A method and system for health status identification for thoracic surgery patients

Abstract

This invention relates to the field of state recognition technology, specifically a method and system for health status recognition of thoracic surgery patients. The method includes the following steps: acquiring postoperative heart rate, blood pressure, and respiratory rate data; normalizing and constructing a recovery rhythm sequence; using a variational autoencoder to map and calculate the residuals to extract the maximum deviation; calibrating the unstable intervals; determining the synchronous transition critical point; constructing a recovery sequence; identifying abnormal chains; and outputting the state result. In this invention, by constructing the original recovery rhythm sequence and introducing a latent space prediction mechanism, and using differential calculation of residual vectors combined with a fluctuation tolerance threshold to calibrate the unstable intervals, rhythm abnormalities are more accurately judged. By extracting directional consistency and average dwell time indicators from the recovery sub-paths composed of continuous states, and combining this with standard path matching to form a potential abnormal chain identification method, this invention effectively solves the problems of inaccurate identification caused by multi-parameter disconnection, missed transition nodes, and lack of temporal trend in traditional methods.

Description

Technical Field

[0001] This invention relates to the field of status recognition technology, and in particular to a method and system for health status recognition of thoracic surgery patients. Background Technology

[0002] State recognition technology involves perceiving, analyzing, and judging the physiological, behavioral, operational, or system states of objects in specific spatiotemporal environments. Its core aspects include state information acquisition, feature extraction, state classification, and the construction of recognition models. This technology encompasses state discrimination mechanisms based on signals, images, videos, text, or multi-source data. By constructing static or dynamic feature vector spaces, it employs pattern recognition, sequence modeling, probabilistic reasoning, and temporal analysis to achieve accurate classification of target states. State recognition is widely used in scenarios such as healthcare, industrial monitoring, human-computer interaction, and safety early warning. In practical applications, machine learning algorithms are often combined for state modeling and recognition, such as using support vector machines for state boundary delineation or recurrent neural networks to handle temporal dependencies, thereby improving recognition accuracy and generalization ability. In the process of technological evolution, state recognition has gradually shifted from rule-based recognition to data-driven intelligent recognition, emphasizing real-time performance, high-dimensional data processing capabilities, and cross-modal collaborative recognition capabilities.

[0003] Traditional methods for assessing the health status of thoracic surgery patients involve observing and recording changes in vital signs, behavior, and subjective feelings before, during, or after surgery through manual rounds, intermittent instrument monitoring, or subjective patient feedback to determine their current health status. Traditional methods primarily rely on regular measurements of basic parameters such as heart rate, respiratory rate, blood pressure, and body temperature by healthcare professionals. Behavioral status assessment typically depends on observing patient activity, rest status, or self-reported symptoms. In some scenarios, bedside monitoring equipment is used for continuous signal acquisition, but this is usually limited to specific parameters and lacks structured data modeling and cross-time period analysis. These methods are primarily based on healthcare professionals' experience, lacking data-driven reasoning mechanisms and failing to capture dynamic characteristics of status changes in real time.

[0004] Traditional identification methods rely mainly on observation and patient feedback to identify behaviors and subjective states, lacking an effective correspondence with physiological data. Although some continuous data collection is done with the help of monitoring equipment, it is limited to a single parameter and fails to achieve dynamic linkage analysis between multiple indicators. Furthermore, it does not model and process the suddenness and nonlinearity of physiological signal changes, making it impossible to identify potential risk nodes and lacking the ability to characterize the temporal evolution trend of the state. This results in delayed response and insufficient intervention timeliness in actual disease monitoring. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method and system for identifying the health status of thoracic surgery patients.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for identifying the health status of thoracic surgery patients, comprising the following steps: S1: Obtain postoperative heart rate, blood pressure and respiratory rate data of thoracic surgery patients, perform normalization on the data, construct the original recovery rhythm sequence, call the variational autoencoder to map the original recovery rhythm sequence to the latent space, and construct the predicted data sequence. S2: Calculate the residual between the original recovery rhythm sequence and the predicted data sequence, extract the maximum deviation value from the difference residual vector, compare the maximum deviation value with the fluctuation tolerance threshold, and calibrate the rhythm instability interval; S3: Obtain the electrocardiogram (ECG) and respiratory signals in the non-stable interval of the rhythm, locate the abrupt change points of the ECG and respiratory signals, and extract the waveform segments of the abrupt change points; S4: Calculate the conditional mutual information value between the waveform segment at the mutation point and the entire signal segment on the opposite side. When the conditional mutual information value changes, determine the location of the mutation point and identify the critical point for the transition of the cardiopulmonary synchronization state. S5: Based on the critical point of cardiopulmonary synchronous state transition, construct a rehabilitation state sequence and divide it into equally weighted sub-path segments. Calculate the consistency score of the state transition direction and the average dwell time of each stage for the equally weighted sub-path segments. By comparing the standard path matching degree, identify potential abnormal chain segments and output the health status identification result.

[0007] As a further aspect of the present invention, the predicted data sequence includes reconstructed heart rate time series values, predicted blood pressure fluctuation trajectory, and fitted respiratory rate curve; the rhythm instability interval includes abnormal start timestamp, distortion termination timestamp, and non-steady-state confidence marker; the mutation point waveform segment includes local high-frequency amplitude vector, symmetrical time window span, and signal channel identifier; the cardiopulmonary synchronization state transition critical point includes state switching time index, mutual information gradient change rate, and coupling mode classification label; and the health status identification result includes abnormal rehabilitation chain coding, recovery mode deviation matrix, and postoperative risk level assessment.

[0008] As a further aspect of the present invention, the step of obtaining the predicted data sequence specifically includes: S101: Obtain postoperative heart rate, blood pressure, and respiratory rate data of thoracic surgery patients, calculate the arithmetic mean and standard deviation of each data item, perform point-by-point subtraction to remove bias based on the arithmetic mean, perform division normalization on the removed data using the standard deviation to eliminate dimensional differences, and perform temporal alignment and matrix concatenation on the normalized multidimensional data according to a unified timestamp index to construct the original recovery rhythm sequence; S102: For the original recovery rhythm sequence, perform time-series feature encoding, extract nonlinear fluctuation features in the sequence through a multi-layer neural network and map them to a latent variable space of a preset dimension, calculate the mean vector and log-variance vector of the feature distribution in the latent space, and perform Gaussian distribution reparameterization sampling operation based on the mean vector and log-variance vector to obtain the variational sequence embedding vector. S103: The generative decoding network is invoked to perform inverse mapping and upsampling reconstruction operations on the variational sequence embedding vector, restoring the latent space features to the dimensional space of the original time series data. The reconstructed output data is subjected to inverse normalization transformation based on prior statistical features to restore the physical dimension amplitudes of heart rate, blood pressure and respiratory rate. The restored multidimensional data stream is combined according to the time step sequence to generate the predicted data sequence.

[0009] As a further aspect of the present invention, the step of obtaining the rhythmic instability interval specifically includes: S201: Perform point-by-point difference operation on the original recovery rhythm sequence and the predicted data sequence at each time node, calculate the amplitude difference value between the two at the corresponding time step, take the absolute value of the difference value to quantify the single-point reconstruction error, arrange the reconstruction error data of each time node in time sequence and combine them into a vector form to generate a multidimensional amplitude residual sequence. S202: Perform first-order difference operation on the elements of three consecutive time nodes in the multidimensional amplitude residual sequence, calculate the rate of change between adjacent residual values, capture the error fluctuation trend, retrieve the absolute value of the difference operation result within a sliding window of a preset width, filter the difference result with the largest value within the window, and extract the difference extreme value feature. S203: Call the preset fluctuation tolerance threshold, compare the differential extreme value feature with the fluctuation tolerance threshold, detect whether the differential extreme value feature exceeds the fluctuation tolerance threshold, record the start and end timestamps of the current time window when the detection result is exceeded, mark the time domain segment covered by the start and end timestamps as an abnormal state, and mark the rhythm unstable interval.

