Intelligent processing method and system for cross-disease critical illness monitoring data

Abstract

The application discloses an intelligent processing method and system for cross-disease critical illness monitoring data, relates to the field of cross-disease data processing, and constructs multiple-source monitoring data into a disease perception data matrix under the constraint of a disease label, and further introduces a trend decomposition and false probability modeling mechanism, so that an abnormality recognition process originally based on a numerical threshold is converted into a trend, fluctuation and semantic joint discrimination process, so that when an index short-term fluctuation and a long-term trend deviate, the deviation behavior can be probabilistically expressed and classified, the recognition path of complex states such as surface improvement but trend deterioration or surface deterioration but trend recovery is reconstructed, and specifically, a real trend and a disturbance signal are separated by hierarchical decomposition, and then the abnormality is promoted from whether abnormal to abnormal credibility by using deviation modeling, so that the problem of false improvement and false deterioration in the cross-disease data which is difficult to identify is structurally corrected.

Description

Technical Field

[0001] This invention relates to the field of cross-disease data processing, specifically to an intelligent processing method and system for cross-disease critical illness monitoring data. Background Technology

[0002] In the process of cross-disease monitoring of acute and severe illnesses, due to frequent treatment interventions, complex changes in disease status, and nonlinear coupling relationships among multiple indicators, short-term fluctuations and long-term trends often appear in the monitoring data. For example, some indicators may show numerical improvement in a short period of time, but the overall trend is still on the deterioration track, or there may be a short-term abnormal increase but a stable long-term trend. Such phenomena are difficult to distinguish in existing methods based on thresholds or single trend analysis and are easily misidentified as real changes in disease status. At the same time, the clinical meaning of the same monitoring indicator varies significantly in different disease contexts. For example, an increase in a certain indicator may represent increased risk in one type of disease, while in another type of disease it may be a normal adjustment or treatment response. However, existing methods mostly use a uniform judgment logic and lack semantic differentiation mechanisms for disease differences, resulting in inconsistent and inaccurate abnormal identification results.

[0003] The aforementioned problems overlap in actual operation. On the one hand, the inability to distinguish between false improvement or false deterioration and real changes in the condition can easily lead to misjudgment of the status during continuous monitoring, causing fluctuations and deviations in risk assessment results over time. On the other hand, without the introduction of disease-specific semantic constraints, the mixed use of judgment criteria for similar indicators across different diseases leads to the incorrect construction of correlations between multiple indicators, further affecting subsequent risk integration and judgment processes. Based on this, when risk aggregation or intervention decisions are made based on the aforementioned inaccurate anomaly identification results, it is easy to encounter situations such as distorted risk assessment and amplification or masking of the contributions of key indicators, thereby reducing the reliability and consistency of the overall monitoring and analysis results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent processing method and system for cross-disease critical illness monitoring data, solving the problems mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Firstly, an intelligent processing method for cross-disease critical illness monitoring data includes: Physiological data of patients are acquired from different monitoring devices, and preliminary abnormality screening is performed. Indicator mapping is performed in combination with disease labels to construct a disease perception data matrix. Based on the disease perception data matrix, a set of trend vectors is obtained by performing hierarchical decomposition on the time series of each indicator, and a false probability vector and cross-indicator correlation matrix are generated through short-term fluctuation and long-term trend deviation analysis. By performing threshold dynamic modeling through cross-indicator correlation matrix, the initial anomaly screening markers are corrected and a multi-dimensional risk indicator matrix is ​​constructed. Perform component expansion, direction unification, and time-series cumulative analysis on the multidimensional risk indicator matrix to generate an indicator contribution matrix and obtain a multidimensional risk score matrix after correlation constraint correction. By reading historical intervention records and combining them with a multidimensional risk scoring matrix, a risk contribution prediction sequence is constructed through intervention record matching and contribution change path deduction, and the optimal intervention is determined based on time stability analysis.

[0006] Secondly, an intelligent processing system for cross-disease critical illness monitoring data includes: The data acquisition and processing module is used to acquire patients' physiological data from different monitoring devices, perform preliminary abnormality screening, combine disease labels to map indicators, and construct a disease-aware data matrix. The data decomposition and analysis module is used to perform hierarchical decomposition of the time series of each indicator based on the disease perception data matrix to obtain a set of trend vectors, and to generate false probability vectors and cross-indicator correlation matrices through short-term fluctuation and long-term trend deviation analysis. The cross-disease dynamic anomaly identification module is used to perform threshold dynamic modeling through cross-indicator correlation matrix, correct the initial anomaly screening markers and construct a multi-dimensional risk indicator matrix; The cross-disease risk analysis module is used to perform component expansion, directional unification, and time-series cumulative analysis on the multidimensional risk indicator matrix, generate an indicator contribution matrix, and obtain a multidimensional risk score matrix after correlation constraint correction. The intervention prediction module is used to read historical intervention records, combine them with a multidimensional risk score matrix, construct a risk contribution prediction sequence by matching intervention records and extrapolating contribution change paths, and determine the optimal intervention based on time stability analysis.

[0007] The above-described solution of the present invention has at least the following beneficial effects: By constructing a disease-aware data matrix from multi-source monitoring data under disease-specific labels, and further introducing trend decomposition and false probability modeling mechanisms, the anomaly identification process, which was originally based solely on numerical thresholds, is transformed into a joint discrimination process of trend, fluctuation, and semantics. This enables the probabilistic expression and classification of deviations when short-term fluctuations and long-term trends of indicators deviate, thus reconstructing the identification path for complex states such as apparent improvement but trend deterioration or apparent deterioration but trend recovery. Specifically, by layered decomposition to separate the true trend from the disturbance signal, and then using deviation modeling to elevate anomalies from whether they are abnormal to their credibility, the structural correction is directly applied to address the difficulty in identifying false improvements and false deteriorations in cross-disease data.

[0008] By constructing a cross-indicator correlation matrix and incorporating it into the dynamic threshold modeling and anomaly labeling calibration process, the judgment logic of a single indicator is transformed from isolated judgment to multi-indicator coupled constraint judgment. On this basis, false probability is combined to suppress unstable signals, so that anomaly judgment not only depends on the changes of the indicator itself, but also on the synchronicity and consistency of its associated indicators. Specifically, structural constraints are constructed by utilizing the synergistic change relationship between indicators, and the differences in the meaning of the same indicator under different diseases are indirectly expressed through the correlation structure, thereby avoiding the problem of threshold misuse caused by disease differences and realizing adaptive correction of the risk of misjudgment of the same indicator with large differences in clinical meaning in different diseases.

[0009] By performing risk component unification, contribution decomposition, and correlation redundancy removal on a multidimensional risk indicator matrix, and further combining historical intervention records to construct an intervention impact matrix and risk evolution prediction sequence, risk assessment is extended from static results to a traceable and predictable dynamic decision-making link. Specifically, the contribution matrix first clarifies the source structure of each indicator for risk, then eliminates duplicate contributions through correlation constraints, and finally maps historical responses to future evolution through intervention path deduction. This forms a stable basis for intervention selection in complex multi-indicator coupling scenarios, avoids intervention decision deviations due to short-term fluctuations or local anomalies, and achieves effective control over the uncertainty of risk evolution in complex cross-disease monitoring environments. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an intelligent processing method for cross-disease-specific acute and critical illness monitoring data according to the present invention. Figure 2 This is a schematic diagram of the structure of an intelligent processing system for monitoring critical and severe illness data across different diseases, according to the present invention.

