Thoracic surgery postoperative complication early warning system

Abstract

The invention relates to the technical field of medical information, and discloses a chest surgery postoperative complication early warning system which comprises a screening module and an early warning module. The screening module is used for performing multi-dimensional feature integration on the initial physiological data, and covering vital sign feature data (such as heart rate variability parameter grouping tags and blood pressure fluctuation parameter clustering tags), medical record feature data and postoperative monitoring feature data; and the early warning module performs dynamic noise filtering on the real-time physiological data stream, identifies an abnormal mode through a multi-source fusion layer, and generates a complication early warning instruction. The multi-source fusion layer comprises a data normalization module and a feature recombination module, and the feature recombination module realizes multi-dimensional weight fusion and dynamic threshold optimization based on the physiological parameter analysis layer, the risk self-adaption layer and the complication trigger layer. Through space-time correlation analysis, compensation decision matrix modeling and high-risk scene positioning, the real-time performance and accuracy of complication early warning are remarkably improved, and the method is suitable for multi-parameter cooperative monitoring of thoracic surgery postoperative patients.

Description

Technical Field

[0001] The present invention relates to the field of medical information technology, and particularly to a postoperative complication warning system for thoracic surgery. Background Art

[0002] In the field of thoracic surgery, the timely detection and intervention of postoperative complications are crucial for the prognosis of patients. Traditional postoperative monitoring mainly relies on the manual observation and experience judgment of medical staff on single physiological indicators, which has many limitations.

[0003] From the perspective of data processing, traditional monitoring modes are difficult to effectively integrate multi-dimensional physiological data. The physiological state of postoperative patients is usually reflected by multiple types of data such as vital signs, medical records, and postoperative monitoring. However, traditional methods often analyze single vital sign parameters such as heart rate and blood pressure in isolation, lacking the ability to integrate and process multi-dimensional features such as grouping labels generated by heart rate variability parameters and clustering labels generated by blood pressure fluctuation parameters, and unable to capture the potential associations between different data dimensions. For example, simply focusing on abnormal heart rate may ignore its co-variation relationship with blood pressure fluctuations and the medication history in medical records, resulting in missed detection of early signs of complications.

[0004] In terms of real-time monitoring, traditional systems have insufficient processing capabilities for real-time physiological data streams. In a clinical environment, physiological data is often interfered by noise. Traditional static filtering methods are difficult to adapt to the dynamic characteristics of physiological signals, easily leading to a decrease in the accuracy of abnormal pattern recognition. At the same time, the lack of an efficient multi-source data fusion mechanism makes it impossible to jointly analyze real-time monitoring data and historical medical data, and it is difficult to timely identify abnormal patterns in complex physiological trajectories, resulting in delayed warnings.

[0005] From the perspective of the warning mechanism, traditional methods lack a dynamic and adaptive risk assessment model. Existing warning systems usually adopt fixed thresholds or simple rule models, and are unable to dynamically adjust assessment parameters according to multi-scenario factors such as patient individual differences, surgical types, and postoperative recovery stages. For example, the physiological index change rules of postoperative patients in different thoracic surgeries (such as radical resection of lung cancer and esophagectomy) are different. Traditional fixed-threshold models are difficult to accurately adapt to the risk assessment requirements in different scenarios, and are prone to false alarms or missed alarms.

[0006] In addition, traditional systems have obvious defects in feature processing and modeling. For the spatio-temporal correlation features in physiological data, traditional methods lack effective means of analysis and recombination and cannot generate coupled feature data that can reflect the overall state of the physiological system. In terms of modeling dynamic correlation relationships, it is difficult to quantitatively analyze the real-time interaction between data in each dimension and unable to generate key data such as a compensation decision matrix, resulting in insufficient scientificity and reliability of early warning instructions. At the same time, for data normalization processing, traditional methods lack a refined time window segmentation and redundant feature elimination mechanism, making it difficult to ensure data consistency and usability and affecting the accuracy of subsequent analysis.

[0007] With the development of medical informatization, there is an urgent need for a postoperative complication early warning system that can integrate multi-dimensional physiological data, adapt to dynamic physiological changes, and have the ability to accurately identify abnormal patterns to address the deficiencies of traditional monitoring methods in data integration, real-time processing, dynamic modeling, etc., improve the early warning ability of postoperative complications in thoracic surgery, and improve the prognosis of patients. Summary of the Invention

[0008] The purpose of the present invention is to provide a postoperative complication early warning system for thoracic surgery to solve the problems raised in the above background technology.

[0009] To achieve the above purpose, the present invention provides the following technical solution: A postoperative complication early warning system for thoracic surgery, the system includes: A screening module for performing multi-dimensional feature integration processing on initial physiological data, where the initial physiological data includes an analysis data set corresponding to vital sign feature data, an associated feature group corresponding to medical record feature data, and an evaluation index set corresponding to postoperative monitoring feature data, and the vital sign feature data includes a grouping label generated by a heart rate variability parameter and a clustering label generated by a blood pressure fluctuation parameter; An early warning module for performing dynamic noise filtering processing on real-time physiological data streams and inputting them into a multi-source fusion layer for abnormal pattern recognition, and generating a complication early warning instruction according to the output result of the multi-source fusion layer; The multi-source fusion layer includes a data normalization module and a feature recombination module. Among them, the data normalization module is used for time window segmentation and redundant feature elimination of the original physiological data stream, and the feature recombination module is obtained by joint training based on multi-scenario historical medical data and real-time physiological trajectories; the feature recombination module includes a physiological parameter analysis layer, a risk adaptation layer, and a complication trigger layer connected in sequence.

[0010] Preferably, the physiological parameter analysis layer is used to perform spatio-temporal correlation processing on different analysis data sets in the original physiological data stream to generate physiological coupling feature data; the risk adaptation layer is used to model the dynamic correlation relationship between the physiological coupling feature data corresponding to each analysis data set to generate compensation decision matrix data; the complication trigger layer is used to perform multi-dimensional weight fusion based on the compensation decision matrix data and the physiological coupling feature data to generate complication warning instructions.