[0010] As a further aspect of the present invention, the process of setting the fluctuation tolerance threshold specifically involves dividing the rehabilitation cycle into multiple characteristic time periods based on the postoperative recovery time process, obtaining historical steady-state sample data corresponding to the current characteristic time period, calculating the first-order difference statistical features of the reconstruction error of the historical steady-state sample data in the variational space, extracting its distribution mean and standard deviation, and establishing the sum of the mean and the standard deviation of the preset confidence coefficient weight as the fluctuation tolerance threshold for the current time period.

[0011] As a further aspect of the present invention, the step of obtaining the waveform segment at the mutation point specifically includes: S301: Based on the timestamp range of the non-stable interval of the rhythm, extract the electrocardiogram (ECG) signal and respiratory signal of the corresponding time period from the monitoring database, perform first-order discrete difference operation on the ECG signal and respiratory signal respectively, calculate the amplitude change rate between adjacent sampling points, quantize the instantaneous fluctuation rate of the signal, arrange the change rate data according to the time series, and generate a physiological signal gradient sequence. S302: For the physiological signal gradient sequence, call the bilateral partial derivative calculation function to calculate the average gradient difference between the left and right neighborhoods of each time node, compare the absolute value of the difference with the preset gradient jump threshold, filter the time node index where the absolute value of the difference exceeds the gradient jump threshold, and identify the location of the abrupt change point. S303: Using the mutation point location as the central time anchor point, set a symmetrical sliding window of a preset width, determine the start and end boundaries of the window on the time axis, and perform local truncation of the ECG and respiratory signals in the rhythm instability interval according to the start and end boundaries to obtain signal segments including mutation features and their surrounding states, and extract the waveform segment of the mutation point.

[0012] As a further aspect of the present invention, the process of setting the gradient jump threshold specifically involves selecting a historical physiological signal segment of the patient in a stable resting state, performing a first-order difference operation on it to generate a reference gradient sequence, calculating the standard deviation statistic of the amplitude of the reference gradient sequence, multiplying the standard deviation statistic with a preset sensitivity gain coefficient, and the resulting product value is the gradient jump threshold.

[0013] As a further aspect of the present invention, the step of obtaining the critical point for the transition of the cardiopulmonary synchronization state specifically includes: S401: For the waveform segment at the abrupt change point, retrieve the full data of the opposite signal in the corresponding time window, construct a joint probability distribution model of the two signals using the kernel density estimation method, calculate the edge entropy of the waveform segment at the abrupt change point and the conditional entropy relative to the opposite signal based on the distribution model, and generate the conditional mutual information value. S402: Perform a first-order gradient operation on the conditional mutual information value, calculate the slope and magnitude of the change of the mutual information index in the time domain, extract the positive and negative signs of the slope, identify the direction of evolution of the coupling strength increase or decrease, combine the magnitude of the deviation to quantify the dynamic variation of the synchronization relationship, and generate a mutual information evolution feature vector. S403: Based on a preset deviation standard threshold, the deviation amplitude in the mutual information evolution feature vector is numerically compared with the deviation standard threshold. When the deviation amplitude exceeds the deviation standard threshold, it is confirmed that a qualitative change has occurred in the signal coupling mode. The location of the mutation point that triggers the qualitative change is marked as the time of reorganization, and the critical point of cardiopulmonary synchronization state transition is determined.

[0014] As a further aspect of the present invention, the step of obtaining the health status identification result specifically includes: S501: Based on the critical point of the cardiopulmonary synchronization state transition, construct a temporally continuous rehabilitation state topology sequence, and perform equal division of the topology sequence according to the preset node number span to obtain equally weighted sub-path segments; S502: For the equal-weighted sub-path segments, extract the actual state transition vector and the corresponding standard reference vector between each node, monitor the dwell time of each node in maintaining the current state, and count the total duration of the sub-path. Calculate the consistency score of the state transition direction, calculate the average dwell time of each stage, and establish a tuple including state transition features. S503: Call the preset standard path matching degree benchmark range, perform numerical comparison between the data items in the state transition feature tuple and the standard path matching degree benchmark range, identify potential abnormal chain segments, mark the risk level, and generate health status identification results.

[0015] A health status recognition system for thoracic surgery patients, the system being used to implement the aforementioned health status recognition method for thoracic surgery patients, the system comprising: The data sequence processing module acquires postoperative heart rate, blood pressure and respiratory rate data of thoracic surgery patients, performs normalization processing on the data, constructs the original recovery rhythm sequence, calls the variational autoencoder to map the original recovery rhythm sequence to the latent space, and constructs the predicted data sequence. The unstable interval identification module calculates the residual between the original recovery rhythm sequence and the predicted data sequence, extracts the maximum deviation value from the difference residual vector, compares the maximum deviation value with the fluctuation tolerance threshold, and calibrates the unstable interval of the rhythm. The abrupt change band segmentation module acquires the electrocardiogram (ECG) and respiratory signals in the non-stable interval of the rhythm, locates the abrupt change points in the ECG and respiratory signals, and extracts the waveform segments of the abrupt change points. The synchronous transition identification module calculates the conditional mutual information value between the waveform segment at the mutation point and the entire signal segment on the opposite side. When the conditional mutual information value changes, it determines the location of the mutation point and identifies the critical point of cardiopulmonary synchronization state transition. The state judgment and segmentation module constructs a rehabilitation state sequence and divides it into equally weighted sub-path segments based on the cardiopulmonary synchronous state transition critical point. It calculates the consistency score of the state transition direction and the average dwell time of each stage for the equally weighted sub-path segments. By comparing the standard path matching degree, it identifies potential abnormal chain segments and outputs the health status identification result.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by constructing the original recovery rhythm sequence and introducing a latent space prediction mechanism, a unified encoding process for changes in multi-parameter physiological rhythms is achieved, enhancing the identifiability of state evolution trends. The differential calculation of residual vectors combined with fluctuation tolerance thresholds is used to calibrate unstable intervals, making rhythm abnormalities more credible. The first derivative jump points of ECG and respiratory signals are used to identify abrupt change regions. Waveform segments are formed under a symmetrical sliding window, and conditional mutual information is introduced to evaluate the synchronous transition relationship between the two signals. This allows for mutual verification between cardiopulmonary state abrupt changes and rhythm fluctuations, establishing a state transition critical point judgment mechanism. By extracting directional consistency and average dwell time indicators from the rehabilitation sub-paths composed of continuous states, and combining this with standard path matching to form a potential abnormal chain identification method, this invention effectively solves the problems of inaccurate identification caused by multi-parameter disconnection, missed transition nodes, and lack of temporal trend in traditional methods. 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 these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a detailed flowchart of S1 of the present invention; Figure 3 This is a detailed flowchart of the S2 process of the present invention; Figure 4 This is a detailed flowchart of the S3 process of the present invention; Figure 5 This is a detailed flowchart of the S4 process of the present invention; Figure 6 This is a detailed flowchart of S5 of the present invention; Figure 7 This is a system flowchart 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 technical solution, a method for identifying the health status of thoracic surgery patients, comprising the following steps: S1: Obtain postoperative heart rate, blood pressure and respiratory rate data of thoracic surgery patients, perform mean skewing and variance normalization on the postoperative heart rate, blood pressure and respiratory rate data to construct the original recovery rhythm sequence, call the variational autoencoder to map the original recovery rhythm sequence to the latent space, and generate the prediction data sequence. S2: For each time node, perform residual calculation on the amplitude offset between the original recovery rhythm sequence and the predicted data sequence to generate a residual vector. Perform first-order difference operation on the residual vector to extract the maximum deviation value. Compare the maximum deviation value with the fluctuation tolerance threshold defined based on the recovery period segmentation to calibrate the rhythm instability interval. S3: Obtain the ECG and respiratory signals corresponding to the rhythm instability interval, use the bilateral partial derivative calculation function to identify the abrupt change points where the first derivative of the ECG and respiratory signals jumps in the time series, and set a symmetrical sliding window at the abrupt change point to extract the waveform segment of the abrupt change point. S4: Call the conditional mutual information algorithm to calculate the conditional mutual information value between the waveform segment of the mutation point and the entire signal segment on the opposite side. When the conditional mutual information value increases or decreases within the standard threshold, the location of the mutation point is determined to be the critical point for the transition of the cardiopulmonary synchronization state. S5: Construct a continuous rehabilitation state sequence based on the critical point of cardiopulmonary synchronous state transition and divide it into sub-path segments with equal weights. Extract the consistency score of state transition direction and the average dwell time of each stage for each sub-path segment. Compare the consistency score of state transition direction and the average dwell time of each stage with the matching degree of the standard path to identify potential abnormal chain segments and output the health status identification results.