[0011] In the attached diagram, the components represented by each number are as follows: Data acquisition and processing module 11, data decomposition and analysis module 12, cross-disease dynamic anomaly identification module 13, cross-disease risk analysis module 14, intervention prediction module 15. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Example 1, such as Figure 1 As shown, this invention provides an intelligent processing method for cross-disease critical illness monitoring data, including: S100: Acquire patient physiological data from different monitoring devices, perform preliminary abnormality screening, combine disease labels to map indicators, and construct a disease perception data matrix; S200: Based on the disease perception data matrix, the time series of each indicator is decomposed into a set of trend vectors, and a false probability vector and cross-indicator correlation matrix are generated through short-term fluctuation and long-term trend deviation analysis. S300: Performs dynamic threshold modeling through cross-indicator correlation matrix, corrects preliminary anomaly screening markers and constructs a multi-dimensional risk indicator matrix; S400: Perform component expansion, direction unification and time series cumulative analysis on the multidimensional risk indicator matrix to generate the indicator contribution matrix and obtain the multidimensional risk score matrix after correlation constraint correction; S500: Read historical intervention records, combine them with a multidimensional risk score matrix, construct a risk contribution prediction sequence by matching intervention records and extrapolating contribution change paths, and determine the optimal intervention based on time stability analysis.

[0014] In this embodiment, the present invention introduces a progressive mechanism of disease perception, trend decomposition, false identification, correlation constraint, risk reconstruction, and intervention inference into the data processing flow, so that short-term fluctuations that were originally mixed in multi-source monitoring data can be separated from the real trend. On this basis, the judgment logic of the same type of indicators under different diseases is distinguished and modeled, thereby avoiding misjudging local fluctuations as real changes in the condition during the anomaly identification stage. Furthermore, by constructing a cross-indicator correlation matrix and participating in threshold dynamic modeling, each indicator is no longer judged in isolation, but rather a consistent correction is made based on the coupling relationship of multiple indicators, thereby forming a more stable risk expression structure. Building upon this foundation, by performing component expansion and contribution attribution on the risk vector, the overall risk is decomposed into interpretable indicator contributions. Furthermore, correlation constraints eliminate the amplification effect caused by duplicate contributions or synchronous shifts. Finally, during the intervention simulation phase, historical intervention records are used to predict and screen risk evolution paths, ensuring that intervention selection has a reliable temporal evolutionary logic. For example, in patients with septic shock, blood pressure may temporarily rise under the influence of vasopressors, but lactate levels may continue to increase. Traditional methods might interpret this as an improvement, but this method identifies the blood pressure rise as a short-term fluctuation and assigns a higher probability of false improvement through trend decomposition. Simultaneously, the correlation between lactate and blood pressure is used to correct for the risk, thus reflecting the unstable state of the condition in the overall risk score and prioritizing treatment pathways targeting inadequate perfusion in the intervention simulation.

[0015] Specifically: S100 includes: the patient's physiological data includes at least hemodynamic parameters, respiratory parameters, blood oxygen index, laboratory indicators and drug intervention information; and preliminary abnormality screening is performed on each indicator in the patient's physiological data, including threshold detection, missing data marking and abnormal fluctuation identification; Hemodynamic parameters refer to physiological indicators that reflect the state of the circulatory system, including systolic blood pressure, diastolic blood pressure, mean arterial pressure, heart rate, central venous pressure, cardiac output, etc. These parameters are used to reflect whether blood circulation is stable. Respiratory parameters are indicators that reflect the function of the respiratory system, including respiratory rate, tidal volume, end-expiratory carbon dioxide concentration, ventilator settings and trigger status, etc. Blood oxygenation indicators mainly include blood oxygen saturation, arterial blood oxygen partial pressure, and oxygenation index, which reflect the oxygenation level. Laboratory indicators include blood chemistry, blood gas analysis, lactate, blood glucose, creatinine, and white blood cell count, which are used to reflect the function and metabolic status of multiple organs.

[0016] Drug intervention information includes records of infusion volume, vasoactive drug dosage, antibiotic dosage, and time of analgesic and sedative drug use. This information is used to explain the relationship between changes in indicators and treatment intervention.

[0017] Specifically, for each indicator, it is compared with the historical range or reference physiological interval of the same disease. If the data exceeds the upper and lower limits, it is marked as abnormal. This is used to output the upper and lower limit marking status of each indicator, which is convenient for trend analysis and judgment of false features. If an indicator at a certain point in time is not collected or data is missing due to equipment failure, a missing marker is recorded, and time window information is generated to determine whether interpolation or resampling is needed in trend analysis.

[0018] Short-term fluctuation calculations are performed on the time series of indicators to identify data points with excessively large amplitudes or abnormal continuous fluctuations, and the fluctuation amplitude and direction characteristics are output for pseudo-improvement / deterioration analysis.

[0019] Based on the patient's previous medical records and admission diagnosis information, disease labels are mapped to each indicator to form a disease perception data matrix. The disease perception data matrix records the preliminary abnormal screening mark, time series and clinical semantic label of each indicator under the corresponding disease label. Clinical semantic labels refer to the meaning or explanatory labels of each monitoring indicator under a specific disease. They convert raw numerical values ​​into disease-aware semantic information for subsequent dynamic thresholding, trend analysis, and risk scoring.

[0020] Time series characteristics include the trend of indicator changes over time, short-term fluctuations, directionality, and periodicity, which are used to identify false improvements / deteriorations, perform trend analysis, and conduct cross-indicator correlation analysis. Clinical semantic labels refer to the meaning of each indicator in a specific disease. For example, elevated heart rate is a compensatory response in sepsis, but may indicate increased risk in myocardial infarction. They are used for dynamic threshold generation and risk scoring module interpretation. The specific acquisition method is as follows: Each patient's monitoring indicator data is constructed as a feature vector, including indicator value, time-series fluctuations, short-term trends, equipment source, and sampling time. Simultaneously, the patient's disease information is embedded into the feature vector, forming a disease-aware feature embedding matrix to ensure that each indicator contains disease-related contextual information. The output feature embedding matrix serves as input for subsequent model training or inference, automatically generating a clinical semantic representation for each indicator. Then, a multi-disease indicator embedding model is trained using self-supervised or contrastive learning methods. The model learns the latent semantics of the indicators by predicting changes in indicators at future time points, cross-indicator dependencies, and trend consistency. During training, the model does not require manual semantic annotation; instead, it extracts semantic features by maximizing the similarity of indicator embeddings within the same disease or similar clinical states and minimizing embedding differences between different diseases. The output latent semantic vector for each indicator reflects its clinical significance in a specific disease context and can be used to generate specific semantic labels. Finally, the latent semantic vector is mapped to interpretable semantic labels using clustering or classification algorithms, such as indicators indicating normal, warning, abnormal trends, or indicators showing increases, decreases, or stability. The generated semantic identifiers are combined with preliminary anomaly markers and time-series features to form a complete disease-aware data matrix, giving each indicator disease-specific clinical semantic information. The output disease-aware data matrix serves as input for dynamic threshold generation, risk scoring, and intervention analysis, while also supporting subsequent cyclical feedback and iterative optimization to achieve closed-loop processing.

[0021] In this embodiment, hemodynamic parameters, respiratory parameters, blood oxygen index, laboratory index and drug intervention information are uniformly accessed at the data source stage, and structured preliminary screening results are formed by combining over-threshold detection, missing marker and abnormal fluctuation identification. On this basis, disease labels are introduced to semantically map each indicator, so that the original values ​​not only have time series attributes, but also have the ability to interpret disease context. Furthermore, by constructing a disease-aware data matrix that includes time-series features and clinical semantic identifiers, the changes in indicators no longer depend solely on the numerical values ​​themselves, but are instead bound to their meaning within a specific disease, thus providing a unified input basis for subsequent trend decomposition, false positive identification, and risk modeling. Among them, clinical semantic labels are automatically generated through multi-disease feature embedding and comparative learning, so that the semantic differences of the same indicators under different diseases are internalized into computable features, thereby avoiding the misjudgment problem caused by uniform threshold.

