi-dimensional joint feature space. The polarity inversion density feature and autocorrelation decay feature datasets generated in the system cache are read separately. For these two types of independent feature data, the column length within their data matrices is scanned and extracted as the dimension parameter, while the total number of sampling points in the dataset is counted as the record length parameter. The record length parameters corresponding to the two types of features are compared, and the smaller value is used as the truncation benchmark. All feature nodes within their respective dimensions are aligned to ensure complete consistency in temporal span and total number of sample points. A multidimensional spliced ​​data architecture is constructed, initially represented as an empty tensor array with two parallel input channels. According to a fixed loading order, the polarity inversion density feature numerical sequence after record length alignment is mapped to the first parallel input channel of the multidimensional spliced ​​data architecture as the first column data; the aligned autocorrelation decay feature numerical sequence is mapped to the second parallel channel as the second column data. During the loading process, nodes at the same row index position are extracted horizontally, and the attribute labels of each independent node are merged to establish a multidimensional joint feature space containing composite dimensional information. For example, the polarity reversal density feature sequence generated earlier is extracted, with a record length of 1000 nodes and a single-dimensional dimension for each node. The autocorrelation decay feature sequence is extracted, with a record length of 1050 nodes and a single-dimensional dimension. After comparison and judgment, the 50 nodes exceeding the tail of the autocorrelation decay feature are pruned, and the record lengths of both are strictly aligned to 1000 nodes. A 1000-row, 2-column two-dimensional matrix is ​​constructed as the multi-dimensional spliced ​​data architecture. The polarity reversal density feature representative node 0.80 obtained in the previous example is mapped and filled into the channel position of the 1st row and 1st column; the autocorrelation decay feature representative node 0.67 is mapped and filled into the channel position of the 2nd column in the same row. 0.80 and 0.67 in the same row are bound together by position index. Synchronous traversal and assignment operations are performed on the remaining 999 nodes. After filling the channel units, the 1000-row, 2-column matrix constitutes a complete multi-dimensional joint feature space.

[0034] S402: Call the multi-dimensional joint feature space, obtain the preset support vector set and kernel function scale parameter, calculate the Euclidean distance between the node in the feature space and each vector in the support vector set, construct the proportional relationship between the squared distance value and the kernel function scale parameter to obtain the intermediate quotient value, perform the natural exponent negative operation to extract nonlinear mapping features, and aggregate to establish a regression mapping matrix; The system invokes the constructed multidimensional joint feature space and extracts a preset support vector set trained from historical real anesthesia monitoring data samples from the built-in database. Simultaneously, it retrieves the scaling parameters used to control the influence range of the Gaussian radial basis function kernel. The system extracts the test sample node vector row by row from the multidimensional joint feature space and calculates the Euclidean distance between the test sample node vector and each independent support vector in the preset support vector set. The system calculates the difference between the corresponding values ​​of each dimension of the test sample vector and the support vectors, squares the differences, and sums them to obtain a squared sum. Due to the canceling property of subsequent kernel function operations, this sum of squares is directly retained as the squared distance value. The kernel function scaling parameter is extracted, and the squared distance value is divided by twice the squared kernel function scaling parameter value to establish a proportional relationship and obtain a median quotient. This median quotient is then negativeened, and the negative median quotient is used as the exponent for natural exponential operations, extracting nonlinear mapping feature values. The nonlinear mapping feature values ​​obtained from all support vectors are sequentially aggregated to establish a complete regression mapping matrix. For example, the sample node vector extracted from the multidimensional joint feature space contains two dimensions with feature values ​​of 0.80 and 0.67. The first support vector extracted from the preset support vector set contains feature values ​​of 0.50 and 0.40. The kernel function scaling parameter is set to 1.5. First, the differences between each dimension are calculated: 0.80 minus 0.50 equals 0.30, and 0.67 minus 0.40 equals 0.27. Squaring these values ​​yields 0.09 and 0.0729 respectively, summing them to 0.1629. This value is directly used as the squared distance value. The square of the kernel function scaling parameter 1.5 is 2.25, and twice that is 4.5. Dividing the squared distance value 0.1629 by 4.5 yields a quotient of 0.0362. Negating this quotient yields -0.0362. Using the natural constant 2.718 as the base and -0.0362 as the exponent, a natural exponentiation operation is performed, yielding an exponent of approximately 0.9644. This value represents the nonlinear mapping characteristic between the current sample and the first support vector. This exponentiation operation is repeated for all support vectors, and all exponent results are concatenated and arranged to construct a regression mapping matrix.