[0022] The predicted data sequence includes reconstructed heart rate time series values, predicted blood pressure fluctuation trajectory, and fitted respiratory rate curve. The rhythm instability interval includes abnormal start timestamps, distortion termination timestamps, and non-steady-state confidence markers. The abrupt change waveform segment includes local high-frequency amplitude vectors, symmetrical time window spans, and signal channel identifiers. The cardiopulmonary synchronization state transition critical points include state switching time indexes, mutual information gradient change rates, and coupling pattern classification labels. The health status identification results include abnormal rehabilitation chain coding, recovery pattern deviation matrix, and postoperative risk level assessment.

[0023] Please see Figure 2 The specific steps for obtaining the predicted data sequence are as follows: S101: Obtain postoperative heart rate, blood pressure, and respiratory rate data of thoracic surgery patients, calculate the arithmetic mean and standard deviation of each data item, perform point-by-point subtraction to remove bias based on the arithmetic mean, perform division normalization on the removed data using the standard deviation to eliminate dimensional differences, and perform temporal alignment and matrix concatenation on the normalized multidimensional data according to a unified timestamp index to construct the original recovery rhythm sequence; A high-precision bedside monitor is used as the physical signal acquisition source, and the system connects to the hospital's clinical monitoring database in real time using HL7 or FHIR standard communication protocol interfaces. At this stage, object recognition technology is introduced to automatically identify and classify waveform morphology objects in the real-time incoming raw signal stream, intelligently distinguishing between normal sinus rhythm objects, paced rhythm objects, and various artifact / noise objects, ensuring that subsequent processing targets only valid physiological signal entities. The data acquisition strategy was set to high-frequency sampling at the second level, specifically a sampling frequency of 500Hz to ensure the integrity of the original waveform. Then, on the server side, downsampling was performed to 1Hz to obtain characteristic values ​​per second. For heart rate data, the RR interval was obtained through a three-lead or five-lead ECG unit and converted to beats per minute. For blood pressure data, if the patient wore an invasive arterial catheter, real-time waveform data of systolic and diastolic blood pressure were directly read; if non-invasive cuff monitoring was used, data was acquired at a set 15-minute interval, and the intermediate values ​​were supplemented using cubic spline interpolation to form a continuous time series. For respiratory rate data, chest impedance or carbon dioxide waveform data was extracted. After data acquisition, rigorous data cleaning and preprocessing were performed, with a physiological rationality filter to remove extreme values ​​caused by lead loss or body motion artifacts, and outliers identified and removed using the sliding interquartile range method. Subsequently, for the three independent time series data of heart rate, blood pressure, and respiratory rate after cleaning, their arithmetic mean and standard deviation statistics were calculated respectively. A time length was set as... A rolling statistical window (e.g., 30 minutes) slides along the time axis in 1-minute increments; within the current window, heart rate sequences are extracted. Blood pressure sequence and respiratory rate sequence All valid data points; for each type of data, use the formula Calculate the arithmetic mean over this time period, which reflects the patient's physiological baseline level at the current recovery stage; and simultaneously use the formula... Calculate the standard deviation statistic, which quantifies the dispersion and variability of physiological parameters; here... This represents the total number of valid sampling points within the window. Based on the calculated arithmetic mean, a point-by-point subtraction operation is performed on the postoperative heart rate, blood pressure, and respiratory rate data to remove bias. This step aims to eliminate individual baseline differences between different patients, focusing on the relative changes of physiological indicators relative to their own baseline. Specifically, for each raw data point in the time series... Subtract the arithmetic mean of the corresponding window. Obtain the biased data This operation shifts the data center to near zero, preserving the signal's fluctuation pattern while removing the DC component. Next, the standard deviation statistic is used to perform division normalization on the debiased data. Since heart rate, blood pressure, and respiratory rate have completely different physical dimensions and numerical ranges, directly performing multidimensional fusion would lead to larger numerical indicators dominating the model weights. Therefore, a Z-Score normalization strategy is adopted to normalize the debiased data. Divide by the corresponding standard deviation ,Right now This operation transforms all physiological indicators into dimensionless standardized values, ensuring they follow a standard normal distribution with a mean of 0 and a variance of 1, thus eliminating dimensional differences. Finally, the normalized multidimensional data is time-series aligned and matrix-stitched according to a unified timestamp index. Since the internal clocks of different monitoring devices may have slight deviations, a standard timeline sequence needs to be created based on the unified time on the server. For each type of normalized physiological data, nearest neighbor interpolation or linear interpolation algorithms are used to map it to... At each point in time, ensure that the three types of data are strictly aligned in the time dimension; then standardize the aligned heart rate sequence. Blood pressure standardized sequence and respiratory rate normalization sequence As a column vector, construct a dimension of The multidimensional feature matrix, where The time step of the sequence is represented by this matrix, which is the constructed "original recovery rhythm sequence" that fully preserves the synchronous fluctuation characteristics of multimodal physiological signals. Table 1 lists the key data acquisition parameters and preprocessing settings involved in this step. Table 1 Physiological signal acquisition and preprocessing parameters

[0024] Table 1 shows the standardized parameters for data preprocessing in the test results of the embodiment. The experimental data shows that when the parameter statistical window length is set to 30 minutes, compared with the traditional fixed baseline method, the accuracy of this method in capturing postoperative physiological fluctuation trends is improved by 12.5%, effectively smoothing out short-term noise interference.