[0022] Specifically: S200 includes: based on the disease perception data matrix, by performing hierarchical decomposition analysis on the time series of each indicator, a set of trend vectors containing long-term trend components, short-term fluctuation components and noise components is obtained, which is used as the input basis for subsequent false feature identification; In practice, the disease-sensing data matrix is ​​used as input, and the corresponding time series data is extracted for each indicator. Continuous sequence segments are constructed in units of time windows. Subsequently, within each time window, a sliding window segmentation process is performed on the indicator sequence to obtain multiple local sub-sequences. The mean sequence and residual sequence are calculated for each sub-sequence to form a preliminary smoothing decomposition result. Next, a weighted cumulative operation is performed on the smoothed sequence. By assigning different weights to different time scales, the low-frequency change part is extracted as the long-term trend component. At the same time, the difference between the original sequence and the long-term trend component is used as the short-term fluctuation component, and the remaining unexplained part is marked as a noise component. After the component splitting is completed, the long-term trend component, short-term fluctuation component, and noise component are re-aligned according to the time index, and combined with the existing preliminary anomaly markers, a ternary feature structure is constructed for each time point; then, the ternary structure is spliced ​​on the indicator dimension to form a set of trend vectors, where each vector contains a combination of trend, fluctuation, and noise features. The trend vector set is a set of vectors formed by hierarchically decomposing the time series of each indicator in the disease perception data matrix. For each indicator at each time point, a vector containing three types of components is constructed. The combination of all time points and the corresponding vectors of all indicators constitutes the trend vector set.

[0023] Based on the trend vector set, pattern decomposition analysis is performed on the deviation relationship between short-term fluctuations and long-term trends to obtain the false probability vector of each indicator within the corresponding time window, and this vector is used as the key input for anomaly correction and subsequent correlation analysis. The probability vectors for false improvement and false deterioration represent the probability sequence of each indicator exhibiting false improvement or false deterioration at each point in time. Essentially, they are a probabilistic expression of short-term fluctuations deviating from long-term trends, used to identify non-true trend signals in indicator changes. Based on the probability vectors of false improvement and false deterioration, as well as the set of trend vectors, this study analyzes the degree of synchronization of changes among multiple indicators, calculates the correlation strength, and constructs a cross-indicator correlation matrix. This matrix is ​​then used as input for subsequent threshold adjustment and trend reconstruction.

[0024] In practice, the false improvement probability vector and the false deterioration probability vector are fused with the trend vector set. For each indicator at each time point, a composite feature vector containing trend component, fluctuation component, and false probability is constructed. Subsequently, all indicators are paired at the indicator dimension to calculate the degree of synchronization of their transformation within the same time window, including the consistency of trend direction, the correlation of fluctuation amplitude, and the degree of coordinated change of false probability. These results are used as similarity measures between indicators.

[0025] After obtaining the similarity measure, adjacency relationships are constructed for all indicators, and an initial association network structure is generated using similarity as edge weights. Then, connectivity analysis and weight selection are performed on the network structure to retain highly correlated indicator connections, forming a stable indicator dependency subgraph. Subsequently, the subgraph is expanded along the time dimension to obtain a multidimensional association matrix that changes over time, which can reflect the dynamic dependencies between different indicators.

[0026] The cross-indicator correlation matrix describes the matrix structure of the relationship between multiple indicators within the same time window. The matrix elements are the correlation strengths, representing the degree of synchronization of changes, trend consistency, and degree of spurious probability co-change between any two indicators. It is used to characterize the linkage structure between multiple indicators and provide multi-indicator coupling information for subsequent dynamic threshold generation and risk scoring. Based on the short-term fluctuation components in the trend vector set, the dynamic time warping algorithm is used to calculate the degree of synchronization of changes between different indicators through sequence alignment analysis within a sliding time window. Based on the long-term trend components in the trend vector set, correlation analysis is performed on the trend direction sequence, and the Pearson correlation coefficient method is used to calculate the trend consistency between different indicators.

[0027] Based on the probability vectors of false improvement and false deterioration, the joint distribution relationship of the probability sequences is modeled, and the mutual information algorithm is used to calculate the degree of synergistic change of false probabilities among different indicators. The correlation strength is obtained by normalizing and weighting the degree of synchronization of changes, consistency of trends, and degree of spurious probability of coordinated changes. Based on the trend vector set, pattern decomposition analysis is performed on the deviation relationship between short-term fluctuations and long-term trends to obtain the false probability vector of each indicator within the corresponding time window, including: Based on the trend vector set, by constructing the difference between the long-term trend component and the short-term fluctuation component, a set of difference sequences reflecting the degree of local deviation is obtained as candidate analysis input; In practice, the trend vector set is expanded index by index, and the long-term trend component sequence and short-term fluctuation component sequence are extracted for each index and aligned under a unified time index. Then, a difference operation is performed on each time point to construct the offset value sequence between the short-term fluctuation component and the long-term trend component, so as to obtain a difference sequence reflecting the degree of deviation of local changes. Next, a sliding window segmentation process is performed on the difference sequence to divide the continuous time interval into multiple candidate segments, and the offset mean, variance and consistency of change direction are calculated in each segment to form a local offset feature set.

[0028] After obtaining the local offset feature set, preliminary anomaly labeling information is added to each time segment and bound to the differential sequence to form a joint feature structure containing the offset degree and preliminary anomaly state; then, this structure is output as a candidate analysis sequence.

[0029] The difference sequence set is a set of sequences obtained by subtracting the short-term fluctuation components and the long-term trend components from the trend vector set point by point. This sequence reflects the degree of deviation of local changes from the overall trend at each point in time. It is used to characterize the intensity of local changes and is a direct input for identifying abnormal fluctuations. Based on the differential sequence set, a set of potential abnormal fluctuation intervals is obtained by jointly analyzing the local offset amplitude and time continuity. In practice, the candidate analysis sequence is used as input. Within each time segment, the difference value is filtered by amplitude threshold, and the time point with the offset amplitude higher than the local mean is selected as the initial anomaly candidate point. Then, the temporal continuity of these candidate points is detected. By judging the interval relationship between adjacent time points, the discrete points are connected into continuous intervals to form a preliminary set of abnormal fluctuation intervals. Next, the duration, maximum offset value and rate of change of each interval are calculated to construct an interval-level feature vector.

[0030] After the interval-level features are constructed, the directional information of the trend component is introduced, and a directional consistency test is performed on each abnormal interval to determine the relationship between the fluctuation direction within the interval and the long-term trend direction. Subsequently, the intervals that meet the conditions of high offset, continuity and directional anomaly are marked as potential false change intervals. The potential abnormal fluctuation interval set is a set of time intervals obtained from the difference sequence through amplitude screening and time continuity analysis. Each interval represents a fluctuation with a significant deviation from the trend within a continuous time period, which is used to locate the time period in which abnormal fluctuations occur.

[0031] Local offset amplitude refers to the degree of deviation of the short-term fluctuation component from the long-term trend component at a certain point in time or within a certain time window. It is used to characterize the strength of the deviation of local changes from the overall trend. Extract the features of each potential abnormal fluctuation interval within the set of potential abnormal fluctuation intervals, measure the degree of deviation between them and the historical trend pattern, and obtain the false change discrimination feature vector of each potential abnormal fluctuation interval. In practice, potential abnormal fluctuation intervals and their interval feature vectors are used as inputs. For each interval, the duration, rate of change of amplitude, frequency of change of direction, and connection relationship with neighboring intervals are extracted. Then, these interval features are mapped to a unified feature space, and the interval features between different indicators are standardized to ensure the consistency of subsequent measurements. Next, a deviation function based on distance metric is constructed, and the historical trend pattern of each interval feature vector is compared with the corresponding indicator. The Mahalanobis distance algorithm is used to calculate the degree of deviation. After the deviation calculation is completed, a multi-dimensional feature description containing the degree of deviation, duration and direction of change is generated for each interval, and the description is encoded into a pseudo-change discrimination feature vector. Then, the vector is fused with the trend vector set to form an extended feature structure.