[0011] Preferably, the modeling of the dynamic correlation relationship between the physiological coupling feature data corresponding to each analysis data set to generate compensation decision matrix data includes: Using a physiological segmentation algorithm to identify the complication key nodes in the physiological coupling feature data, and determining the decision compensation feature groups corresponding to each analysis data set based on the physiological scenario type corresponding to each complication key node; Calculating the deviation coefficient between the physiological nodes of the same scenario in the decision compensation feature groups corresponding to any two analysis data sets, and generating the compensation decision matrix data between the any two analysis data sets based on the deviation coefficient.

[0012] Preferably, the calculating of the deviation coefficient between the physiological nodes of the same scenario in the decision compensation feature groups corresponding to any two analysis data sets includes: When the number of psychological nodes in the decision compensation feature groups corresponding to the any two analysis data sets is inconsistent, performing virtual physiological point interpolation based on the scenario parameters corresponding to the terminal physiological nodes in the one with fewer physiological nodes, and calculating the deviation coefficient between the physiological nodes of the same scenario based on the interpolated data.

[0013] Preferably, the data normalization module is specifically used for: Performing equal-frequency band division on the analysis data set, the associated feature group and the evaluation index set according to a preset time window to generate normalized physiological data, normalized medical data and normalized monitoring data; Performing real-time alignment on the normalized physiological data and the normalized medical data by using a dynamic feature clustering method, and performing steady-state optimization on the normalized monitoring data by using a fixed sliding window mechanism, and outputting a first calibration feature group, a second calibration feature group and a third calibration feature group; wherein, the first calibration feature group includes the calibrated grouping label and the calibrated clustering label.

[0014] Preferably, the data normalization module is further used for: Calculating the physiological deviation coefficient between the calibrated grouping label and the calibrated clustering label in the historical monitoring period; Predict the expected distribution value of the calibrated clustering label in the real-time monitoring period according to the physiological offset coefficient and the scene parameters of the calibrated grouping label in the real-time monitoring period; Generate target decision compensation data based on the calibrated clustering label and its expected distribution value, and use the analysis data set corresponding to the target decision compensation data as the first calibration feature group.

[0015] Preferably, the risk adaptation layer specifically includes: A pattern traceability unit, configured to perform physiological pattern traceability on each analysis data set in the physiological coupling feature data respectively, so as to extract the corresponding physiological propagation chain from each analysis data set; A scene matching unit, configured to perform spatio-temporal superposition on the physiological propagation chain extracted from each analysis data set and the corresponding physiological coupling feature data to generate compensation decision matrix data.

[0016] Preferably, the risk adaptation layer further includes: A noise suppression unit, configured to perform time lag effect elimination processing on the compensation decision matrix data.

[0017] Preferably, the complication trigger layer specifically includes: A multi-dimensional weight collaboration unit, including a plurality of warning decision nodes, and each warning decision node is connected to each analysis data set in the compensation decision matrix data and the physiological coupling feature data through parameter configuration; A dynamic threshold optimization unit, configured to iteratively optimize the parameter configuration through a dynamic threshold adjustment algorithm to minimize the error between the complication warning instruction and the actual physiological distribution; An abnormal pattern recognition unit, configured to perform high-risk scenario positioning based on the compensation decision matrix data and the physiological coupling feature data to generate a complication warning instruction.

[0018] Preferably, the dynamic feature clustering method specifically includes: Construct a dynamic similarity matrix based on the physiological distribution characteristics of real-time scene parameters; Perform dimensionality reduction clustering processing on the normalized physiological data by using feature space mapping.

[0019] Compared with the prior art, the beneficial effects of the present invention are: The screening module of the system can perform multi-dimensional feature integration processing on the initial physiological data. It integrates the analysis data set corresponding to the vital sign feature data, the associated feature group corresponding to the medical record feature data, and the evaluation index set corresponding to the postoperative monitoring feature data. In particular, it deeply processes key features such as the grouping labels generated by heart rate variability parameters and the clustering labels generated by blood pressure fluctuation parameters, breaking through the limitations of traditional single-index analysis. By integrating multi-dimensional data, the system can capture the comprehensive changes in the patient's physiological state as a whole, discover the potential associations between different data dimensions, such as the co-variation pattern between heart rate variability grouping and blood pressure fluctuation clustering, thus providing a more comprehensive basis for complication warning and effectively reducing the occurrence of missed detections.

[0020] The warning module performs dynamic noise filtering on the real-time physiological data stream. By using a filtering algorithm that adapts to the dynamic changes of physiological signals, it can more accurately remove noise interference and retain the true and effective physiological signal features. The processed data stream is input into the multi-source fusion layer for abnormal pattern recognition. The multi-source fusion layer performs time window segmentation and redundant feature elimination on the original physiological data stream through the data normalization module to ensure the standardization and usability of the input data. The data normalization module divides various types of data into equal frequency bands through a preset time window, generates normalized physiological data, medical data, and monitoring data, and uses a dynamic feature clustering method to align the physiological data and medical data in real time, while performing steady-state optimization on the monitoring data, effectively improving the consistency and analysis efficiency of the data.

[0021] The feature recombination module of the multi-source fusion layer is jointly trained based on multi-scenario historical medical data and real-time physiological trajectories. The physiological parameter analysis layer, risk adaptation layer, and complication trigger layer it contains form a complete feature processing and modeling link. The physiological parameter analysis layer performs spatio-temporal association processing on different analysis data sets in the original physiological data stream, generates physiological coupling feature data, which can effectively capture the association features of physiological indicators in time and space and reflect the overall state of the physiological system. The risk adaptation layer models the dynamic association relationships between the physiological coupling feature data corresponding to each analysis data set, identifies the key nodes of complications through a physiological segmentation algorithm, determines the decision compensation feature group, and calculates the deviation coefficient to generate the compensation decision matrix data, realizing the quantitative analysis of the dynamic interaction of data. It can dynamically adjust the evaluation model according to the individual differences of patients and surgical scenarios, improving the accuracy and adaptability of the warning. The noise suppression unit performs time lag effect elimination processing on the compensation decision matrix data, further improving the real-time performance and reliability of the data.