[0025] S102: For the original recovery rhythm sequence, perform time-series feature encoding, extract nonlinear fluctuation features in the sequence through a multi-layer neural network and map them to a latent variable space of a preset dimension, calculate the mean vector and log-variance vector of the feature distribution in the latent space, perform Gaussian distribution reparameterization sampling operation based on the mean vector and log-variance vector, and obtain the variational sequence embedding vector that can characterize the potential manifold structure of the physiological signal. Will The multidimensional feature matrix is ​​divided into time slices of fixed length (e.g., time step). The input tensor of the encoder network is used as the input tensor. Each input time slice is treated as an independent "physiological time sequence object," and the feature extraction capability of the neural network is used to perform deep semantic recognition on these objects. The encoder network adopts a deep architecture composed of a Long Short-Term Memory network or a one-dimensional convolutional neural network. In this embodiment, a three-layer stacked 1D-CNN structure is specifically selected, with each convolutional layer followed by a batch normalization layer and a LeakyReLU activation function. The first convolutional kernel size is set to 3, and the number of filters is 32, used to capture short-term local fluctuations; the second convolutional kernel size is 5, and the number of filters is 64, used to extract mesoscale rhythmic features; the third convolutional kernel size is 7, and the number of filters is 128, used to integrate long-term dependencies; after multiple layers of nonlinear mapping, the original time series is compressed into a high-dimensional abstract feature map; subsequently, the extracted features are mapped to a latent variable space of a preset dimension. The dimension of the latent variable space is... This is a key hyperparameter, which is set to 16 in this embodiment after experimental verification. To construct the probabilistic model structure of the variational autoencoder, the encoder output is not directly used as the latent vector, but is mapped to the mean vector of the feature distribution in the latent space through two parallel fully connected layers. and log-variance vector This essentially constructs a probability distribution description of the physiological state object within the latent space, where the mean vector determines the object's center position and the log-variance vector determines the object's boundary range. Gaussian distribution reparameterization sampling is then performed based on the mean and log-variance vectors; this is a core step in variational inference, aiming to solve the problem of non-differentiability of the sampling operation leading to the inability to backpropagate gradients. The specific process is as follows: first, from the standard normal distribution... Randomly sample an auxiliary noise vector Its dimensions and hidden space dimensions Consistent. Then, using the formula... Calculate the latent variables Finally, a variational sequence embedding vector capable of characterizing the underlying manifold structure of physiological signals is obtained. This vector... This method condenses the deep temporal dependencies and nonlinear correlation features of the input physiological sequences. Through the above process, it completes the recognition and transformation from the original data stream to high-dimensional latent space objects. This embedding vector is the digital fingerprint of the identified physiological state object. To train this encoding network, a composite loss function including reconstruction error and KL divergence is used. Experimental data shows that when the latent space dimension... When the network parameters are set to 16 and the number of convolutional layers is 3, the model achieves the highest encoding efficiency for early signs of postoperative complications, with an information retention rate of 98.2%. The specific network parameter configurations are shown in Table 2.

[0026] Table 2. Variational Encoder Network Architecture Parameters

[0027] S103: The generative decoding network is invoked to perform inverse mapping and upsampling reconstruction operations on the variational sequence embedding vector, restoring the latent space features to the dimensional space of the original time series data. The inverse normalization transformation based on the preceding statistical features is performed on the reconstructed output data to restore the physical dimension amplitudes of heart rate, blood pressure and respiratory rate. The restored multidimensional data stream is combined according to the time step sequence to generate the predicted data sequence. The generative decoding network is invoked to perform inverse mapping and upsampling reconstruction operations on the variational sequence embedding vector. The architecture of the decoding network is mirror-symmetric to that of the encoding network, aiming to incorporate low-dimensional latent variables. The dimensions are restored to those of the original time series. First, the 16-dimensional variational sequence is embedded into a vector. The input is fed into a fully connected layer, mapping it back to the same tensor shape as the output of the encoder's last convolutional layer. Subsequently, upsampling is performed through a series of one-dimensional transposed convolutional layers. Each transposed convolutional layer gradually restores the length of the time series while performing feature expansion. This process is equivalent to using the decoder to reconstruct the abstract objects identified in the latent space, attempting to restore the object's physical form in the temporal domain. After restoring the latent space features to the dimensional space of the original time series data, an inverse normalization transformation based on prior statistical features is performed on the reconstructed output data; the recorded and stored arithmetic mean of the corresponding time window is then retrieved. and standard deviation The inverse transform is the mathematical inverse process of normalization, which is applied to each standardized value in the reconstructed sequence. First, perform the multiplication operation. The data fluctuation range is restored; then the addition operation is performed. The absolute physical baseline values ​​of the data are restored. This process ensures that the generated data is not only realistic in waveform morphology but also has practical clinical physical significance in numerical value. After restoring the physical dimensions of heart rate, blood pressure, and respiratory rate, the restored multidimensional data stream is combined according to the time step sequence. Since the decoder outputs segmented time-slice data, it is necessary to use the overlapping addition method or splicing method to integrate the continuously generated segments into a complete time series, generating a predicted data sequence. This sequence represents the "theoretical steady state" or "reconstruction path" deduced by the model based on the current potential physiological state manifold structure. Through this generative reconstruction, an "ideal physiological object" is actually constructed as a reference system for subsequent comparative analysis. Experimental verification shows that when the parameters are set as shown in Table 2, the root mean square error (RMSE) of the reconstruction of patient data in the steady period is as low as 0.05 (normalized scale).

[0028] Please see Figure 3 The specific steps for obtaining the non-stable intervals of the rhythm are as follows: S201: Perform point-by-point difference calculation on the original recovery rhythm sequence and the predicted data sequence at each time node, calculate the amplitude difference between the two at the corresponding time step, take the absolute value of the difference value to quantify the single-point reconstruction error, arrange the reconstruction error data of each time node in time sequence and combine them into a vector form to generate a multidimensional amplitude residual sequence that represents the distribution of the model prediction bias over time. Ensure the original sequence With predicted sequence The timelines are fully aligned and of uniform length. This comparison process aligns the "actual observed objects" with the "ideal predicted objects" generated by the model, aiming to identify anomalous object components hidden in the data through difference analysis; for each time point in the sequence... Extract the original multidimensional vector corresponding to that moment. and predicting multidimensional vectors Perform vector subtraction operation The magnitude difference between the two at the corresponding time steps is calculated; then, the absolute value of the difference is taken to quantify the single-point reconstruction error. For the difference vector... For each element in the array, calculate its absolute value to obtain... Using absolute values ​​instead of squared values ​​is to maintain consistency between the dimensions of the error and the original physical quantities, and to avoid excessive amplification of the error metric by extreme values; the reconstructed error data at each time point are arranged chronologically and combined into a vector form to generate a multidimensional amplitude residual sequence representing the distribution of the model prediction bias over time; the error vectors at all times are... Stacked in chronological order, forming a structure with dimension [missing information]. residual matrix Each row of the matrix corresponds to a time point, and each column corresponds to the reconstruction error flow of a physiological indicator. This multidimensional amplitude residual sequence is the core input for subsequent anomaly object recognition. The high-amplitude continuous residuals capture nonlinear dynamic features that the model cannot effectively reconstruct, clearly indicating potential pathophysiological abnormalities.

[0029] S202: Perform first-order difference operation on the elements of three consecutive time nodes in the multidimensional amplitude residual sequence, calculate the rate of change between adjacent residual values, capture the error fluctuation trend, retrieve the absolute value of the difference operation result within a sliding window of preset width, filter the difference result with the largest value within the window as an indicator of the degree of local variation, and extract the difference extreme value feature. Let the multidimensional amplitude residual sequence be... For any point in time Select it and the previous moment and the next moment Form a group of three consecutive points; calculate the rate of change between adjacent residual values: as well as This first-order difference operation can eliminate the constant bias component in the residual, which is a key step in identifying transient change objects; then, the absolute value of the difference operation result is retrieved within a sliding window of a preset width; the width of the sliding window is set. The window, with a duration of 10 seconds, slides point-by-point across the difference sequence; at each window position, all difference values ​​are extracted. absolute value The maximum difference value within the window is selected as an indicator of the degree of local variation, and the extreme value feature of the difference is extracted. Among the 10 absolute difference values ​​in the current window, the maximum value is selected using a maximum value function. Filter out peak values The differential extreme value feature is actually a quantitative description of the most significant abnormal fluctuations within a local range. The maximum value is chosen instead of the average value as the feature because, in clinical monitoring, transient and dramatic fluctuations are often more alarming than the average value.