[0035] S403: Based on the regression mapping matrix, read the preset mapping coefficient column and intercept, perform the inner product operation of the regression mapping matrix and the preset mapping coefficient column to obtain the linear combination value, and add it to the intercept to obtain the preliminary trend value. Perform smoothing filtering on the preliminary trend value to extract trend nodes and generate a risk trend inference vector. Based on the generated regression mapping matrix, the preset mapping coefficient column and bias intercept value saved in the historical training configuration file are read. The nonlinear mapping feature values ​​corresponding to each support vector in the regression mapping matrix are retrieved, and this matrix is ​​paired with the data items at the same index position in the preset mapping coefficient column. For all paired data combinations, the corresponding position values ​​are multiplied, and the results of each multiplication are globally accumulated and summed to complete the inner product operation, obtaining the linear combination value representing the current sample's main predicted state within the model. After obtaining the linear combination value, it is directly added to the retrieved bias intercept value to obtain the preliminary trend value after compensating for the baseline bias. To eliminate abnormal fluctuations caused by short-term noise during numerical calculation, a moving average window with a length of 5 nodes is applied to perform a smoothing filter operation of summing and averaging the current and five consecutive preliminary trend values ​​of the preceding and following periods. Finally, the average value of the filtered output is extracted as a stable trend node, and all trend nodes in the continuous time series are sequentially concatenated to generate a risk trend inference vector representing the changes in the patient's awakening process. For example, based on the previously obtained regression mapping matrix, which contains three nonlinear mapping feature values ​​of 0.9644, 0.8520, and 0.7230, the training configuration is read to obtain the preset mapping coefficient column, which contains three weight values ​​of 2.5, -1.2, and 1.8. The bias intercept value is also obtained as 0.5. Pairwise multiplication is performed: 0.9644 multiplied by 2.5 yields 2.411, 0.8520 multiplied by -1.2 yields -1.0224, and 0.7230 multiplied by 1.8 yields 1.3014. The product results 2.411, -1.0224, and 1.3014 are then summed to obtain a linear combination value of 2.69. This linear combination value of 2.69 is then added to the intercept value of 0.5 to obtain a preliminary trend value of 3.19. The two preceding and two following historical preliminary trend values ​​are retrieved: 3.10 and 3.15, and 3.22 and 3.29, respectively. These, along with the current value of 3.19, form a window containing five data points. The sum is 15.95, which, divided by the window length of 5, yields the smoothed trend node of 3.19. As time progresses, these smoothed trend nodes are stored and combined in chronological order to generate a risk trend inference vector.

[0036] Please see Figure 6 The specific steps of S5 are as follows: S501: Read the numerical sequence inside the risk trend inference vector, extract the state inference value of the first node representing the risk level, retrieve the time offset parameter of the last node representing the time span, aggregate the state inference value and the time offset parameter, fuse the node position index, and construct the parsing parameter pair. After completing the trend prediction, the newly generated risk trend inference vector in the storage space is read. The numerical sequence generated by smoothing filtering within this vector is traversed from beginning to end. Since the sequence data structure is defined as a special dictionary object with composite fields, the first attribute field is extracted. This field records the numerical value representing the severity of the awakening state calculated by the system within the current prediction period. This field is extracted and used as the state inference value. The attribute field at the end of the sequence is retrieved. This field records the expected duration (in milliseconds) from the current time point to the occurrence of the awakening event. This field is extracted and used as the time offset parameter. After extracting these two core fields, the state inference value representing the severity and the time offset parameter representing the duration are correlated and aggregated. During aggregation, the position index number corresponding to the current inference object on the global event timeline is extracted, and the node position index is forcibly merged into the aggregated data packet as a location tag. By encapsulating the state inference value, time offset parameter, and position identifier into a single data structure, a parsing parameter pair for subsequent window judgment is generated. For example, a single frame of risk trend inference vector data is read. Using a key-value pair retrieval method, the specific value 3.19 stored in the first key-value field of the vector is extracted. This value reflects the probability level of the patient's body being in a state of imminent awakening. 3.19 is copied and established as the state inference value. The last key-value field of the vector is scanned, and the recorded duration span value of 180,000 milliseconds is extracted. This value represents the estimated countdown time before the appearance of obvious signs of awakening, and is established as the time offset parameter. Simultaneously, the node position index number currently recorded by the system is extracted, assuming the current sequence frame is number 5680. The state inference value 3.19, the time offset parameter 180,000, and the position index 5680 are loaded together into a separate memory structure unit. This encapsulation action transforms the fragmented inference results into a compact parsed parameter pair data packet.