[0022] The complication trigger layer performs multi-dimensional weight fusion based on compensation decision matrix data and physiological coupling feature data. Through the collaborative action of the multi-dimensional weight coordination unit, dynamic threshold optimization unit, and abnormal pattern recognition unit, it can accurately locate high-risk scenarios and generate complication warning instructions. The dynamic threshold optimization unit minimizes the error between the warning instructions and the actual physiological distribution by iteratively optimizing parameter configurations, enabling the system to adapt to the physiological change characteristics of different patients and different postoperative stages and reducing the false alarm rate. The dynamic feature clustering method of the feature recombination module constructs a dynamic similarity matrix based on real-time scenario parameters and performs dimensionality reduction clustering processing, which can more efficiently process high-dimensional physiological data and improve the efficiency and accuracy of feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is the working principle diagram of the postoperative complication warning system described in the present invention; Figure 2 is the working principle diagram of the feature recombination module; Figure 3 is the design diagram of the compensation decision matrix generation method; Figure 4 is the design diagram of the risk adaptive layer modeling process; Figure 5 is the working principle diagram of the complication trigger layer. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0025] Please refer to Figures 1 - 5 , a postoperative complication warning system for thoracic surgery involved in the present invention, the system includes: a screening module and a warning module. The specific steps are as follows: The screening module performs multi-dimensional feature integration processing on the initial physiological data. The initial physiological data includes an analysis data set corresponding to vital sign feature data, an associated feature group corresponding to medical record feature data, and an evaluation index set corresponding to postoperative monitoring feature data. The vital sign feature data includes a grouping label generated by a heart rate variability parameter and a clustering label generated by a blood pressure fluctuation parameter.

[0026] The early warning module performs dynamic noise filtering on the real-time physiological data stream and inputs it into the multi-source fusion layer for abnormal pattern recognition, and generates a complication early warning instruction according to the output result of the multi-source fusion layer. The multi-source fusion layer includes a data normalization module and a feature recombination module. The data normalization module is used to segment the original physiological data stream by time window and eliminate redundant features. The feature recombination module is obtained by joint training based on multi-scenario historical medical data and real-time physiological trajectories, and includes a physiological parameter analysis layer, a risk adaptation layer, and a complication trigger layer connected in sequence.

[0027] Embodiment 1:

[0028] In this embodiment, the physiological parameter analysis layer, the risk adaptation layer, and the complication trigger layer work together in sequence when the system runs. The physiological parameter analysis layer performs spatio-temporal correlation processing on different analysis data sets in the original physiological data stream. Specifically, for each piece of data in the analysis data set, correlation analysis is carried out from two dimensions: time series and spatial distribution. For example, for the group labels generated by the heart rate variability parameters and the clustering labels generated by the blood pressure fluctuation parameters in the vital sign feature data, it is necessary to analyze the numerical changes of these parameters at different time points and the correlations in space, such as the mutual influences between different monitoring positions or different physiological systems. Through this analysis, the internal correlation relationships between different data in terms of time sequence and spatial position are identified, and then these spatio-temporally correlated data are integrated to finally generate physiological coupling feature data.

[0029] The risk adaptation layer models the dynamic correlation relationships between the physiological coupling feature data corresponding to each analysis data set. The physiological segmentation algorithm is used to process the physiological coupling feature data to identify the key nodes of complications. These key nodes refer to the important data points or data segments in the physiological data that can reflect the possible occurrence of complications. Each complication key node corresponds to a specific physiological scenario type, such as a specific postoperative recovery stage scenario corresponding to a certain heart rate variability pattern, or a potential complication risk scenario corresponding to a certain type of blood pressure fluctuation feature. Based on the physiological scenario type corresponding to each complication key node, a decision compensation feature group corresponding to each analysis data set can be determined. The decision compensation feature group is a set of features extracted from the analysis data set related to specific physiological scenarios and complication key nodes, which is used for subsequent compensation decision calculation.

[0030] Calculate the deviation coefficient between physiological nodes in the same scenario of the decision compensation feature groups corresponding to any two analysis datasets. During the calculation process, it is possible that the number of physiological nodes in the decision compensation feature groups corresponding to any two analysis datasets is inconsistent. When this occurs, virtual physiological point interpolation needs to be performed based on the scenario parameters corresponding to the terminal physiological nodes in the one with fewer physiological nodes. For example, assume that the decision compensation feature group of analysis dataset A has 10 physiological nodes, and the corresponding group of analysis dataset B has 8 physiological nodes, and they both contain the same "blood pressure monitoring scenario 24 hours after surgery". At this time, using the scenario parameters corresponding to the terminal physiological node (i.e., the 8th node) in analysis dataset B, such as the blood pressure value and monitoring time of this node, as a reference, 2 virtual physiological nodes are added after the corresponding position in analysis dataset A to make the number of physiological nodes in the two feature groups consistent. After completing the interpolation, calculate the deviation coefficient between the physiological nodes in the same scenario. The calculation of the deviation coefficient needs to consider multiple dimensions of factors, such as the numerical difference of physiological nodes, like the difference in heart rate values or blood pressure values between two nodes; the time difference, that is, the interval between two nodes on the time axis; and the spatial correlation difference, such as the difference in the impact of different monitoring positions on physiological indicators, etc. By comprehensively considering these factors, a value that can reflect the deviation degree between two physiological nodes is obtained. Based on this deviation coefficient, compensation decision matrix data between any two analysis datasets is generated. The compensation decision matrix data is a matrix structure used to represent the association relationship and compensation strategy between different analysis datasets. Each element in the matrix corresponds to the deviation coefficient between two analysis datasets and the corresponding compensation decision information.