[0030] S203: Call the preset fluctuation tolerance threshold, compare the differential extreme value feature with the fluctuation tolerance threshold, detect whether the differential extreme value feature exceeds the fluctuation tolerance threshold, record the start and end timestamps of the current time window when the detection result is exceeded, mark the time domain segment covered by the start and end timestamps as an abnormal state, and mark the rhythm non-stable interval. The process of setting the fluctuation tolerance threshold is as follows: the rehabilitation cycle is divided into multiple characteristic periods according to the postoperative recovery time process, historical steady-state sample data corresponding to the current characteristic period is obtained, the first-order difference statistical features of the reconstruction error of the historical steady-state sample data in the variational space are calculated, the distribution mean and standard deviation are extracted, and the sum of the mean and the standard deviation of the preset confidence coefficient weight is established as the fluctuation tolerance threshold of the current period. The difference extreme value characteristics of each time window With dynamically set fluctuation tolerance threshold A comparison is performed. This comparison logic is a Boolean condition: if... If the detection result exceeds the threshold, it is determined that the detection result is excessive. When the detection result exceeds the threshold, the start and end timestamps of the current time window are recorded, and the time domain segment covered by the start and end timestamps is marked as an abnormal state, and the non-steady interval of the rhythm is identified. Through this determination process, the binary classification and recognition task of "physiological rhythm objects" within the current time window is completed, that is, it is identified as "steady-state objects" or "non-steady-state abnormal objects". The start time corresponding to the window that triggers the threshold will be automatically extracted. and end time The data for that time period is then tagged in the database. The process of setting the fluctuation tolerance threshold involves: dividing the recovery period into multiple characteristic time periods based on the postoperative recovery timeline; obtaining historical steady-state sample data corresponding to the current patient's specific characteristic time period from the hospital's historical database; calculating the first-order difference statistical characteristics of the reconstruction error of the historical steady-state sample data in the variational space, and extracting its distribution mean. with standard deviation Based on statistical principles, the sum of the standard deviation of the mean and the pre-set confidence coefficient weights is established as the allowable fluctuation threshold for the current period. The specific calculation logic is as follows: the obtained mean parameter... (For example, a value of 0.02) and the standard deviation parameter (e.g., a value of 0.01), and preset confidence coefficient weights. (In this embodiment, 3.5 is selected), substituting into the formula The final result was obtained through calculation. for Experimental data show that when the confidence coefficient When the value is set to 3.5, the detection rate of atrial fibrillation precursors reaches 94%.

[0031] Please see Figure 4 The specific steps for obtaining the waveform segment at the mutation point are as follows: S301: Based on the timestamp range of the non-stable interval of the rhythm, extract the corresponding time period's electrocardiogram (ECG) and respiratory signals from the monitoring database, perform first-order discrete difference operations on the ECG and respiratory signals respectively, calculate the amplitude change rate between adjacent sampling points, quantify the instantaneous fluctuation rate of the signal, arrange the change rate data according to the time series, and generate a physiological signal gradient sequence that reflects the dynamic change trend of cardiopulmonary function. The start time of the non-stationary interval of the rhythm read from the marker. and termination time Furthermore, a buffer window is extended before and after this buffer. Based on this extended time range, original ECG and respiratory waveform data with a sampling rate of 500Hz are indexed and loaded from the high-frequency raw database; first-order discrete difference operations are performed on the ECG and respiratory signals respectively; for the extracted discrete time series signals... Calculate the rate of change of amplitude between adjacent sampling points, i.e. This operation aims to extract the gradient features of the signal, thereby providing enhanced feature input for subsequent recognition of specific morphological objects. For electrocardiogram signals, large differential values ​​correspond to the rapid depolarization phase of the QRS complex; for respiratory signals, the differential values ​​reflect the instantaneous flow rate changes during inspiration and expiration. The rate of change data are arranged in a time series to generate a physiological signal gradient sequence reflecting the dynamic changes in cardiopulmonary function. Each calculated differential point... Reorganized according to the original sampling order, two parallel time series were constructed: the cardiac elevator sequence. and respiratory gradient sequence These two sequences visually demonstrate the instantaneous energy bursts and morphological change rates of the cardiopulmonary system, making potential pathological mutations (such as premature beat waveforms and apnea segments) more prominent in the gradient domain, facilitating subsequent identification.

[0032] S302: For the physiological signal gradient sequence, call the bilateral partial derivative calculation function to calculate the average gradient difference between the left and right neighborhoods of each time node, compare the absolute value of the difference with the preset gradient jump threshold, filter the time node index where the absolute value of the difference exceeds the gradient jump threshold, and identify the location of the abrupt change point. The specific process of setting the gradient jump threshold is as follows: select a historical physiological signal segment of the patient in a stable resting state, perform a first-order difference operation on it to generate a baseline gradient sequence, calculate the standard deviation statistic of the amplitude of the baseline gradient sequence, multiply the standard deviation statistic with the preset sensitivity gain coefficient, and the calculated product value is the gradient jump threshold. To accurately capture the instant when the signal properties change, a value is defined based on the current detection point. A local observation window centered on the center, with a window width of Divide the window into a left half and a right half; calculate the mean of the gradient sequence in the left neighborhood for each half. and the mean of the gradient sequence in the right neighborhood Calculate the difference between the two. This calculation process essentially constructs an edge detection operator to identify inflection points in gradient change trajectories. It compares the absolute value of the difference with a preset gradient jump threshold, filters time points where the absolute value of the difference exceeds the threshold, and identifies abrupt change locations. The absolute value of the difference is then calculated. and compare it with dynamic thresholds Compare; if Then determine the current point This is considered a potential signal abrupt change point. The specific process for setting the gradient jump threshold is as follows: Select a historical physiological signal segment from when the patient is in a stable resting state, perform the same first-order difference operation as described above on it, and generate a baseline gradient sequence. Calculate the standard deviation statistic of the amplitude of the baseline gradient sequence. The standard deviation statistic is compared with the preset sensitivity gain coefficient. Perform multiplication; in this embodiment, Set to 6.0; the specific calculation logic is as follows: the obtained baseline gradient standard deviation is... (For example, a value of 0.05mV / ms) and the preset sensitivity gain coefficient (Value 6.0) Substitute into the formula The gradient jump threshold is calculated. It is 0.3 mV / ms; experimental data show that, at Under the specified settings, the algorithm achieves a positioning accuracy of less than 2ms for the R-wave peak and the onset of abnormal premature beats.

[0033] S303: Using the mutation point as the central time anchor point, set a symmetrical sliding window of preset width, determine the start and end boundaries of the window on the time axis, and perform local truncation of the ECG and respiratory signals in the rhythm instability interval according to the start and end boundaries to obtain signal segments including mutation features and the states of the surrounding neighborhoods, and extract the waveform segment of the mutation point. Based on mutation point index Define a time window width The chosen window width must ensure complete coverage of the physiological event to be identified; for ECG signals, the window is typically set to 600ms; for respiratory signals, the window is set to 4 seconds. Determine the starting boundary of the window on the time axis. Termination Boundary Based on the start and end boundaries, local truncation of ECG and respiratory signals within the rhythm instability interval is performed; using the calculated boundary index, array slicing is performed on the original high-frequency signal sequence to obtain signal segments including mutation features and their preceding and following neighborhood states; the waveform segment of the mutation point is extracted. This waveform segment is not just a piece of data, but a clearly identified and encapsulated "mutation event object", containing the original voltage value, timestamp, and sampling rate metadata, which serves as an independent sample unit for subsequent mutual information analysis.