[0037] S502: Call the parsing parameter pair, extract the state inference value and time offset parameter, calculate the state inference value and subtract the preset light anesthesia judgment threshold to obtain the difference parameter, filter the difference parameter as the positive value corresponding to the inference value, add the current time and time offset parameter to obtain the cutoff time node, integrate the current time and the cutoff time node, and establish a high-risk monitoring window. The system calls the encapsulated parsing parameters to extract the state inference value and the corresponding time offset parameter from the data packet. A global scan is performed on the numerical sequence within the risk trend inference vector for all recent periods, collecting all historical risk level nodes to form a value range. The value at the exact middle position after arranging the data within this range is defined as the preset light anesthesia judgment threshold. The extracted current state inference value is subtracted from the preset light anesthesia judgment threshold, and a difference operation is performed to obtain the difference parameter. The polarity of the difference parameter is used for judgment and filtering; only when the difference parameter is positive does it indicate that the current state exceeds the light anesthesia benchmark, and the data of the period corresponding to the inference value is retained. Under the condition that the difference parameter is positive, the absolute system timestamp of the current server is retrieved as the current time. This current time is added to the time offset parameter extracted from the parsing parameter pair to obtain the absolute time node predicted to occur of awakening reaction, which is set as the cutoff time node. On the timeline, the current time is used as the starting boundary, and the cutoff time node is used as the ending boundary to integrate the two ends of the time, thereby establishing a high-risk monitoring window. For example, for the preceding parsing parameter pair, the extracted state inference value is 3.19, and the time offset parameter is 180,000 milliseconds (180 seconds). The risk level node set within the past 30 minutes is queried, sorted in ascending order, and the median value is extracted to obtain the preset light anesthesia judgment threshold of 2.50. Subtracting the threshold 2.50 from the state inference value 3.19 yields a difference parameter of +0.69. Since +0.69 is considered positive, the high-risk window construction mechanism is triggered. The current system's absolute timestamp is set to 14:30:00. This current time is added to the time offset parameter of 180,000 milliseconds, resulting in a cutoff time node including the 180-second time span of 14:33:00. The complete 180-second duration from 14:30:00 to 14:33:00 is delineated to establish a high-risk monitoring window.

[0038] S503: Based on the high-risk monitoring window, collect the vital signs data sequence within the window, perform a matching operation between the vital signs data sequence and the awakening feature image library to obtain the waveform overlap rate, extract the waveform segments corresponding to the waveform overlap rate being greater than the reference ratio, encapsulate the waveform segments into control code, configure the emergency flag bit inside the control code, and generate anesthesia awakening warning command. Based on the established high-risk monitoring window, a specific acquisition channel is activated to continuously acquire real-time EEG and respiratory wave data sequences of patients falling within the start and end boundaries of this time window. The system's built-in awakening feature library is retrieved, which stores various historical waveform templates that have verified patients' conscious states. A sliding time window is used to extract the acquired vital sign data sequences, matching them frame-by-frame with the templates in the awakening feature library. Specifically, the waveform diagrams of the real-time vital sign data sequences and the template waveform diagrams are extracted, and the size of the common area overlapping in a two-dimensional coordinate system is calculated. This common area is divided by the total coverage area of ​​the vital sign data sequence itself, and the percentage value is taken as the waveform overlap rate. A reference ratio limit of 75% is set for all waveform overlap rates. Vital sign data sequences with waveform overlap rates strictly greater than the 75% reference ratio are extracted and filtered as valid waveform segments. These valid waveform segments are converted to hexadecimal format according to the communication protocol specifications and encapsulated into the payload field of the instruction packet to form the underlying control code. An emergency flag is configured at the beginning of the control code, and its value is forcibly changed from 0 to 1. The control code with the flag and other check bits are integrated to generate an anesthesia recovery warning command to activate the external blocking mechanism. For example, a 10-second sequence of vital signs data is collected within a high-risk monitoring window. A template from the recovery feature library is selected for matching. Integration yields a total waveform coverage area of ​​500 square units. The common overlap area between the real-time waveform and the template waveform is calculated to be 410 square units. Dividing 410 by 500 yields a waveform overlap rate of 82%. Since 82% is greater than the set reference ratio of 75%, the 10-second waveform segment is extracted and converted into a hexadecimal 16-bit control code array using an encoding library. The emergency flag is changed from the default 0 to 1 at the beginning of the command protocol frame. Finally, a cyclic redundancy check (CRC) code is appended to generate the anesthesia recovery warning command.