[0031] The complication trigger layer performs multi-dimensional weight fusion based on the compensation decision matrix data and the physiological coupling feature data. The multi-dimensional weight fusion process needs to consider the association relationship and deviation degree between different analysis datasets reflected in the compensation decision matrix data, as well as the spatio-temporal association information contained in the physiological coupling feature data. For each relationship between a group of analysis datasets in the compensation decision matrix data and each feature dimension in the physiological coupling feature data, corresponding weights need to be assigned. The assignment of weights needs to comprehensively consider factors such as the importance, reliability of the data, and its association degree with complications. For example, for physiological feature data that is closely associated with common complications, its weight will be increased accordingly; while for data with lower reliability, the weight will be appropriately reduced. By performing fusion calculations on these multi-dimensional weights, a complication warning instruction is finally generated. This warning instruction contains an assessment of the patient's current physiological state and judgment information on whether there is a risk of complications. The system will take corresponding measures according to the level of the warning instruction, such as issuing an alarm or prompting medical staff to conduct further examinations and treatments.

[0032] Example 2:

[0033] In this embodiment, the data normalization module performs the normalization processing and calibration operations on the initial physiological data, specifically including processes such as equal-frequency band division, dynamic feature clustering, fixed sliding window optimization of the analysis data set, associated feature group, and evaluation index set, as well as calibration processing based on the physiological offset coefficient.

[0034] The data normalization module performs equal-frequency band division on the analysis data set, associated feature group, and evaluation index set according to a preset time window. The preset time window can be set according to the regular cycle of clinical monitoring. For example, a time window of 10 minutes, 30 minutes, or 1 hour can be used. For the analysis data set, which contains various parameters corresponding to vital sign feature data, such as the grouping labels generated by heart rate variability parameters and the clustering labels generated by blood pressure fluctuation parameters, it is divided into multiple equal-length frequency bands according to the preset time window, and the data within each frequency band constitutes the normalized physiological data. Taking the heart rate variability parameters as an example, if the preset time window is 30 minutes, then the heart rate variability data within every 30 minutes will be grouped into a frequency band to form the normalized physiological data for that time period. Similarly, the associated feature group corresponding to the medical record feature data, such as the surgical type, anesthesia method, medication record, etc., is processed, and after being divided according to the same time window, normalized medical data is generated; for the evaluation index set corresponding to the postoperative monitoring feature data, such as blood oxygen saturation, body temperature, respiratory rate, etc., equal-frequency band division is performed to generate normalized monitoring data.

[0035] After completing the equal-frequency band division, a dynamic feature clustering method is used to perform real-time alignment on the normalized physiological data and the normalized medical data. The first step of the dynamic feature clustering method is to construct a dynamic similarity matrix based on the physiological distribution characteristics of real-time scenario parameters. Real-time scenario parameters include the patient's current vital sign parameters (such as the immediate heart rate and blood pressure values), postoperative recovery stage (such as 12 hours, 24 hours after surgery, etc.), medication status, etc. The physiological distribution characteristics are the distribution laws of these parameters within the current time period, such as the fluctuation range of the heart rate, the mean and standard deviation of the blood pressure, etc. By analyzing the physiological distribution characteristics of these real-time scenario parameters, the similarity degree between different data points is calculated, and a dynamically changing similarity matrix is constructed. The element values in this matrix represent the similarity between the corresponding data points, and the larger the value, the higher the similarity.

[0036] Dimensionality reduction clustering is performed on the normalized physiological data using feature space mapping. Since the normalized physiological data may contain features in multiple dimensions (such as heart rate, blood pressure, blood oxygen, etc.), directly processing high-dimensional data will increase the computational complexity. Through feature space mapping, the high-dimensional physiological data is mapped into a low-dimensional space while retaining the main features and differences of the data. For example, methods such as principal component analysis (PCA) or linear discriminant analysis (LDA) are used for dimensionality reduction, converting multi-dimensional data into two-dimensional or three-dimensional data for easy clustering analysis. In the reduced-dimensional space, the data is clustered according to the dynamic similarity matrix, and similar data points are grouped into one category to achieve real-time alignment of the normalized physiological data. For example, the heart rate and blood pressure data of the same category are associated so that they correspond consistently in the time series for subsequent analysis.

[0037] For the normalized monitoring data, a fixed sliding window mechanism is used for steady-state optimization. The fixed sliding window has a specific window size and sliding step. For example, the window size is set to 15 minutes and the sliding step is set to 5 minutes. As time goes by, the window slides on the time axis of the monitoring data, and the data within the window is analyzed for steady state each time it slides. By calculating statistics such as the mean and variance of the data within the window, it is judged whether the data is in a stable state. If the data fluctuates greatly and exceeds the preset threshold, it is considered that there is noise or abnormality in the monitoring data during this time period, and optimization processing is required, such as using smoothing filtering and other methods to correct the data to improve the stability and reliability of the data. After the above processing, the first calibrated feature group, the second calibrated feature group, and the third calibrated feature group are output. Among them, the first calibrated feature group contains the calibrated grouping label and the calibrated clustering label, that is, the calibrated data obtained after the normalization and clustering alignment processing of the original heart rate variability grouping label and blood pressure fluctuation clustering label; the second calibrated feature group is the calibrated medical data; the third calibrated feature group is the calibrated monitoring data.

[0038] In addition, the data normalization module also performs the following operations: calculating the physiological offset coefficient between the calibrated grouping label and the calibrated clustering label in the historical monitoring period. The historical monitoring period can be multiple consecutive time periods after surgery, such as each hour within the first 24 hours after surgery as a historical monitoring period. The physiological offset coefficient is used to measure the degree of change difference between the grouping label and the clustering label in the historical period, and the deviation of both in terms of numerical value and time series needs to be considered when calculating. For example, for each historical monitoring period, calculate the difference between the mean of the calibrated grouping label and the mean of the calibrated clustering label within this period, and then perform weighted processing in combination with time factors (such as the time point of this period after surgery) to obtain the physiological offset coefficient of this period.