[0034] Please see Figure 5 The specific steps for obtaining the critical point of cardiopulmonary synchronization transition are as follows: S401: For the waveform segment at the abrupt change point, retrieve the full data of the opposite signal in the corresponding time window, construct a joint probability distribution model of the two signals using the kernel density estimation method, calculate the marginal entropy of the waveform segment at the abrupt change point and the conditional entropy relative to the opposite signal based on the distribution model, calculate the statistical index characterizing the instantaneous coupling strength between signals according to Shannon's information theory formula, and generate the conditional mutual information value. The electrocardiogram (ECG) waveform object was paired with the respiratory signal object to construct a multimodal joint analysis object; the joint probability distribution model of the two signals was constructed using kernel density estimation (KDE); assuming the ECG signal segment is... The respiratory signal segment is Using the Gaussian kernel function, through the formula Estimate the joint probability density function of the two. and their respective marginal probability density functions and The marginal entropy and conditional entropy relative to the opposite signal of the waveform segment at the abrupt change point are calculated based on the distribution model. First, the marginal entropy is calculated. Then calculate the joint entropy. Based on Shannon's information theory formula, a statistical index characterizing the instantaneous coupling strength between signals is calculated, generating conditional mutual information values. This embodiment calculates instantaneous mutual information. This value directly reflects the intensity of the modulation effect of respiratory movements on heart rate variability within this small time window; through this indicator, the functional coupling tightness between two physiological objects is quantified; for example, the mutual information value of a certain mutation segment is calculated to be 0.8 bits, while that of the adjacent normal segment is 1.2 bits. This significant decrease in value suggests that the mutation point may be accompanied by a temporary failure of the cardiopulmonary coupling mechanism.

[0035] S402: Perform first-order gradient operation on the conditional mutual information value, calculate the slope and magnitude of the change of mutual information index in the time domain, extract the positive and negative signs of the slope, identify the evolution direction of the increase or decrease of coupling strength, combine the magnitude of the deviation to quantify the dynamic variation of the synchronization relationship, and generate a mutual information evolution feature vector. Since the mutual information value is a continuous sequence that slides with the time window. It is necessary to analyze its dynamic evolution pattern; calculate the slope of the mutual information index in the time domain: Simultaneously calculate the magnitude of the numerical deviation: Extract the sign of the slope to indicate the direction of evolution of increasing or decreasing coupling strength; a positive sign indicates that cardiopulmonary synchrony is increasing; a negative sign indicates that synchrony is disintegrating; combine the deviation amplitude to quantify the dynamic variation of the synchronization relationship and generate a mutual information evolution feature vector; this feature vector It is a mathematical abstraction of the "coupled state evolution object," comprehensively describing "how fast the cardiopulmonary coupled state changes," "in what direction it changes," and "how far it deviates"; for example, eigenvectors. This indicates that the mutual information is decreasing rapidly at a rate of 0.5 units / s, and the current value has deviated from the baseline by 0.4 units.

[0036] S403: Based on the preset deviation standard threshold, the deviation amplitude in the mutual information evolution feature vector is numerically compared with the deviation standard threshold. When the deviation amplitude exceeds the deviation standard threshold and the evolution direction is clear, the signal coupling mode is confirmed to have undergone a qualitative change. The location of the mutation point that triggers the qualitative change is marked as the moment of dynamic reorganization, and the critical point of cardiopulmonary synchronization state transition is calibrated. The specific process for setting the deviation standard threshold is as follows: select historical conditional mutual information data in the stable recovery stage as defined by clinical medicine to construct a benchmark sample set; perform normal distribution fitting operation on the benchmark sample set to obtain mean and variance parameters; calculate the statistical boundary value covering the preset confidence interval based on the mean and variance parameters; and establish the statistical boundary value as the deviation standard threshold. Read the deviation magnitude component in the feature vector and compare it with the deviation standard threshold. Logical judgment is performed; when the deviation exceeds the deviation standard threshold and the evolution direction is clear, a qualitative change in the signal coupling mode is confirmed; the location of the mutation point that triggers the qualitative change is marked as the moment of dynamic reorganization, and the critical point of cardiopulmonary synchronization state transition is calibrated; this critical point is identified as a key "state transition object", marking the transition from one steady-state manifold to another manifold structure that may have pathological risks; the process of setting the deviation standard threshold is as follows: historical conditional mutual information data in the clinically defined stable recovery stage are selected to construct a benchmark sample set; normal distribution fitting operation is performed on the benchmark sample set to obtain the mean parameter. With variance parameter ; Calculate the statistical boundary values ​​covering the pre-set confidence interval based on the mean and variance parameters; the specific calculation logic is as follows: The obtained mean parameter... (e.g., a value of 1.2 bits) and standard deviation (For example, a value of 0.15 bits), substitute into the formula ;Right now The results indicate that if the mutual information deviates from the baseline by more than 0.387 bits, there is a 99% confidence level that this is not just random noise. Table 3 lists the threshold setting references for different confidence levels.

[0037] Table 3 Reference Table for Mutual Information Deviation Threshold Setting

[0038] Table 3 presents the test results of the embodiments. Through backtracking verification of historical data, when 0.387 bits corresponding to 99% confidence level is selected as the threshold, the average advance prediction time for cardiopulmonary decoupling events caused by postoperative lung infection reaches 2.5 hours.

[0039] Please see Figure 6 The specific steps for obtaining the health status recognition results are as follows: S501: Based on the critical point of cardiopulmonary synchronous state transition, construct a temporally continuous rehabilitation state topology sequence, and perform equal division of the topology sequence according to the preset node number span to obtain equally weighted sub-path segments with a unified metric scale. All critical points are arranged chronologically, dividing the entire postoperative recovery process into several distinct stages. Each segment is identified as an independent "rehabilitation state object" and abstracted as a node in a topological graph. Transitions between states are treated as directed edges, thus constructing a topological sequence reflecting the patient's recovery trajectory. The topological sequence is then equally divided according to a preset node span to obtain equally weighted sub-path segments with a uniform metric. To facilitate quantitative evaluation, natural stages of varying lengths cannot be directly compared; a fixed node span is defined. Alternatively, over a fixed time span, the entire rehabilitation trajectory is resampled into a fixed number of equidistant virtual nodes; each path connecting adjacent virtual nodes is a weighted sub-path segment. This completes the process of transforming a continuous time stream into discrete "path segment objects," making the recovery processes of different patients comparable.

[0040] S502: For equally weighted sub-path segments, extract the actual state transition vector and corresponding standard reference vector between each node, monitor the dwell time of each node in maintaining its current state, and calculate the total duration of the sub-path using the formula: ; The consistency score of the state transition direction is obtained by calculation, and the average dwell time of each stage is calculated to establish a state transition feature tuple that includes the consistency score and the duration statistics. in, The consistency score represents the direction of state transition. This represents the total number of state nodes within a sub-path segment. This parameter is obtained by counting the number of nodes contained in each of the equally weighted sub-path segments. Representing the The normalized actual state transition vector of the nth node, this parameter is extracted by... The coordinate displacements from one node to the next node are obtained by performing normalization in the vector space. Representing the The normalized standard reference vector corresponding to each node is obtained by calling the ideal transfer direction of the corresponding stage from the preset standard rehabilitation path library and performing unitization processing. Representing the The dwell time value at the nth node, this parameter is calculated by... Obtain the difference between the start and end timestamps of each node's state. This represents the total duration of the sub-path segment. This parameter is obtained by summing the dwell time values ​​of all nodes within the sub-path segment. For the first in the subpath Each node is used to determine its coordinates in the multidimensional physiological feature space. and the coordinates of the next node ; Calculate the actual displacement vector Perform normalization on the vector to obtain This vector is a mathematical representation of the current patient's "actual recovery direction object"; simultaneously, it calls a pre-defined standard rehabilitation path library; this library stores "ideal trajectories" generated based on thousands of successful rehabilitation cases; based on the patient's current postoperative time and physiological state, it retrieves the corresponding ideal next-state coordinates from the standard library. The standard reference vector is calculated and normalized. This refers to the "ideal reference direction object"; it monitors the duration of each node's stay in its current state and calculates the total duration of sub-paths; it records the patient's current state. Actual time length ; Calculate the sum of dwell times for all nodes within the sub-path segment. .