[0039] Please see Figure 7 A lightweight anesthesia recovery early warning system based on the temporal characteristics of heart-brain linkage includes: The signal waveform analysis module collects electrocardiogram (ECG) waveform data from patients in the anesthesia recovery period, segments the ECG waveform data according to a time window, detects the peak timestamps of the ECG waveform data within the time window, performs a difference operation on adjacent timestamps, and generates a beat interval sequence. The polarity feature extraction module performs phase space reconstruction operation on the jump interval sequence, calculates the probability log ratio of neighboring point pairs in the multidimensional state space to extract sample entropy values, counts the number of times adjacent sign bits of sample entropy values ​​switch between positive and negative, and generates polarity flip density features. The autocorrelation assessment module performs decentralization processing on the jump interval sequence and extracts the third-order cumulant that characterizes the asymmetric distribution of the waveform. It evaluates the autocorrelation coefficient of the third-order cumulant under continuous delay steps and performs attenuation amplitude convergence processing on the autocorrelation coefficients under adjacent delay steps to generate autocorrelation attenuation features. The risk trend inference module concatenates the polarity reversal density feature and the autocorrelation decay feature to construct a multi-dimensional joint feature space. The multi-dimensional joint feature space is then input into the support vector regression model to perform kernel function mapping operations. Preset mapping coefficients are used for bias compensation to generate a risk trend inference vector. The awakening status early warning module analyzes the risk trend inference vector to obtain the status inference value and time offset parameter. When the status inference value exceeds the preset light anesthesia judgment threshold, it determines the high-risk monitoring window based on the current time and time offset parameter, and generates an anesthesia awakening early warning command within the high-risk monitoring window.

[0040] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage, characterized in that, Includes the following steps: S1: Collect ECG waveform data of patients in the anesthesia recovery period, segment the ECG waveform data according to the time window, detect the peak timestamp of the ECG waveform data within the time window, and perform a difference operation on adjacent timestamps to generate a beat interval sequence; S2: Perform phase space reconstruction operation on the jump interval sequence, calculate the probability log ratio of neighboring point pairs in the multidimensional state space to extract sample entropy values, count the number of times adjacent sign bits of sample entropy values ​​switch between positive and negative, and generate polarity flip density features. S3: Perform decentralization processing on the jump interval sequence and extract the third-order cumulant that represents the asymmetric distribution of the waveform. Evaluate the autocorrelation coefficient of the third-order cumulant under continuous delay steps. Perform attenuation amplitude convergence processing on the autocorrelation coefficient under adjacent delay steps to generate autocorrelation attenuation features. S4: Concatenate the polarity reversal density feature and the autocorrelation decay feature to construct a multi-dimensional joint feature space. Input the multi-dimensional joint feature space into the support vector regression model to perform kernel function mapping operation. Combine the preset mapping coefficients to perform bias compensation and generate a risk trend inference vector. S5: Analyze the risk trend inference vector to obtain the state inference value and time offset parameter. When the state inference value exceeds the preset light anesthesia judgment threshold, determine the high-risk monitoring window based on the current time and time offset parameter, and generate an anesthesia awakening warning instruction within the high-risk monitoring window.