[0039] Predict the expected distribution value of the calibrated cluster label in the real-time monitoring period based on the physiological offset coefficient and the scenario parameters of the calibrated group label in the real-time monitoring period. The real-time monitoring period is the time period currently under monitoring, and the scenario parameters include the time points within this period (such as the 25th hour after surgery), the current status of the patient (such as whether active, medication situation, etc.). Use the physiological offset coefficient of the historical monitoring period to establish a relationship model between the group label and the cluster label, and then combine the scenario parameters of the real-time monitoring period and substitute them into the model to calculate the expected distribution range and possible numerical trend of the calibrated cluster label within this period.

[0040] Generate target decision compensation data based on the calibrated cluster label and its expected distribution value. Compare the actual calibrated cluster label with the expected distribution value, analyze the differences between the two. If there is a deviation, generate corresponding compensation data according to the degree and direction of the deviation. For example, if the actual value of the calibrated cluster label is higher than the upper limit of the expected distribution value, generate a positive compensation value to adjust the subsequent data analysis model; if it is lower than the lower limit, generate a negative compensation value. Finally, use the analysis data set corresponding to the target decision compensation data as the first calibration feature group, so that the first calibration feature group not only includes the calibrated group label and cluster label, but also incorporates decision compensation information, providing more accurate data support for subsequent early warning analysis.

[0041] Example 3:

[0042] In this example, the risk adaptive layer is composed of a pattern traceability unit, a scenario matching unit, and a noise suppression unit working together to achieve dynamic association modeling and noise processing of physiological coupling feature data. The pattern traceability unit performs physiological pattern traceability on each analysis data set in the physiological coupling feature data to extract the corresponding physiological propagation chain from each analysis data set. For the analysis data set containing vital sign feature data, such as the data set where the group label generated by the heart rate variability parameter is located, it is necessary to parse the change trajectory of its physiological indicators according to the time series. Taking the heart rate variability parameter as an example, starting from the initial postoperative moment, record the change of its group label at each time point, such as the evolution process from the "normal fluctuation group" to the "mild abnormality group" and then to the "high-risk group", and at the same time combine the change of the cluster label generated by the blood pressure fluctuation parameter to analyze the mutual influence path between different physiological indicators, so as to form a propagation chain reflecting the evolution of the physiological state. This propagation chain covers the physiological feature changes at each time node and the correlation between indicators. For example, at the 12th hour after surgery, when the group label of the heart rate variability changes, the cluster label of the blood pressure shows a corresponding fluctuation, and the two form a node record of a causal or concomitant relationship in the propagation chain.

[0043] The scene matching unit performs spatio-temporal superposition on the physiological propagation chains extracted from each analysis data set and the corresponding physiological coupling feature data to generate compensation decision matrix data. The spatio-temporal superposition process needs to perform matching synchronously in the time dimension and the space dimension. In the time dimension, align each time node in the physiological propagation chain with the time stamps in the physiological coupling feature data to ensure that the propagation chain features and coupling features at the same time point correspond; in the space dimension, consider the monitoring positions or physiological systems corresponding to different physiological indicators, such as the cardiovascular system corresponding to heart rate, the respiratory system corresponding to blood oxygen, etc., and associate the changes of each indicator in the propagation chain with the features at the corresponding spatial positions in the coupling feature data. For example, the node with an abnormal increase in heart rate 24 hours after surgery in the physiological propagation chain is superimposed with the coupling features related to the cardiovascular system (such as the spatio-temporal correlation features of heart rate and blood pressure) in the physiological coupling feature data at the same time point to form a composite feature point containing time and space information. By processing all the composite feature points, compensation decision matrix data is generated. The rows and columns of this matrix correspond to different analysis data sets respectively, and the matrix elements reflect the correlation degree and compensation strategy parameters after spatio-temporal superposition of different data sets.

[0044] The noise suppression unit performs time lag effect elimination processing on the compensation decision matrix data. The time lag effect refers to the fact that the mutual influence between physiological indicators may have a time delay. For example, the change in blood pressure after medication may lag behind the change in heart rate. If this delay is not processed, it will interfere with the accuracy of the compensation decision matrix. To eliminate this effect, a time lag factor is introduced, and its calculation formula is:

[0045] where represents the number of time nodes used for calculation, is the time point of the actual change of a certain physiological indicator, is the time point at which another physiological indicator associated with it should theoretically change, is the weight of each time node, and the weight can be determined according to clinical experience or historical data. For example, higher weights are assigned to the change nodes of indicators closely related to common complications. By calculating the time lag factor , the elements in the compensation decision matrix are adjusted. Specifically, for the elements in the matrix reflecting the correlation relationship between two indicators, if there is a time lag, the correlation intensity is corrected according to the lag factor to make the matrix data more accurately reflect the real-time correlation state of physiological indicators.

[0046] In specific operations, for the physiological pattern tracing of each analysis dataset, all key features of the dataset need to be covered. Taking the associated feature group corresponding to the medical record features as an example, it is necessary to extract the time series changes of data such as the type of surgery, anesthesia method, and medication records, such as the duration of surgery, the injection time and dose changes of anesthetic drugs, the types and time nodes of postoperative medications, etc., to form a physiological propagation chain related to medical records, which reflects the impact path of medical intervention measures on the patient's physiological state. When the scenario matching unit processes the medical record propagation chain and physiological coupling feature data, it is necessary to associate the time nodes of medical intervention with the time nodes of vital sign changes. For example, the time nodes of anesthetic drug injection are spatially and temporally superimposed with the time nodes of postoperative heart rate and blood pressure changes to analyze the associated features between medical intervention and physiological response, and generate the corresponding compensation decision matrix elements.

[0047] When the noise suppression unit processes the time lag effect, different weights need to be set for different types of physiological indicators and medical intervention measures . For example, for the association between heart rate and blood pressure, since their physiological correlation is strong, the weight can be set to 0.8; while for the association between body temperature and blood oxygen saturation, the physiological correlation is relatively weak, and the weight can be set to 0.5. The time lag factor is obtained through weighted calculation . Then, the time dimension parameters of each element in the compensation decision matrix are adjusted so that the matrix can accurately reflect the immediate correlation between physiological indicators and avoid misjudgment caused by time lag.