[0041] Formula used: The calculation obtains a consistency score for the state transition direction; the core logic of this formula lies in calculating the cosine similarity (i.e., consistency of direction) between the actual object and the reference object, and combining it with time weights to quantitatively assess the compliance of the rehabilitation process.

[0042] in: The dot product of two unit vectors, i.e., cosine similarity, has a value range of [-1, 1]; the absolute value indicates the parallelism of the directions of interest. Weighted by the length of stay; Normalize the total duration; The square root is used to smooth the distribution of numerical values.

[0043] Calculate the average dwell time during the calculation phase and establish a state transition feature tuple that includes consistency score and duration statistics. .

[0044] For example: Assume a subpath contains 2 nodes; node 1's actual vector is exactly the same as the reference vector (dot product = 1), and it stays for 30 minutes; node 2's vector is perpendicular to the reference vector (dot product = 0), and it stays for 10 minutes; the total duration is 40 minutes. Substituting these values ​​into the formula yields the result. The result of 0.866 indicates that the overall rehabilitation consistency score remains high.

[0045] S503: Call the preset standard path matching degree benchmark range, compare the data items in the state transition feature tuple with the standard path matching degree benchmark range, identify potential abnormal chain segments, mark the risk level, and generate health status identification results; The process of setting the standard path matching degree benchmark range is as follows: access the historical rehabilitation database to retrieve multiple sets of historical state sequence samples that have been verified as normal recovery categories, calculate the similarity value of each set of historical state sequence samples relative to the standard reference path, construct a benchmark sample similarity set, perform normal distribution fitting operation on the values ​​in the set, obtain the distribution mean and standard deviation statistics, calculate the confidence interval boundary based on the distribution mean and the standard deviation of the preset multiple, and establish the numerical range defined by the confidence interval boundary as the standard path matching degree benchmark range; The calculated consistency score With the preset reference range Compare. If If it is determined to be a "normal trajectory"; If it is determined to be "recovery lag or directional deviation"; The process involves identifying "over-rapid recovery," identifying potential abnormal chain segments, marking them with risk levels, and generating health status identification results. Through the above comparisons, the overall identification and classification of the patient's "health status object" is completed, and detailed risk reports are generated for abnormal objects. The process of setting the standard path matching benchmark range is as follows: Accessing the historical rehabilitation database to retrieve multiple sets of historical status sequence samples that have been verified as belonging to the normal recovery category. For each set of historical status sequence samples, the similarity value relative to the standard reference path is calculated, constructing a benchmark sample similarity set. Perform a normal distribution fitting operation on the values ​​in the set to obtain the distribution mean. and standard deviation statistic The confidence interval boundaries are calculated based on the standard deviation of the distribution mean and a preset multiple. The specific calculation logic is as follows: the obtained mean parameter... (For example, a value of 0.90) and standard deviation (For example, a value of 0.04), substitute into the interval formula The calculated numerical range This result indicates that for a new postoperative patient, if their state transition direction consistency score is high... If the score falls between 0.82 and 0.98, there is a 95% probability that the recovery process is consistent with historical successful cases, indicating a healthy recovery state.

[0046] Please see Figure 7 A health status recognition system for thoracic surgery patients, the system being used to execute the aforementioned health status recognition method for thoracic surgery patients, the system comprising: The data sequence processing module acquires postoperative heart rate, blood pressure and respiratory rate data of thoracic surgery patients, performs normalization processing on the data, constructs the original recovery rhythm sequence, calls the variational autoencoder to map the original recovery rhythm sequence to the latent space, and constructs the predicted data sequence. The unstable interval identification module calculates the residual between the original recovery rhythm sequence and the predicted data sequence, extracts the maximum deviation value from the difference residual vector, compares the maximum deviation value with the fluctuation tolerance threshold, and calibrates the unstable interval of the rhythm. The abrupt change segmentation module acquires ECG and respiratory signals during rhythm instability intervals, locates abrupt change points in the ECG and respiratory signals, and extracts waveform segments from the abrupt change points. The synchronous transition identification module calculates the conditional mutual information value between the waveform segment at the mutation point and the entire signal segment on the opposite side. When the conditional mutual information value changes, it determines the location of the mutation point and identifies the critical point of cardiopulmonary synchronization state transition. The state judgment and segmentation module constructs a rehabilitation state sequence based on the cardiopulmonary synchronous state transition critical point and divides it into equally weighted sub-path segments. It calculates the consistency score of the state transition direction and the average dwell time of each stage for the equally weighted sub-path segments. By comparing the matching degree of the standard path, it identifies potential abnormal chain segments and outputs the health status identification result.

[0047] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the related objects are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for a more specific understanding.

[0048] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0049] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0050] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware; whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0051] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0052] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways; for example, the apparatus embodiments described above are merely illustrative, for example, the division of the units is merely a logical functional division, and there may be other division methods in actual implementation, for example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed; another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of the devices or units may be electrical, mechanical or other forms.

[0053] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0054] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0055] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0056] 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 method for identifying the health status of thoracic surgery patients, characterized in that, Includes the following steps: S1: Obtain postoperative heart rate, blood pressure and respiratory rate data of thoracic surgery patients, perform normalization on the data, construct the original recovery rhythm sequence, call the variational autoencoder to map the original recovery rhythm sequence to the latent space, and construct the predicted data sequence. S2: Calculate the residual between the original recovery rhythm sequence and the predicted data sequence, extract the maximum deviation value from the difference residual vector, compare the maximum deviation value with the fluctuation tolerance threshold, and calibrate the rhythm instability interval; S3: Obtain the electrocardiogram (ECG) and respiratory signals in the non-stable interval of the rhythm, locate the abrupt change points of the ECG and respiratory signals, and extract the waveform segments of the abrupt change points; S4: Calculate the conditional mutual information value between the waveform segment at the mutation point and the entire signal segment on the opposite side. When the conditional mutual information value changes, determine the location of the mutation point and identify the critical point for the transition of the cardiopulmonary synchronization state. S5: Based on the critical point of cardiopulmonary synchronous state transition, construct a rehabilitation state sequence and divide it into equally weighted sub-path segments. Calculate the consistency score of the state transition direction and the average dwell time of each stage for the equally weighted sub-path segments. By comparing the standard path matching degree, identify potential abnormal chain segments and output the health status identification result.

2. The health status identification method for thoracic surgery patients according to claim 1, characterized in that, The predicted data sequence includes reconstructed heart rate time series values, predicted blood pressure fluctuation trajectory, and fitted respiratory rate curve. The rhythm instability interval includes abnormal start timestamps, distortion termination timestamps, and non-steady-state confidence markers. The abrupt change waveform segment includes local high-frequency amplitude vectors, symmetrical time window spans, and signal channel identifiers. The cardiopulmonary synchronization state transition critical points include state switching time indexes, mutual information gradient change rates, and coupling mode classification labels. The health status identification results include abnormal rehabilitation chain coding, recovery mode deviation matrix, and postoperative risk level assessment.