[0032] The pseudo-change discrimination feature vector is a combination of multi-dimensional features extracted from each potential abnormal fluctuation interval, including but not limited to deviation magnitude, duration, frequency of directional changes, and connection relationship with adjacent intervals. These features are uniformly encoded into a vector to describe the change attributes of the interval and to quantify the degree of difference between abnormal intervals and normal trend patterns.

[0033] Historical trend patterns refer to a set of reference patterns extracted from the past time series of the same indicator that can represent the change pattern of the indicator under stable or typical conditions. They are used as a benchmark for comparison of the current abnormal interval characteristics and are derived from the historical time period screening results in the trend vector set.

[0034] Based on the set of spurious change discriminant feature vectors and trend vectors, the multidimensional features are subjected to time expansion and normalization fusion processing, and the deviation degree and trend direction are jointly encoded. A probability mapping model is used to generate continuous probability values, and the probabilities are classified according to the trend direction to obtain spurious probability vectors, including spurious improvement probability vectors and spurious deterioration probability vectors.

[0035] In practice, the discriminative feature vectors are expanded in chronological order, and the feature information of the interval to which each time point belongs is summarized. Normalization is performed on features of different dimensions to eliminate differences in units. Then, a weighted fusion process is performed on the normalized features. By constructing a feature combination function, the degree of deviation, duration, and frequency of directional changes are mapped to a unified scalar representation. Next, a probability mapping operation is performed on the scalar representation to convert it into a false change probability value in the interval [0,1].

[0036] After the probability is generated, the probability value at each time point is classified according to the direction information of the trend component, and it is divided into false improvement probability or false deterioration probability, forming a complete false feature probability vector. In this embodiment, by performing hierarchical decomposition on the time series of each indicator in the disease perception data matrix, the original changes are split into long-term trends, short-term fluctuations and noise components. Furthermore, differential sequences and abnormal intervals are constructed around the deviation relationship between short-term fluctuations and long-term trends, so that the local disturbances originally mixed in the time series are expressed in a structured way. Based on this, feature encoding is performed on the duration, offset magnitude and direction of abnormal intervals, and the degree of deviation is calculated in combination with historical trend patterns to form quantifiable features for identifying false changes. Then, probability mapping is used to generate probability vectors for false improvement and false deterioration, so that a distinguishable probability expression can be formed between short-term anomalies and the real trend. Simultaneously, by integrating trend components and pseudo-probability information, and utilizing dynamic time warping, Pearson correlation coefficient, and mutual information, a cross-indicator correlation matrix is ​​constructed. This allows for a unified characterization of the synchronous relationships and collaborative deviations among multiple indicators, thus providing structured constraints for subsequent threshold modeling. The logic lies in first breaking down the changes, then quantifying the deviations, and finally reconstructing the relationships, forming a progressive analysis chain from single indicators to multiple indicators.

[0037] For example, in clinical practice, if a patient's blood oxygen level drops briefly but recovers quickly, and respiratory parameters and hemodynamics do not change synchronously, this method can identify the drop as a local abnormal range by comparing short-term fluctuations with long-term trend deviations, and assign it a higher probability of false deterioration. At the same time, it can identify the lack of synergistic changes in cross-index correlation analysis, thereby avoiding misjudging the instantaneous drop as a true deterioration of the condition.

[0038] Specifically: S300 includes: performing structural constraint analysis on the coupling relationship between various indicators based on the cross-indicator correlation matrix, and obtaining a dynamic threshold sequence by combining the trend vector set and the false probability vector; In practice, the cross-indicator correlation matrix is ​​used as the structural constraint input to analyze the correlation strength between indicators. The correlation strength is used as the weight to perform a weighted coupling operation on the long-term trend components in the trend vector set. This makes the trend expression of a certain indicator at the current moment no longer depend solely on its own changes, but integrates the trend information of its related indicators within the same time window, thereby forming a coupled trend reference value. Based on this, a false improvement probability vector and a false deterioration probability vector are introduced to perform suppression modulation processing on the coupled trend reference value, weakening the trend contribution corresponding to the time point with a higher false probability, so that the subsequent threshold calculation process will preferentially retain stable and reliable change components, and obtain the suppressed effective trend sequence. Subsequently, a sliding window segmentation process is performed on the effective trend sequence on the time axis. Within each local time window, the trend value is statistically modeled to extract its local concentration and dispersion features. The concentration is characterized by the local average level, and the dispersion is characterized by the local fluctuation range or change amplitude. Next, based on the above statistical results, upper and lower boundaries are constructed within each time window. An upper threshold is formed by superimposing an offset proportional to the degree of dispersion on the local average level, and a lower threshold is formed by subtracting the corresponding offset from the local average level, thus obtaining the threshold interval within the time window. Finally, the threshold intervals obtained from each time window are spliced ​​and aligned in chronological order, so that each time point corresponds to a set of upper and lower threshold boundaries, thereby forming a complete dynamic threshold sequence. This sequence is then output to the subsequent anomaly screening and consistency calibration steps, while the threshold change trend is fed back to the preceding trend decomposition module for adjusting the parameters in the trend component extraction process.

[0039] Dynamic threshold sequence refers to a set of upper and lower limit threshold pairs generated on the time axis for each indicator and each time point. Its value is not a fixed constant, but a time-varying boundary obtained by dynamic modulation of the coupling relationship of multiple indicators. It is used to replace the fixed threshold and realize the adaptive judgment boundary that changes with time, providing constraints for subsequent anomaly screening. Based on dynamic threshold sequences, trend vector sets, and false probability vectors, preliminary anomaly screening markers are obtained by performing threshold constraints and consistency calibration analysis on the current indicators. In practice, a dynamic threshold sequence is applied to the time series of each indicator to identify points exceeding the threshold for a single indicator. Then, a cross-indicator correlation matrix is ​​introduced to perform propagation analysis on these outliers, determining whether an anomaly in a given indicator occurs synchronously in its associated indicators, thereby identifying isolated and linked anomalies. Next, a false probability vector is used as a filtering condition to suppress outliers with high false probabilities and enhance outliers with low false probabilities but exhibiting cross-indicator linkage characteristics. Based on this, the anomaly markers are reconstructed to obtain a corrected anomaly marker sequence. This sequence, obtained by applying dynamic threshold constraints and cross-indicator consistency calibration to the original anomaly detection, is the final anomaly identification sequence used to replace the initial screening results of single indicators, forming a more reliable anomaly determination and serving as the core input for subsequent risk matrix construction. Based on the revised preliminary anomaly screening markers and cross-indicator correlation matrix, a multidimensional risk index matrix for disease perception is obtained by performing unified mapping and structured integration analysis on multidimensional features. Each element in the multidimensional risk index matrix represents the comprehensive risk vector of the corresponding indicator at each time point.

[0040] In practice, the modified preliminary anomaly screening markers are expanded according to the indicator dimension to obtain a binary anomaly sequence for each indicator on the full time axis. Then, continuous interval identification processing is performed on the anomaly sequence to merge adjacent anomaly points into anomaly intervals and record the start time, end time and interval length of each anomaly interval. Next, within each abnormal interval, the duration is normalized, and the abnormal intensity value is calculated by combining the deviation between the original index value and the dynamic threshold within the interval. This ensures that each abnormal point not only has a label indicating whether it is abnormal, but also a quantitative result of the degree of abnormality. The above abnormal intensity is then remapped back to the original time axis according to the time index to construct a time-point-level abnormal feature sequence. Then, the cross-index correlation matrix is ​​used as input to filter index pairs with correlation strength higher than a set threshold to construct an index adjacency relationship set; subsequently, the abnormal feature sequences are aligned among these correlated indicators so that abnormal states at the same time point can be compared among different indicators. Next, for each pair of related indicators, the difference in their abnormal intensity is calculated within the same time window, and it is determined whether there is a synchronous or delayed abnormal relationship. If both indicators show abnormalities in adjacent time points, an abnormality propagation calculation is performed, and the abnormal intensity of one indicator is allocated to the other indicator according to the correlation weight, thereby forming cross-indicator abnormality compensation. The abnormal intensity sequence after propagation and compensation is re-summarized to obtain the cross-indicator aligned abnormal correlation feature sequence, which is used to correct the abnormal interval division.