[0048] After the sequential processing of the pattern tracing unit, the scenario matching unit, and the noise suppression unit, the risk adaptive layer can generate compensation decision matrix data that accurately reflects the dynamic correlation between each analysis dataset, providing reliable input data for the multi-dimensional weight fusion of the subsequent complication trigger layer, and ensuring the accuracy and timeliness of the complication warning instructions.

[0049] Example 4: In this embodiment, the complication trigger layer generates a complication warning instruction based on the compensation decision matrix data and the physiological coupling feature data through the collaborative operation of the multi-dimensional weight collaborative unit, the dynamic threshold optimization unit, and the abnormal pattern recognition unit. The multi-dimensional weight collaborative unit includes multiple warning decision nodes, and each warning decision node is connected to each analysis data set in the compensation decision matrix data and the physiological coupling feature data through parameter configuration. Taking the analysis data sets where the group labels generated by the heart rate variability parameters and the clustering labels generated by the blood pressure fluctuation parameters are located as an example, the warning decision node needs to receive the data reflecting the correlation degree between these two data sets in the compensation decision matrix, and at the same time obtain the spatio-temporal correlation features between the two in the physiological coupling feature data. For example, when the compensation decision matrix shows that the deviation coefficient between the heart rate group label and the blood pressure clustering label within 24 hours after surgery reaches 0.6, the corresponding warning decision node will, according to the preset parameter configuration, assign weights to the features of these two data sets. For example, the heart rate feature weight is set to 0.4, and the blood pressure feature weight is set to 0.6, and then fuse and process these two types of feature data.

[0050] The dynamic threshold optimization unit iteratively optimizes the parameter configuration through the dynamic threshold adjustment algorithm to minimize the error between the complication warning instruction and the actual physiological distribution. At the initial stage of system operation, the preset parameter configuration may deviate from the actual physiological state of the patient. For example, initially it is set that when the heart rate group label enters the "high-risk group" and the blood pressure clustering label fluctuates beyond the preset range, a warning is triggered. However, some patients may not develop complications under similar index changes due to individual differences. At this time, the dynamic threshold optimization unit will collect the actual physiological distribution data, such as the heart rate, blood pressure indicators and the occurrence of complications of all patients within 72 hours after surgery, and compare the warning instruction with the actual occurrence of complications. If it is found that the warning is too frequent (such as multiple warnings are triggered when no complications occur), the parameter configuration is adjusted to increase the trigger threshold of the heart rate or blood pressure indicators; if the warning is lagging, the threshold is lowered, and through multiple iterations of optimization, the parameter configuration is made more suitable for the actual physiological distribution of the patient group.

[0051] The abnormal pattern recognition unit locates high-risk scenarios based on the compensation decision matrix data and the physiological coupling feature data, and generates a complication warning instruction. Specifically, this unit will analyze the deviation coefficients between the analysis data sets in the compensation decision matrix, and combine the spatio-temporal correlation features in the physiological coupling feature data to identify abnormal patterns that may cause complications. For example, when the compensation decision matrix shows that the deviation coefficient between the heart rate group label and the blood oxygen saturation evaluation index set suddenly rises to 0.7 within 48 hours after surgery, and at the same time the physiological coupling feature data shows that the spatio-temporal correlation features between the heart rate and the blood oxygen in this period are abnormal (such as the heart rate continuously increases while the blood oxygen saturation continuously decreases), the abnormal pattern recognition unit will determine this as a high-risk scenario and generate a corresponding complication warning instruction to prompt medical staff to pay attention to the possible respiratory and circulatory system complications of the patient.

[0052] In the specific operation process, the warning decision nodes of the multi-weight collaborative unit need to cover different types of physiological indicators and medical data. Taking the associated feature group corresponding to the medical record feature data as an example, when the surgical type is radical resection of lung cancer and a certain vasoactive drug appears in the postoperative medication record, the corresponding warning decision node will receive the associated data between the medical record data set and the vital sign data set in the compensation decision matrix, and at the same time obtain the spatio-temporal association features between the surgical type, medication time and heart rate and blood pressure changes in the physiological coupling feature data. Through weight fusion processing, it is judged whether the drug use has an impact on the patient's vital signs.

[0053] The iterative optimization process of the dynamic threshold optimization unit needs to continuously collect multiple groups of patient data. For example, collect the physiological data and the occurrence of complications of 100 postoperative patients in the thoracic surgery department, compare each warning instruction with the actual occurrence of complications, and calculate the error value. If the error value exceeds the preset range (such as the warning accuracy rate is lower than 70%), the parameter configuration is adjusted, such as changing the weight ratio of the heart rate grouping label and the blood pressure clustering label, or modifying the index combination conditions for triggering the warning, and then process the data with the new parameter configuration until the error value is reduced to an acceptable range, making the warning instruction more in line with the actual physiological distribution.

[0054] When the abnormal pattern recognition unit locates high-risk scenarios, it needs to synthesize features from multiple dimensions. In addition to vital sign data, it also needs to consider the evaluation index set corresponding to the postoperative monitoring feature data, such as body temperature, respiratory rate, etc. For example, when the compensation decision matrix shows that the deviation coefficient of the association feature group between the body temperature evaluation index set and the white blood cell count increases, and the physiological coupling feature data shows that the body temperature continuously exceeds 38.5°C and the white blood cell count exceeds the normal range, the abnormal pattern recognition unit will determine this as a high-risk infection scenario and generate a warning instruction for infection complications, reminding medical staff to conduct infection-related examinations.

[0055] Through the weight fusion of multi-source data by the multi-weight collaborative unit, the iterative adjustment of parameters by the dynamic threshold optimization unit, and the accurate positioning of high-risk scenarios by the abnormal pattern recognition unit, the complication trigger layer can generate accurate complication warning instructions based on the patient's real-time physiological data and historical medical data, providing data support for the prevention and timely intervention of complications in postoperative thoracic surgery patients.