2. The lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage according to claim 1, characterized in that, The beat interval sequence includes heart rate interval, heart rate variability, and periodic volatility; the polarity reversal density feature includes sign change rate, sample entropy value, and complexity index; the autocorrelation decay feature includes autocorrelation decay coefficient, autocorrelation function value, and time delay correlation coefficient; the risk trend inference vector includes risk score, trend slope, and deviation correction value; and the anesthesia awakening early warning instruction includes early warning level identifier, risk duration, and monitoring response category.

3. The lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Collect continuous electrocardiogram (ECG) data from the patient's body surface, retrieve the potential amplitude and corresponding time sequence markers within the continuous ECG data, perform a truncation operation on the continuous ECG data with a fixed step size using window scale limiting parameters, aggregate all sampled feature points falling within the same time span, and obtain a segmented waveform feature matrix; S102: Extract the extreme values ​​of potential amplitude of each row item in the segmented waveform feature matrix, compare the extreme values ​​with the preset peak determination benchmark threshold, retain the items corresponding to extreme values ​​greater than the preset peak determination benchmark threshold and retrieve the corresponding time stamps, arrange the corresponding time stamps according to the incrementing rule, and obtain the peak point timestamp sequence. S103: Extract the record values ​​of adjacent nodes in the timestamp sequence of the peak point, perform difference calculation by subtracting the timestamp value of the preceding node from the timestamp value of the subsequent node, and write the time difference variables output by all difference calculations into the chained cache space one by one according to the node arrangement order to generate the jump interval sequence.

4. The lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Call the jump interval sequence, obtain the embedding dimension parameter and delay time parameter, perform a truncation and shift operation on the jump interval sequence, construct the reconstructed phase space matrix, calculate the Chebyshev distance between row vectors in the reconstructed phase space matrix, aggregate vector pairs whose Chebyshev distance is less than a preset tolerance threshold, divide the number of vector pairs by the total number of pairings, and obtain the probability set of neighboring points. S202: Based on the step size parameter, embed the dimension parameter, re-execute vector recombination and distance comparison to obtain the expanded dimension probability set, determine the ratio of the values ​​in the neighboring point probability set to the values ​​in the expanded dimension probability set to obtain the probability ratio, perform natural logarithm operation and extract the absolute value, and establish the sample entropy state sequence. S203: Based on the sample entropy state sequence, extract the difference between the node value and the overall mean, establish a positive and negative sign label column and retrieve adjacent items in the time series, compare the polarity state, accumulate the reverse switching times to obtain the cumulative flip times, and calculate the ratio of the cumulative flip times to the total number of node records to generate the polarity flip density feature.

5. The lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage according to claim 4, characterized in that, The preset tolerance threshold is determined based on the ascending order of the Chebyshev distances between row vectors in the reconstructed phase space matrix. The preset tolerance threshold is selected between 1 / 10 and 1 / 2 of the total number of Chebyshev distances between row vectors in the reconstructed phase space matrix.

6. The lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage according to claim 4, characterized in that, The specific steps for S3 are as follows: S301: For the jump interval sequence, calculate the expected parameter, extract the difference between the node value and the expected parameter, map the node difference state to construct a centralized deviation matrix, perform a cube operation on the elements in the centralized deviation matrix to obtain the sum parameter, calculate the skewness parameter based on the ratio of the sum parameter to the total number of nodes, and obtain the third-order cumulant parameter. S302: Call the third-order cumulant parameter to obtain the step size set, extract the mapping feature column based on the numerical translation of the third-order cumulant parameter within the step size set, calculate the sum of the product of the third-order cumulant parameter and the corresponding nodes of the mapping feature column, perform standardization processing based on the ratio of the product sum to the node variance, and establish an autocorrelation coefficient sequence. S303: Retrieve adjacent nodes in the autocorrelation coefficient sequence arranged in ascending order of time, perform numerical subtraction operation between adjacent nodes to obtain discrete differences, extract the absolute value of discrete differences to obtain the attenuation amplitude, summarize the attenuation amplitudes corresponding to the step size to construct an attenuation feature library, and solve the mean value in the attenuation feature library to generate autocorrelation attenuation features.