[0041] The abnormal correlation feature sequence is expanded in chronological order, and the abnormal intensity value corresponding to each indicator is summarized at each time point. At the same time, the trend component in the trend vector set and the probability value in the false probability vector are introduced. These features are combined according to the indicator dimension to form a multi-dimensional feature set, and unified encoding processing is performed. The data at each time point is marked according to the three-dimensional index of indicator number, time index, and feature type, and normalization processing is performed on different features to make them comparable in a unified numerical space. Subsequently, by combining the cross-indicator correlation matrix, the structure of the indicator dimensions is rearranged, and indicators with high correlation strength are arranged adjacently in the matrix, so that the matrix structure can reflect the coupling relationship between indicators. Finally, the data of all time points are stacked according to the above structure to form a complete multi-dimensional risk indicator matrix. This matrix serves as a unified input structure for subsequent risk component decomposition, contribution calculation, and correlation analysis. Each element in the matrix represents the comprehensive risk vector of a certain indicator at a certain time point, including the anomaly strength and false probability. In this embodiment, by introducing a cross-index correlation matrix to weightedly couple the trends of each index, and combining false probability to suppress unstable changes, the dynamic threshold no longer originates from the historical fluctuations of a single index, but is driven by the synergistic relationship of multiple indicators and the effective trend, thereby forming a threshold boundary that adaptively adjusts with state changes on the time axis. Based on this, by combining dynamic thresholds with anomaly propagation mechanisms, cross-index consistency calibration is performed on single-point anomalies, enabling the distinction between isolated anomalies and linked anomalies. Furthermore, unreliable anomalies are filtered out through false probabilities, transforming anomaly labeling from whether anomaly is present into a composite expression of anomaly strength and credibility. Furthermore, by propagating and compensating for anomalies among related indicators, the originally scattered anomaly information is structurally aligned and integrated, ultimately constructing a multidimensional risk indicator matrix that reflects the coupling relationship between indicators. The logic is to first reshape trends through correlation constraints, then filter effective information through probability modulation, and finally achieve a unified expression of multi-indicator risk through propagation and structural rearrangement.

[0042] For example, in real-world scenarios, when a patient's blood pressure drops briefly but heart rate and blood oxygen levels do not show synchronous abnormalities, this drop might be considered abnormal in traditional methods. However, this method identifies the lack of cross-index linkage through correlation analysis and suppresses it using false probability, preventing it from being amplified after dynamic threshold correction. Conversely, when a drop in blood pressure is accompanied by an increase in heart rate and a decrease in blood oxygen, the intensity of this linkage abnormality is enhanced through an abnormality propagation mechanism, making it centrally reflected in the risk matrix. This makes abnormality identification and risk expression more consistent with the overall physiological change structure.

[0043] Specifically: S400 includes: based on a multi-dimensional risk indicator matrix, by performing component expansion, time-series accumulation and proportion normalization calculation on the comprehensive risk vector of each indicator in the time dimension, an indicator contribution matrix reflecting the risk contribution degree of each indicator is obtained, and each element in the indicator contribution matrix represents the indicator contribution ratio of each indicator. The indicator contribution ratio represents the relative contribution weight of a particular indicator to the overall risk at a specific point in time, and is used to achieve multi-indicator risk attribution and redundancy removal. The indicator contribution matrix is ​​the distribution structure of the contribution ratios of each indicator, and is used for identifying duplicate contributions and analyzing intervention paths. Based on the indicator contribution matrix, a risk score sequence is obtained by performing time alignment and weighted aggregation analysis on the contribution ratio of each indicator at the same time point. This sequence is a time-based representation of the overall risk intensity as a function of time, obtained by aggregating the risk contributions of each indicator on the time axis. It is used to characterize the change of the overall risk intensity over time and serves as the input basis for subsequent correlation constraint correction and intervention decision analysis.

[0044] The index pairs with a correlation strength exceeding the preset correlation threshold are selected to obtain highly correlated indicators. Synchronous offset detection and duplicate contribution identification calculation are performed on the risk change time series of highly correlated indicators to obtain the corrected risk score series. Specifically, based on highly correlated indicators, the comprehensive risk vectors of these indicators within the same time window are aligned. Subsequently, the aligned sequence is subjected to first-order difference operation to obtain the change difference sequence of each indicator in the time dimension, and it is detected whether there is a situation of continuous change in the same direction and similar magnitude. If the condition is met, it is determined to be a synchronous change relationship. Among them, the magnitude is similar by calculating the absolute difference of the change difference between two indicators and comparing it with a preset difference threshold. When the difference is less than the threshold, it is determined to be a magnitude similar. Next, for the indicator pairs determined to be synchronously changing, their contribution ratio sequence in the indicator contribution matrix is ​​calculated. By comparing the differences in the contribution ratio of each indicator at the same time point, a set of indicators with similar contribution distributions is identified, and a constraint redistribution process is performed on the contribution ratio within this set to avoid multiple indicators repeatedly expressing the same risk source. Finally, the adjusted indicator contribution ratios are re-aggregated in the time dimension to form a corrected risk contribution sequence.

[0045] Based on the corrected risk score sequence, multidimensional index mapping and structural rearrangement calculations are performed on the contribution ratio of indicators to obtain a disease-aware multidimensional risk score matrix.

[0046] Specifically, the corrected risk score sequence is expanded in chronological order, and the contribution ratio of each indicator is extracted at each time point. At the same time, the local contribution distribution characteristics in the indicator contribution matrix are read, and the two are aligned and paired to form a combination feature of contribution ratio and distribution status. Subsequently, the combined features are uniformly indexed and encoded, with time as the first dimension index, indicators as the second dimension index, and feature type as the third dimension index. Normalization is then performed on the data in different dimensions to make them comparable within a unified numerical space. Next, based on the cross-indicator correlation matrix, highly correlated indicators are arranged adjacently in the matrix, so that strongly correlated indicators are structurally adjacent, thereby completing the rearrangement of the matrix structure; finally, the contribution ratios and structural characteristics of all time points are stacked according to the above index relationship to form a complete multidimensional risk scoring matrix. This matrix is ​​a structured risk expression matrix constructed in the contribution space, which is used to support intervention path modeling. Based on a multidimensional risk indicator matrix, by performing component expansion, time-series accumulation, and proportion normalization calculations on the comprehensive risk vector of each indicator in the time dimension, an indicator contribution matrix reflecting the degree of risk contribution of each indicator is obtained, including: Based on a multidimensional risk indicator matrix, a risk component sequence with a unified direction is obtained by performing component decomposition and directional consistency adjustment on the comprehensive risk vector of each indicator at each time point. In practice, firstly, the comprehensive risk vector sequence of each indicator in the time dimension is extracted from the multidimensional risk indicator matrix, and the abnormal intensity component and the false probability component are expanded respectively. Then, the amplitude preservation processing is performed on the abnormal intensity component, and the reverse mapping processing is performed on the false probability component to convert it into a confidence weight, that is, the higher the probability, the lower the weight. Next, for each time point, the anomaly intensity and credibility weight are multiplied to calculate the risk expression of each indicator at that time point, so that the anomaly intensity and credibility are considered simultaneously, thus obtaining a risk component sequence in a unified direction. This sequence is a unidirectional and consistent risk expression sequence obtained by processing the comprehensive risk vector, which is used to avoid conflict between rising and falling directions. Based on the risk component sequence, the cumulative risk of each indicator is obtained by performing cumulative statistics and distribution density calculation on the risk value of each indicator in the time dimension; this cumulative risk is used to calculate the overall contribution capacity of the indicator. In practice, the risk component sequence of each indicator is accumulated over the entire time interval to obtain the total cumulative risk value of the indicator; then, the local cumulative risk value is calculated within the sliding time window to reflect the degree of risk concentration of the indicator in different time periods. Next, the cumulative risk values ​​of all indicators are sorted and their proportion in the overall risk is calculated. At the same time, the distribution density of the risk contribution of each indicator over time is recorded, i.e. whether the risk is concentrated in a few time points or evenly distributed. Finally, the cumulative risk amount and distribution characteristics of each indicator over time are obtained. Based on the cumulative risk of each indicator over time, the indicator contribution matrix is ​​obtained by performing normalization mapping and time alignment processing on its proportion in the overall risk.