[0056] Example 5: In this embodiment, the dynamic feature clustering method is used to perform real-time alignment processing on normalized physiological data and medical data. Its core lies in constructing a dynamic similarity matrix based on real-time scenario parameters and achieving dimensionality reduction clustering through feature space mapping. Specifically, when implementing, it is first necessary to clarify the scope of real-time scenario parameters. For example, the real-time heart rate, blood pressure fluctuation values, postoperative recovery duration, medication conditions (such as whether vasoactive drugs are used) of patients after thoracic surgery, etc. These parameters change dynamically over time and directly affect the distribution characteristics of physiological data.

[0057] Taking a patient within 24 hours after surgery as an example, assume that at 12 hours after surgery for a certain patient, the grouping label generated by the heart rate variability parameter is "moderate abnormality group", and the clustering label generated by the blood pressure fluctuation parameter is "high blood pressure fluctuation group". At the same time, their blood oxygen saturation is 95% and body temperature is 37.2°C. These data constitute the current real-time scenario parameters. The first step of the dynamic feature clustering method is to construct a dynamic similarity matrix based on the physiological distribution characteristics of these parameters. In this process, it is necessary to calculate the similarity degree between different data points. For example, compare the heart rate and blood pressure data of this patient at 12 hours after surgery with the data at 6 hours and 18 hours after surgery, analyze the change trends of the heart rate variability grouping label and blood pressure clustering label, and the correlation degree of indicators such as blood oxygen and body temperature. The calculation of similarity needs to comprehensively consider numerical differences (such as the difference in heart rate values), time intervals (such as the time difference between 6 hours and 12 hours), and indicator correlations (such as the co-variation of heart rate and blood pressure). By setting corresponding weights to quantify the influence of different dimensions, a dynamically updated similarity matrix can be constructed. The element values in this matrix reflect the similarity of physiological data at different time points.

[0058] After completing the construction of the dynamic similarity matrix, dimensionality reduction clustering processing is performed on the normalized physiological data using feature space mapping. The normalized physiological data may contain multiple dimensions, such as heart rate, blood pressure, blood oxygen, body temperature, etc. Each dimension has different characteristic parameters (such as heart rate variability parameters, blood pressure fluctuation parameters, etc.). Directly processing high-dimensional data will increase the computational complexity and may introduce redundant information. The purpose of feature space mapping is to map high-dimensional data to a low-dimensional space while retaining the key features of the data. For example, for the physiological data of the above patient, there may be 10 feature dimensions in the high-dimensional space (heart rate, heart rate variability grouping label, systolic blood pressure, diastolic blood pressure, blood pressure clustering label, blood oxygen saturation, body temperature, respiratory rate, postoperative time, medication type). Through feature space mapping (such as methods like principal component analysis or linear discriminant analysis), it can be mapped to a 3D or 4D low-dimensional space, highlighting the features that have a greater impact on complication warning (such as heart rate variability grouping label, blood pressure clustering label, postoperative time, etc.) and weakening the secondary features (such as small fluctuations in body temperature within the normal range).

[0059] In the reduced low-dimensional space, data clustering is performed based on the dynamic similarity matrix. For example, the physiological data at 12 hours after surgery is grouped with the data at 6 hours and 18 hours after surgery with relatively high similarity in the similarity matrix to form a clustering cluster. The data within this clustering cluster has a similar distribution of physiological characteristics, such as being in the "moderately abnormal" heart rate variability grouping and the "high blood pressure fluctuation" blood pressure clustering label, and the postoperative times are close. Through this clustering process, real-time alignment of the standardized physiological data can be achieved, enabling the data within the same clustering cluster to form a continuous characteristic change trajectory in the time series, facilitating subsequent analysis of the evolution trend of physiological indicators.

[0060] Another example is used to further illustrate this process: At 8 hours after surgery for a certain patient, the heart rate variability grouping label is the "normal group" and the blood pressure clustering label is the "normal fluctuation group". However, at 10 hours after surgery, due to pain stimulation, the heart rate suddenly increases, the grouping label changes to the "mildly abnormal group", and the blood pressure clustering label changes to the "medium blood pressure fluctuation group". At the same time, the blood oxygen saturation drops from 97% to 94%. At this time, the real-time scene parameters change significantly, and the dynamic similarity matrix will recalculate the similarity of the data at each time point, reducing the similarity between the data at 10 hours after surgery and the data at 8 hours after surgery, while increasing the similarity with the data at other time points where the heart rate and blood pressure fluctuate due to pain stimulation. The feature space mapping will prominently display the features corresponding to the factor of pain stimulation (such as sudden increase in heart rate and blood pressure fluctuation) in the low-dimensional space, enabling the new clustering result to accurately reflect the changes in physiological indicators caused by pain in the patient and avoiding misjudging it as a precursor to complications.

[0061] When processing the associated feature groups corresponding to the medical record feature data, the dynamic feature clustering method is equally applicable. For example, for patients undergoing "lobectomy", medical data such as their postoperative medication records (such as the time and dosage of analgesic drugs used) and anesthesia methods need to be aligned in real time with the vital sign feature data. When constructing the dynamic similarity matrix, the degree of association between the surgical type, anesthetic drugs, and vital sign changes needs to be considered, such as the common heart rate and blood pressure responses of lobectomy patients after using a certain type of analgesic drug. Data with similar medical interventions and physiological responses are grouped together. The feature space mapping will highlight key features such as the surgical type and medication time, and after dimensionality reduction, clustering is performed to enable the medical record data and the vital sign data to form a corresponding relationship in the time series, facilitating the analysis of the impact of medical interventions on the patient's physiological state.