7. The lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage according to claim 6, characterized in that, The specific steps of S4 are as follows: S401: For the polarity reversal density feature and the autocorrelation decay feature, extract the dimension and record length, align the nodes within the dimension according to the record length, establish a multi-dimensional spliced ​​data architecture, map the features to the adjacent channels of the multi-dimensional spliced ​​data architecture in order, and fuse the node attributes to establish a multi-dimensional joint feature space. S402: Call the multidimensional joint feature space, obtain the preset support vector set and kernel function scale parameter, calculate the Euclidean distance between the node in the feature space and each vector in the support vector set, construct the proportional relationship between the squared distance value and the kernel function scale parameter to obtain the intermediate quotient value, perform the natural exponent negative operation to extract nonlinear mapping features, and aggregate to establish a regression mapping matrix; S403: Based on the regression mapping matrix, read the preset mapping coefficient column and intercept, perform the inner product operation of the regression mapping matrix and the preset mapping coefficient column to obtain the linear combination value, and add it with the intercept to obtain the preliminary trend value. Perform smoothing filtering on the preliminary trend value to extract trend nodes and generate a risk trend inference vector.

8. The lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage according to claim 7, characterized in that, The specific steps of S5 are as follows: S501: Read the numerical sequence inside the risk trend inference vector, extract the state inference value of the first node representing the risk level, retrieve the time offset parameter of the last node representing the time span, aggregate the state inference value and the time offset parameter, fuse the node position index, and construct the parsing parameter pair. S502: Call the parsing parameter pair, extract the state inference value and time offset parameter, calculate the state inference value minus the preset light anesthesia judgment threshold to obtain the difference parameter, filter the difference parameter as the positive value corresponding to the inference value, add the current time and time offset parameter to obtain the cutoff time node, integrate the current time and the cutoff time node, and establish a high-risk monitoring window. S503: Based on the high-risk monitoring window, collect the vital signs data sequence within the window, perform a matching operation between the vital signs data sequence and the awakening feature image library to obtain the waveform overlap rate, extract the waveform segments corresponding to waveform overlap rates greater than the reference ratio, encapsulate the waveform segments into control code, configure the emergency flag bit inside the control code, and generate an anesthesia awakening early warning command.

9. The lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage according to claim 8, characterized in that, The preset threshold for determining light anesthesia is the median of the range of risk level nodes in the numerical sequence within the risk trend inference vector.

10. A lightweight anesthesia recovery early warning system based on the temporal characteristics of heart-brain linkage, characterized in that, The system is used to implement the lightweight anesthesia recovery early warning method based on the temporal characteristics of heart-brain linkage as described in any one of claims 1-9, the system comprising: The signal waveform analysis module collects electrocardiogram (ECG) waveform data from patients in the anesthesia recovery period, segments the ECG waveform data according to a time window, detects the peak timestamps of the ECG waveform data within the time window, performs a difference operation on adjacent timestamps, and generates a beat interval sequence. The polarity feature extraction module performs phase space reconstruction operation on the jump interval sequence, calculates the probability log ratio of neighboring point pairs in the multidimensional state space to extract sample entropy values, counts the number of times adjacent sign bits of the sample entropy values ​​switch between positive and negative, and generates polarity flip density features. The autocorrelation evaluation module performs decentralization processing on the jump interval sequence and extracts the third-order cumulant that characterizes the asymmetric distribution of the waveform. It evaluates the autocorrelation coefficient of the third-order cumulant under continuous delay steps and performs attenuation amplitude convergence processing on the autocorrelation coefficients under adjacent delay steps to generate autocorrelation attenuation features. The risk trend inference module concatenates the polarity reversal density feature and the autocorrelation decay feature to construct a multi-dimensional joint feature space. The multi-dimensional joint feature space is then input into the support vector regression model to perform kernel function mapping operations. Bias compensation is performed by combining preset mapping coefficients to generate a risk trend inference vector. The awakening status early warning module parses the risk trend inference vector to obtain the status inference value and time offset parameter. When the status inference value exceeds the preset light anesthesia judgment threshold, it determines the high-risk monitoring window based on the current time and time offset parameter, and generates an anesthesia awakening early warning command within the high-risk monitoring window.

Images