[0047] In practice, the cumulative risk of all indicators over time is summed to obtain the total global risk. Then, the cumulative risk value of each indicator is divided by this total to obtain its contribution ratio in the overall risk. Next, this contribution ratio is expanded along the time dimension, that is, the risk ratio of each indicator in a local time window is mapped back to the corresponding time point, so that each time point corresponds to the contribution ratio of an indicator, in order to construct an indicator contribution matrix.

[0048] In this embodiment, based on the multidimensional risk index matrix, the processing further decomposes the comprehensive risk vector composed of anomaly intensity and false probability into a risk component sequence in a unified direction, and forms an index contribution matrix by calculating time accumulation and distribution density, so that the role of each index in the overall risk no longer depends on a single point anomaly, but is measured based on the cumulative performance over time. Subsequently, the contributions of multiple indicators are time-aligned and aggregated through the risk scoring sequence. Then, cross-indicator correlation constraints are combined to perform redundancy correction on indicators that change synchronously and have overlapping contributions, thereby obtaining the corrected risk scoring sequence. Finally, a multi-dimensional risk scoring matrix with structural correlation is constructed. Its internal logic is to first eliminate the interference of false fluctuations on risk expression through credibility modulation, then strengthen the persistence characteristics of real risk signals through time accumulation, and then weaken the problem of multiple indicators repeatedly expressing the same risk source through correlation constraints, ultimately forming a risk expression system that has both time consistency and structural correlation. For example, in the monitoring of a patient, heart rate and blood pressure rise abnormally synchronously within the same time period. However, if one of the changes is caused by drug intervention and has a high probability of being false, then this part is suppressed when constructing the risk component in the same direction. In the contribution matrix calculation and subsequent synchronous detection, only the indicators with higher real contributions are retained, thereby avoiding the repeated amplification of the same risk event by the two indicators, making the final risk score more stable and having an interpretable contribution source structure.

[0049] Specifically, S500 includes: based on a multidimensional risk scoring matrix, by performing time alignment matching analysis and intervention response difference calculation on the correspondence between historical intervention records and indicator contribution distribution, an intervention impact matrix reflecting the intervention path is obtained; this matrix describes the mapping relationship between intervention behavior and indicator contribution changes, and is used to model the intervention path and support intervention prediction. In practice, firstly, the multidimensional risk scoring matrix is ​​used as input to extract the sequence of indicator contribution ratios at each time point, and the historical intervention records within the corresponding time window are read simultaneously. The intervention behaviors are aligned to the sequence of indicator contribution changes according to the time axis, thereby constructing a sequence of correspondence between intervention actions and contribution changes. Historical intervention records represent the intervention behaviors and their timing information implemented along a historical timeline, used to construct an intervention impact matrix and provide a reference for real responses; intervention behaviors refer to the category of operations that regulate the patient's condition, including but not limited to pharmacological interventions, respiratory support interventions, and hemodynamic regulation. Subsequently, within each intervention time window, the difference calculation is performed on the changes in the contribution ratio of each indicator before and after the intervention, and the duration and magnitude of the change are statistically analyzed within the sliding time window to form an intervention response change sequence, which is used to characterize the impact path of the intervention on the contribution structure of different indicators. Next, the above-mentioned intervention response change sequences are grouped and aggregated according to intervention type, and mean and dispersion calculations are performed on the response results of the same type of intervention in different time windows to obtain a stable intervention effect pattern and construct an intervention impact matrix.

[0050] Based on the intervention impact matrix, multi-intervention combination mapping analysis and time evolution extrapolation calculation are performed on the indicator contribution change paths under different interventions to obtain the risk contribution prediction sequence corresponding to various interventions. This sequence represents the predicted trajectory of indicator contribution changes over time under different interventions, and is used to compare the evolutionary effects of different intervention schemes.

[0051] In practice, firstly, the intervention impact matrix is ​​used as input, and combined with the indicator contribution matrix at the current moment, the indicator contribution distribution in the initial state is constructed; then, according to different candidate interventions, the corresponding impact paths are extracted from the intervention impact matrix and mapped to the current indicator contribution state, thereby generating the initial change sequence under each intervention. Next, a recursive calculation is performed on the above change sequence in the time dimension. That is, at each subsequent time point, the predicted contribution of the previous moment is used as input and the effect of the intervention influence matrix is ​​superimposed to form a continuous time evolution sequence, thereby obtaining the complete risk contribution prediction trajectory. Finally, the prediction sequences generated by different intervention schemes are aligned and arranged in a structured manner on a unified time axis to form a set of prediction sequences for risk contribution of multiple schemes.

[0052] Based on the risk contribution prediction sequence, the volatility stability of the contribution distribution of various interventions over time is analyzed to obtain the optimal intervention.

[0053] In practice, firstly, for each intervention plan, the cumulative contribution offset is calculated in the time dimension of the risk contribution prediction sequence. That is, the total change of the plan relative to the initial state is statistically analyzed, and it is recorded whether the trend of change shows monotonic convergence or fluctuation diffusion characteristics. Subsequently, fluctuation stability analysis was performed on the time series of each scheme. By calculating the difference in contribution changes between adjacent time points and statistically analyzing the fluctuation frequency and amplitude distribution, the stability of different schemes during the time evolution process was determined. Next, the cumulative offset and stability results are jointly sorted, and the intervention scheme that satisfies the preset conditions between offset control and fluctuation constraints is selected as the optimal scheme. The optimal scheme represents the intervention scheme that makes the risk contribution evolution relatively conform to the preset constraints among all candidate intervention schemes, and is used as the final decision output of the entire system.

[0054] In this invention, all parameters are dimensionless by using dimensionless processing technology to remove their dimensions, and all thresholds can be obtained by the mean-standard deviation method. In this embodiment, the processing procedure aligns historical intervention records with changes in indicator contributions over time based on the multidimensional risk scoring matrix, and extracts the dynamic change trajectory of contribution distribution before and after intervention through differential calculation and window statistics, thereby constructing an intervention impact matrix and forming a quantifiable mapping relationship between intervention behavior and multi-indicator risk contributions. Based on this, by mapping the intervention impact path to the current indicator contribution status and performing recursive superposition operations in the time dimension, a risk contribution prediction sequence corresponding to various intervention schemes is formed. Then, the intervention schemes are screened and ranked by the joint analysis of cumulative offset and volatility stability. Its internal logic is to first use real historical responses to establish a causal approximate structure of intervention and contribution changes, and then apply this structure to the current state through time evolution, thereby transforming static risk assessment into a dynamic and predictable process. Finally, stability constraints are used to avoid short-term volatility from dominating decision-making. For example, in a patient, both vasopressors and respiratory support can reduce risk contribution, but the former fluctuates more in a short period of time while the latter shows a continuous convergence trend. Predictive sequence analysis can identify that the latter is more stable in the time dimension, thereby avoiding the wrong selection of intervention programs with large fluctuations due to rapid short-term improvement, making the final decision more continuous and controllable.