[0062] The real-time nature of the dynamic feature clustering method is reflected in that as the patient's physiological data is updated, the dynamic similarity matrix is continuously recalculated, and the parameters of the feature space mapping are also adjusted accordingly to ensure that the clustering results can timely reflect the patient's current state. For example, when the patient is transferred from the postoperative intensive care unit to the general ward and the increase in activity level causes physiological fluctuations in heart rate and blood pressure, the dynamic similarity matrix can identify the differences between these fluctuations and those caused by complications, adjust the similarity calculation results, so that the clustering results can accurately distinguish physiological changes from pathological changes and avoid false alarms.

[0063] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0064] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A postoperative complication warning system for thoracic surgery, characterized in that, Comprising: A screening module for performing multi-dimensional feature integration processing on initial physiological data, where the initial physiological data includes an analysis data set corresponding to vital sign feature data, an associated feature group corresponding to medical record feature data, and an evaluation index set corresponding to postoperative monitoring feature data, and the vital sign feature data includes a grouping label generated from heart rate variability parameters and a clustering label generated from blood pressure fluctuation parameters; An early warning module for performing dynamic noise filtering processing on real-time physiological data streams and inputting them into a multi-source fusion layer for abnormal pattern recognition, and generating a complication early warning instruction according to the output result of the multi-source fusion layer; The multi-source fusion layer includes a data normalization module and a feature recombination module. Among them, the data normalization module is used for time window segmentation and redundant feature elimination of the original physiological data stream, and the feature recombination module is obtained by joint training based on multi-scenario historical medical data and real-time physiological trajectories; The feature recombination module includes a physiological parameter analysis layer, a risk adaptation layer, and a complication trigger layer connected in sequence.

2. The postoperative complication warning system for thoracic surgery according to claim 1, wherein The physiological parameter analysis layer is used for spatio-temporal correlation processing of different analysis data sets in the original physiological data stream to generate physiological coupling feature data; the risk adaptation layer is used for modeling the dynamic association relationship between the physiological coupling feature data corresponding to each analysis data set to generate compensation decision matrix data; the complication trigger layer is used for multi-dimensional weight fusion based on the compensation decision matrix data and the physiological coupling feature data to generate a complication early warning instruction.

3. The postoperative complication warning system for thoracic surgery according to claim 2, wherein The modeling of the dynamic association relationship between the physiological coupling feature data corresponding to each analysis data set to generate compensation decision matrix data includes: Using a physiological segmentation algorithm to identify the complication key nodes in the physiological coupling feature data, and determining the decision compensation feature groups corresponding to each analysis data set based on the physiological scenario type corresponding to each complication key node; Calculating the deviation coefficient between the physiological nodes in the same scenario of the decision compensation feature groups corresponding to any two analysis data sets, and generating the compensation decision matrix data between the any two analysis data sets based on the deviation coefficient.

4. The postoperative complication warning system for thoracic surgery according to claim 3, characterized in that, The calculation of the deviation coefficient between the physiological nodes in the same scenario of the decision compensation feature groups corresponding to any two analysis data sets includes: When the number of psychological nodes in the decision compensation feature groups corresponding to the any two analysis data sets is inconsistent, performing virtual physiological point interpolation based on the scenario parameters of the terminal physiological nodes in the one with the fewer physiological nodes, and calculating the deviation coefficient between the physiological nodes in the same scenario based on the interpolated data.

5. The postoperative complication warning system for thoracic surgery according to claim 1, characterized in that, The data normalization module is specifically used for: Performing equal-frequency band division on the analysis data set, the associated feature group, and the evaluation index set according to a preset time window to generate normalized physiological data, normalized medical data, and normalized monitoring data; The normalized physiological data and normalized medical data are aligned in real time using a dynamic feature clustering method, and the normalized monitoring data is optimized for steady state using a fixed sliding window mechanism, and a first calibration feature group, a second calibration feature group, and a third calibration feature group are output; wherein, the first calibration feature group includes a calibrated grouping label and a calibrated clustering label.

6. The postoperative complication warning system for thoracic surgery according to claim 5, wherein The data normalization module is further configured to: calculate a physiological offset coefficient between the calibrated grouping label and the calibrated clustering label in a historical monitoring period; predict an expected distribution value of the calibrated clustering label in the real-time monitoring period according to the physiological offset coefficient and the scenario parameters of the calibrated grouping label in the real-time monitoring period; generate target decision compensation data based on the calibrated clustering label and its expected distribution value, and use the analysis data set corresponding to the target decision compensation data as the first calibration feature group.

7. The postoperative complication warning system for thoracic surgery according to claim 2, wherein The risk adaptive layer specifically includes: a pattern tracing unit, configured to perform physiological pattern tracing on each analysis data set in the physiological coupling feature data respectively, so as to extract a corresponding physiological propagation chain from each analysis data set; a scenario matching unit, configured to perform spatio-temporal superposition on the physiological propagation chain extracted from each analysis data set and the corresponding physiological coupling feature data to generate compensation decision matrix data.

8. The postoperative complication warning system for thoracic surgery according to claim 7, wherein The risk adaptive layer further includes: a noise suppression unit, configured to perform time lag effect elimination processing on the compensation decision matrix data.

9. The postoperative complication warning system for thoracic surgery according to claim 2, characterized in that, The complication trigger layer specifically includes: a multi-dimensional weight collaboration unit, including a plurality of early warning decision nodes, and each early warning decision node is connected to each analysis data set in the compensation decision matrix data and the physiological coupling feature data through parameter configuration; a dynamic threshold optimization unit, configured to iteratively optimize the parameter configuration through a dynamic threshold adjustment algorithm to minimize the error between the complication early warning instruction and the actual physiological distribution; an abnormal pattern recognition unit, configured to perform high-risk scenario positioning based on the compensation decision matrix data and the physiological coupling feature data to generate a complication early warning instruction.

10. The postoperative complication warning system for thoracic surgery according to claim 5, wherein The dynamic feature clustering method specifically includes: constructing a dynamic similarity matrix based on the physiological distribution characteristics of real-time scenario parameters; performing dimensionality reduction clustering processing on the normalized physiological data using feature space mapping.

Images