3. The health status identification method for thoracic surgery patients according to claim 1, characterized in that, The specific steps for obtaining the predicted data sequence are as follows: S101: Obtain postoperative heart rate, blood pressure, and respiratory rate data of thoracic surgery patients, calculate the arithmetic mean and standard deviation of each data item, perform point-by-point subtraction to remove bias based on the arithmetic mean, perform division normalization on the removed data using the standard deviation to eliminate dimensional differences, and perform temporal alignment and matrix concatenation on the normalized multidimensional data according to a unified timestamp index to construct the original recovery rhythm sequence; S102: For the original recovery rhythm sequence, perform time-series feature encoding, extract nonlinear fluctuation features in the sequence through a multi-layer neural network and map them to a latent variable space of a preset dimension, calculate the mean vector and log-variance vector of the feature distribution in the latent space, and perform Gaussian distribution reparameterization sampling operation based on the mean vector and log-variance vector to obtain the variational sequence embedding vector. S103: The generative decoding network is invoked to perform inverse mapping and upsampling reconstruction operations on the variational sequence embedding vector, restoring the latent space features to the dimensional space of the original time series data. The reconstructed output data is subjected to inverse normalization transformation based on prior statistical features to restore the physical dimension amplitudes of heart rate, blood pressure and respiratory rate. The restored multidimensional data stream is combined according to the time step sequence to generate the predicted data sequence.

4. The health status identification method for thoracic surgery patients according to claim 3, characterized in that, The specific steps for obtaining the non-stable interval of the rhythm are as follows: S201: Perform point-by-point difference operation on the original recovery rhythm sequence and the predicted data sequence at each time node, calculate the amplitude difference value between the two at the corresponding time step, take the absolute value of the difference value to quantify the single-point reconstruction error, arrange the reconstruction error data of each time node in time sequence and combine them into a vector form to generate a multidimensional amplitude residual sequence. S202: Perform first-order difference operation on the elements of three consecutive time nodes in the multidimensional amplitude residual sequence, calculate the rate of change between adjacent residual values, capture the error fluctuation trend, retrieve the absolute value of the difference operation result within a sliding window of a preset width, filter the difference result with the largest value within the window, and extract the difference extreme value feature. S203: Call the preset fluctuation tolerance threshold, compare the differential extreme value feature with the fluctuation tolerance threshold, detect whether the differential extreme value feature exceeds the fluctuation tolerance threshold, record the start and end timestamps of the current time window when the detection result is exceeded, mark the time domain segment covered by the start and end timestamps as an abnormal state, and mark the rhythm unstable interval.

5. The health status identification method for thoracic surgery patients according to claim 4, characterized in that, The process of setting the fluctuation tolerance threshold is as follows: the rehabilitation cycle is divided into multiple characteristic time periods according to the postoperative recovery time process; historical steady-state sample data corresponding to the current characteristic time period is obtained; the first-order difference statistical features of the reconstruction error of the historical steady-state sample data in the variational space are calculated; the mean and standard deviation of its distribution are extracted; and the sum of the mean and the standard deviation of the preset confidence coefficient weight is established as the fluctuation tolerance threshold of the current time period.

6. The health status identification method for thoracic surgery patients according to claim 4, characterized in that, The specific steps for obtaining the waveform segment at the mutation point are as follows: S301: Based on the timestamp range of the non-stable interval of the rhythm, extract the electrocardiogram (ECG) signal and respiratory signal of the corresponding time period from the monitoring database, perform first-order discrete difference operation on the ECG signal and respiratory signal respectively, calculate the amplitude change rate between adjacent sampling points, quantize the instantaneous fluctuation rate of the signal, arrange the change rate data according to the time series, and generate a physiological signal gradient sequence. S302: For the physiological signal gradient sequence, call the bilateral partial derivative calculation function to calculate the average gradient difference between the left and right neighborhoods of each time node, compare the absolute value of the difference with the preset gradient jump threshold, filter the time node index where the absolute value of the difference exceeds the gradient jump threshold, and identify the location of the abrupt change point. S303: Using the mutation point location as the central time anchor point, set a symmetrical sliding window of a preset width, determine the start and end boundaries of the window on the time axis, and perform local truncation of the ECG and respiratory signals in the rhythm instability interval according to the start and end boundaries to obtain signal segments including mutation features and their surrounding states, and extract the waveform segment of the mutation point.

7. The health status identification method for thoracic surgery patients according to claim 6, characterized in that, The specific process for setting the gradient jump threshold is as follows: select a historical physiological signal segment of the patient in a stable resting state, perform a first-order difference operation on it to generate a reference gradient sequence, calculate the standard deviation statistic of the amplitude of the reference gradient sequence, multiply the standard deviation statistic with a preset sensitivity gain coefficient, and the calculated product value is the gradient jump threshold.

8. The health status identification method for thoracic surgery patients according to claim 6, characterized in that, The specific steps for obtaining the critical point of cardiopulmonary synchronization state transition are as follows: S401: For the waveform segment at the abrupt change point, retrieve the full data of the opposite signal in the corresponding time window, construct a joint probability distribution model of the two signals using the kernel density estimation method, calculate the edge entropy of the waveform segment at the abrupt change point and the conditional entropy relative to the opposite signal based on the distribution model, and generate the conditional mutual information value. S402: Perform a first-order gradient operation on the conditional mutual information value, calculate the slope and magnitude of the change of the mutual information index in the time domain, extract the positive and negative signs of the slope, identify the direction of evolution of the coupling strength increase or decrease, combine the magnitude of the deviation to quantify the dynamic variation of the synchronization relationship, and generate a mutual information evolution feature vector. S403: Based on a preset deviation standard threshold, the deviation amplitude in the mutual information evolution feature vector is numerically compared with the deviation standard threshold. When the deviation amplitude exceeds the deviation standard threshold, it is confirmed that a qualitative change has occurred in the signal coupling mode. The location of the mutation point that triggers the qualitative change is marked as the time of reorganization, and the critical point of cardiopulmonary synchronization state transition is determined.

9. The health status identification method for thoracic surgery patients according to claim 8, characterized in that, The specific steps for obtaining the health status recognition result are as follows: S501: Based on the critical point of the cardiopulmonary synchronization state transition, construct a temporally continuous rehabilitation state topology sequence, and perform equal division of the topology sequence according to the preset node number span to obtain equally weighted sub-path segments; S502: For the equal-weighted sub-path segments, extract the actual state transition vector and the corresponding standard reference vector between each node, monitor the dwell time of each node in maintaining the current state, and count the total duration of the sub-path. Calculate the consistency score of the state transition direction, calculate the average dwell time of each stage, and establish a tuple including state transition features. S503: Call the preset standard path matching degree benchmark range, perform numerical comparison between the data items in the state transition feature tuple and the standard path matching degree benchmark range, identify potential abnormal chain segments, mark the risk level, and generate health status identification results.

10. A health status recognition system for thoracic surgery patients, characterized in that, The system is used to implement the health status identification method for thoracic surgery patients according to any one of claims 1-9, the system comprising: The data sequence processing module acquires postoperative heart rate, blood pressure and respiratory rate data of thoracic surgery patients, performs normalization processing on the data, constructs the original recovery rhythm sequence, calls the variational autoencoder to map the original recovery rhythm sequence to the latent space, and constructs the predicted data sequence. The unstable interval identification module calculates the residual between the original recovery rhythm sequence and the predicted data sequence, extracts the maximum deviation value from the difference residual vector, compares the maximum deviation value with the fluctuation tolerance threshold, and calibrates the unstable interval of the rhythm. The abrupt change band segmentation module acquires the electrocardiogram (ECG) and respiratory signals in the non-stable interval of the rhythm, locates the abrupt change points in the ECG and respiratory signals, and extracts the waveform segments of the abrupt change points. The synchronous transition identification module calculates the conditional mutual information value between the waveform segment at the mutation point and the entire signal segment on the opposite side. When the conditional mutual information value changes, it determines the location of the mutation point and identifies the critical point of cardiopulmonary synchronization state transition. The state judgment and segmentation module constructs a rehabilitation state sequence and divides it into equally weighted sub-path segments based on the cardiopulmonary synchronous state transition critical point. It calculates the consistency score of the state transition direction and the average dwell time of each stage for the equally weighted sub-path segments. By comparing the standard path matching degree, it identifies potential abnormal chain segments and outputs the health status identification result.

Images