[0055] Example 2, as Figure 2 As shown, based on the same inventive concept as the intelligent processing method for cross-disease critical illness monitoring data provided in Embodiment 1, this embodiment of the invention also provides an intelligent processing system for cross-disease critical illness monitoring data, including: The data acquisition and processing module 11 is used to acquire patient physiological data from different monitoring devices, perform preliminary abnormality screening, combine disease labels to perform indicator mapping, and construct a disease perception data matrix. The data decomposition and analysis module 12 is used to perform hierarchical decomposition of the time series of each indicator based on the disease perception data matrix to obtain a set of trend vectors, and generate a false probability vector and a cross-indicator correlation matrix through short-term fluctuation and long-term trend deviation analysis. The cross-disease dynamic anomaly identification module 13 is used to perform threshold dynamic modeling through cross-indicator correlation matrix, correct the preliminary anomaly screening mark and construct a multi-dimensional risk indicator matrix. The cross-disease risk analysis module 14 is used to perform component expansion, direction unification and time series cumulative analysis on the multidimensional risk indicator matrix, generate the indicator contribution matrix and obtain the multidimensional risk score matrix after correlation constraint correction; The intervention prediction module 15 is used to read historical intervention records, combine them with a multidimensional risk score matrix, construct a risk contribution prediction sequence by matching intervention records and extrapolating contribution change paths, and determine the optimal intervention based on time stability analysis.

[0056] Furthermore, the data acquisition and processing module 11 is also used for: The patient's physiological data includes at least hemodynamic parameters, respiratory parameters, blood oxygen levels, laboratory indicators, and drug intervention information; and each indicator in the patient's physiological data undergoes preliminary anomaly screening, which includes exceeding threshold checks, missing data marking, and abnormal fluctuation identification. Based on the patient's previous medical records and admission diagnosis information, disease labels are mapped to each indicator to form a disease-aware data matrix. The disease-aware data matrix records the preliminary abnormal screening markers and time series of each indicator under the corresponding disease label.

[0057] Furthermore, the data decomposition and analysis module 12 is also used for: Based on the disease-aware data matrix, a set of trend vectors containing long-term trend components, short-term fluctuation components, and noise components is obtained by performing hierarchical decomposition analysis on the time series of each indicator. Based on the trend vector set, pattern decomposition analysis is performed on the deviation relationship between short-term fluctuations and long-term trends to obtain the false probability vector of each indicator within the corresponding time window. Specifically: Based on the trend vector set, a set of difference sequences reflecting the degree of local deviation is obtained by constructing a difference between the long-term trend component and the short-term fluctuation component. Based on the differential sequence set, a set of potential abnormal fluctuation intervals is obtained by jointly analyzing the local offset amplitude and time continuity. Extract the features of each potential abnormal fluctuation interval within the set of potential abnormal fluctuation intervals, measure the degree of deviation between them and the historical trend pattern, and obtain the false change discrimination feature vector of each potential abnormal fluctuation interval. Based on the set of spurious change discriminant feature vectors and trend vectors, the multidimensional features are subjected to time expansion and normalization fusion processing, and the deviation degree and trend direction are jointly encoded. A probability mapping model is used to generate continuous probability values, and the probabilities are classified according to the trend direction to obtain spurious probability vectors, including spurious improvement probability vectors and spurious deterioration probability vectors.

[0058] Based on the set of probability vectors for false improvement and false deterioration, as well as trend vectors, we analyze the degree of synchronization of changes among multiple indicators, calculate the correlation strength, and construct a cross-indicator correlation matrix.

[0059] Furthermore, the cross-disease dynamic anomaly identification module 13 is also used for: Based on the cross-indicator correlation matrix, structural constraint analysis is performed on the coupling relationship between the indicators, and dynamic threshold sequence is obtained by combining the trend vector set and the false probability vector. Based on dynamic threshold sequences, trend vector sets, and false probability vectors, preliminary anomaly screening markers are obtained by performing threshold constraints and consistency calibration analysis on the current indicators. Based on the revised preliminary anomaly screening markers and cross-indicator correlation matrix, a multidimensional risk index matrix for disease perception is obtained by performing unified mapping and structured integration analysis on multidimensional features. Each element in the multidimensional risk index matrix represents the comprehensive risk vector of the corresponding indicator at each time point.

[0060] Furthermore, the cross-disease risk analysis module 14 is also used for: Based on a multidimensional risk indicator matrix, by performing component expansion, time-series cumulative calculation, and proportion normalization on the comprehensive risk vector of each indicator in the time dimension, an indicator contribution matrix reflecting the degree of risk contribution of each indicator is obtained. Each element in the indicator contribution matrix represents the indicator contribution ratio of each indicator. Specifically: Based on a multidimensional risk indicator matrix, a risk component sequence with a unified direction is obtained by performing component decomposition and directional consistency adjustment on the comprehensive risk vector of each indicator at each time point. Based on the risk component sequence, the cumulative risk of each indicator over time is obtained by performing cumulative statistics and distribution density calculation on the risk values ​​of each indicator over time. Based on the cumulative risk of each indicator over time, the indicator contribution matrix is ​​obtained by performing normalization mapping and time alignment processing on its proportion in the overall risk.

[0061] Based on the indicator contribution matrix, time alignment and weighted aggregation analysis are performed on the contribution ratio of each indicator at the same point in time to obtain the risk score sequence. The index pairs with a correlation strength exceeding the preset correlation threshold are selected to obtain highly correlated indices. Synchronous offset detection and duplicate contribution identification calculation are performed on the highly correlated indices to obtain the corrected risk score sequence. Based on the corrected risk score sequence, multidimensional index mapping and structural rearrangement calculations are performed on the contribution ratio of indicators to obtain a disease-aware multidimensional risk score matrix.

[0062] Furthermore, the intervention prediction module 15 is also used for: Based on the multidimensional risk scoring matrix, an intervention impact matrix reflecting the intervention path is obtained by performing time alignment matching analysis and intervention response difference calculation on the correspondence between historical intervention records and indicator contribution distribution. Based on the intervention impact matrix, multi-intervention combination mapping analysis and time evolution extrapolation calculation are performed on the indicator contribution change paths under different interventions to obtain the risk contribution prediction sequence corresponding to various interventions; Based on the risk contribution prediction sequence, the volatility stability of the contribution distribution of various interventions over time is analyzed to obtain the optimal intervention.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent processing method for cross-disease critical illness monitoring data, characterized in that, The method includes: Physiological data of patients are acquired from different monitoring devices, and preliminary abnormality screening is performed. Indicator mapping is performed in combination with disease labels to construct a disease perception data matrix. Based on the disease perception data matrix, a set of trend vectors is obtained by performing hierarchical decomposition on the time series of each indicator, and a false probability vector and cross-indicator correlation matrix are generated through short-term fluctuation and long-term trend deviation analysis. By performing threshold dynamic modeling through cross-indicator correlation matrix, the initial anomaly screening markers are corrected and a multi-dimensional risk indicator matrix is ​​constructed. Perform component expansion, direction unification, and time-series cumulative analysis on the multidimensional risk indicator matrix to generate an indicator contribution matrix and obtain a multidimensional risk score matrix after correlation constraint correction. By reading historical intervention records and combining them with a multidimensional risk scoring matrix, a risk contribution prediction sequence is constructed through intervention record matching and contribution change path deduction, and the optimal intervention is determined based on time stability analysis. Based on the disease-aware data matrix, a set of trend vectors is obtained by performing hierarchical decomposition on the time series of each indicator. Furthermore, false probability vectors and cross-indicator correlation matrices are generated through short-term fluctuation and long-term trend deviation analysis, including: Based on the disease-aware data matrix, a set of trend vectors containing long-term trend components, short-term fluctuation components, and noise components is obtained by performing hierarchical decomposition analysis on the time series of each indicator. Based on the trend vector set, by performing pattern decomposition analysis on the deviation relationship between short-term fluctuations and long-term trends, the false probability vector of each indicator within the corresponding time window is obtained. Based on the set of false improvement probability vectors, false deterioration probability vectors, and trend vectors, we analyze the degree of synchronization of changes among multiple indicators, calculate the correlation strength, and construct a cross-indicator correlation matrix. Based on the trend vector set, pattern decomposition analysis is performed on the deviation relationship between short-term fluctuations and long-term trends to obtain the false probability vector of each indicator within the corresponding time window, including: Based on the trend vector set, a set of difference sequences reflecting the degree of local deviation is obtained by constructing a difference between the long-term trend component and the short-term fluctuation component. Based on the differential sequence set, a set of potential abnormal fluctuation intervals is obtained by jointly analyzing the local offset amplitude and time continuity. Extract the features of each potential abnormal fluctuation interval within the set of potential abnormal fluctuation intervals, measure the degree of deviation between them and the historical trend pattern, and obtain the false change discrimination feature vector of each potential abnormal fluctuation interval. Based on the set of spurious change discriminant feature vectors and trend vectors, the multidimensional features are subjected to time expansion and normalization fusion processing, and the deviation degree and trend direction are jointly encoded. A probability mapping model is used to generate continuous probability values, and the probabilities are classified according to the trend direction to obtain spurious probability vectors, including spurious improvement probability vectors and spurious deterioration probability vectors.

2. The intelligent processing method for cross-disease critical illness monitoring data according to claim 1, characterized in that, Physiological data of patients are acquired from different monitoring devices, and preliminary abnormality screening is performed. This data is then combined with disease-specific tags to map indicators and construct a disease-aware data matrix, including: The patient's physiological data includes at least hemodynamic parameters, respiratory parameters, blood oxygen levels, laboratory indicators, and drug intervention information; and each indicator in the patient's physiological data undergoes preliminary anomaly screening, which includes exceeding threshold checks, missing data marking, and abnormal fluctuation identification. Based on the patient's previous medical records and admission diagnosis information, disease labels are mapped to each indicator to form a disease-aware data matrix. The disease-aware data matrix records the preliminary abnormal screening markers and time series of each indicator under the corresponding disease label.

3. The intelligent processing method for cross-disease critical illness monitoring data according to claim 1, characterized in that, By performing dynamic threshold modeling through cross-indicator correlation matrices, the initial anomaly screening markers are corrected and a multi-dimensional risk indicator matrix is ​​constructed, including: Based on the cross-indicator correlation matrix, structural constraint analysis is performed on the coupling relationship between the indicators, and dynamic threshold sequence is obtained by combining the trend vector set and the false probability vector. Based on dynamic threshold sequences, trend vector sets, and false probability vectors, preliminary anomaly screening markers are obtained by performing threshold constraints and consistency calibration analysis on the current indicators. Based on the revised preliminary anomaly screening markers and cross-indicator correlation matrix, a multidimensional risk index matrix for disease perception is obtained by performing unified mapping and structured integration analysis on multidimensional features. Each element in the multidimensional risk index matrix represents the comprehensive risk vector of the corresponding indicator at each time point.

4. The intelligent processing method for cross-disease critical illness monitoring data according to claim 3, characterized in that, Component expansion, direction unification, and time-series cumulative analysis are performed on the multidimensional risk indicator matrix to generate an indicator contribution matrix. After correlation constraint correction, a multidimensional risk scoring matrix is ​​obtained, including: Based on the multidimensional risk indicator matrix, by performing component expansion, time-series accumulation and proportion normalization calculation on the comprehensive risk vector of each indicator in the time dimension, an indicator contribution matrix reflecting the risk contribution of each indicator is obtained. Each element in the indicator contribution matrix represents the indicator contribution ratio of each indicator. Based on the indicator contribution matrix, time alignment and weighted aggregation analysis are performed on the contribution ratio of each indicator at the same time point to obtain the risk score sequence. The index pairs with a correlation strength exceeding the preset correlation threshold are selected to obtain highly correlated indices. Synchronous offset detection and duplicate contribution identification calculation are performed on the highly correlated indices to obtain the corrected risk score sequence. Based on the corrected risk score sequence, multidimensional index mapping and structural rearrangement calculations are performed on the contribution ratio of indicators to obtain a disease-aware multidimensional risk score matrix.

5. The intelligent processing method for cross-disease critical illness monitoring data according to claim 4, characterized in that, Based on a multidimensional risk indicator matrix, by performing component expansion, time-series accumulation, and proportion normalization calculations on the comprehensive risk vector of each indicator in the time dimension, an indicator contribution matrix reflecting the degree of risk contribution of each indicator is obtained, including: Based on a multidimensional risk indicator matrix, a risk component sequence with a unified direction is obtained by performing component decomposition and directional consistency adjustment on the comprehensive risk vector of each indicator at each time point. Based on the risk component sequence, the cumulative risk of each indicator over time is obtained by performing cumulative statistics and distribution density calculation on the risk values ​​of each indicator over time. Based on the cumulative risk of each indicator over time, the indicator contribution matrix is ​​obtained by performing normalization mapping and time alignment processing on its proportion in the overall risk.

6. The intelligent processing method for cross-disease critical illness monitoring data according to claim 5, characterized in that, By reading historical intervention records and combining them with a multidimensional risk scoring matrix, a risk contribution prediction sequence is constructed through intervention record matching and contribution change path deduction. The optimal intervention is then determined based on time stability analysis, including: Based on the multidimensional risk scoring matrix, an intervention impact matrix reflecting the intervention path is obtained by performing time alignment matching analysis and intervention response difference calculation on the correspondence between historical intervention records and indicator contribution distribution. Based on the intervention impact matrix, multi-intervention combination mapping analysis and time evolution extrapolation calculation are performed on the indicator contribution change paths under different interventions to obtain the risk contribution prediction sequence corresponding to various interventions; Based on the risk contribution prediction sequence, the volatility stability of the contribution distribution of various interventions over time is analyzed to obtain the optimal intervention.

7. An intelligent processing system for cross-disease critical illness monitoring data, used to implement the intelligent processing method for cross-disease critical illness monitoring data as described in any one of claims 1 to 6, characterized in that, include: The data acquisition and processing module is used to acquire patients' physiological data from different monitoring devices, perform preliminary abnormality screening, combine disease labels to map indicators, and construct a disease-aware data matrix. The data decomposition and analysis module is used to perform hierarchical decomposition of the time series of each indicator based on the disease perception data matrix to obtain a set of trend vectors, and to generate false probability vectors and cross-indicator correlation matrices through short-term fluctuation and long-term trend deviation analysis. The cross-disease dynamic anomaly identification module is used to perform threshold dynamic modeling through cross-indicator correlation matrix, correct the initial anomaly screening markers and construct a multi-dimensional risk indicator matrix; The cross-disease risk analysis module is used to perform component expansion, directional unification, and time-series cumulative analysis on the multidimensional risk indicator matrix, generate an indicator contribution matrix, and obtain a multidimensional risk score matrix after correlation constraint correction. The intervention prediction module is used to read historical intervention records, combine them with a multidimensional risk score matrix, construct a risk contribution prediction sequence by matching intervention records and extrapolating contribution change paths, and determine the optimal intervention based on time stability analysis.

Images