Multivariate adaptive weights based medical data anomaly monitoring method and system

Abstract

This invention discloses a method and system for monitoring medical data anomalies using multivariate adaptive weights, relating to the fields of medical artificial intelligence and data analysis. The method includes: acquiring a standard medical data matrix of historical medical data samples and new observation samples, the standard medical data matrix including a first feature; updating the initial weights of the first feature based on the residuals of the first feature to obtain a first weight; and obtaining T based on the first weights and the new observation samples. 2 Statistics; the Q-statistic is obtained based on the standard medical data matrix and the new observation sample; the contribution value of the Q-statistic is obtained based on the residual vector of the Q-statistic and the first weight; the T-statistic is obtained based on the historical medical data sample. 2 Dynamic control threshold and Q dynamic control threshold; if T 2 The statistic is greater than T 2 If the dynamic control threshold or Q statistic is greater than the Q dynamic control threshold, the new observation sample is determined to be abnormal. Abnormal data is obtained based on the contribution value, which solves the problems of adaptive deviation of control limits and fixed variable contribution in the existing MSPC method.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence and data analysis, specifically to a method and system for monitoring medical data anomalies using multivariate adaptive weighting. Background Technology

[0002] With the development of medical informatization, hospitals have generated massive amounts of patient data, including vital signs, laboratory indicators, and real-time monitoring data. This data is characterized by high dimensionality, strong temporal sequence, and complex correlations between variables. Currently, anomaly monitoring in the medical field mainly employs multivariate statistical process control (MSPC) methods; however, existing MSPC methods face the following technical bottlenecks:

[0003] The rigidity of static MSPC models: Although traditional MSPC methods can integrate multivariate information, their control limits are usually static thresholds set once based on historical data. This cannot adapt to the differences in physiological baselines of different departments (such as ICU and general ward), different patient groups (such as the elderly and children), or individual patients, resulting in high false alarm and false negative rates.

[0004] Fixed variable contributions: Once the MSPC model is established, the contribution weight of each variable to the overall statistic is fixed, and it cannot reflect the dynamic changes in the importance of each physiological parameter under different patient conditions or different stages of disease. Summary of the Invention

[0005] To address the issues of adaptive error of control limits and fixed variable contributions in existing medical data anomaly detection methods, this invention provides a multivariate adaptive weighting method for medical data anomaly monitoring, the method comprising:

[0006] Acquire historical medical data samples from different patient groups and new observation samples from individual patients. Construct a historical medical data matrix based on the historical medical data samples. Standardize the historical medical data matrix to obtain a standard medical data matrix. The standard medical data matrix includes several first features.

[0007] The initial weights of the first feature are updated based on the residuals of the first feature to obtain the first weights;

[0008] T is obtained based on the first weight and the new observation sample. 2 Statistic;

[0009] A feature vector matrix is ​​obtained based on the standard medical data matrix. A reconstructed value is obtained based on the new observation sample and the feature vector matrix. A Q statistic is obtained based on the new observation sample and the reconstructed value. The Q statistic is used to quantify the severity of the new observation sample deviating from the normal state.

[0010] Based on the residual vector of the Q statistic and the first weight, the contribution value of the Q statistic is obtained;

[0011] Historical T was obtained based on the aforementioned historical medical data samples. 2 The observed value sequence and the historical Q observed value sequence, based on the historical T 2 The observed value sequence and the historical Q observed value sequence are used to obtain T respectively. 2 Dynamic control threshold and Q-dynamic control threshold;

[0012] If the T 2 The statistic is greater than the stated T 2 If the dynamic control threshold is reached, or if the Q statistic is greater than the Q dynamic control threshold, the new observation sample is determined to be abnormal, and abnormal data is obtained based on the contribution value.

[0013] Traditional weighting schemes rely on prior experience and set static weights for features, requiring frequent manual adjustments. This can lead to delayed responses to critical indicators, masking of sudden abnormal parameters due to low weights, and noise-sensitive imbalances. High-weighted parameters are significantly less accurate when disturbed. Therefore, this method introduces adaptive weights for features to reflect their dynamic contribution to overall abnormalities and the dynamic changes in the importance of each physiological parameter under different patient conditions or disease stages. This can improve the critical condition recognition rate, overcome the bottleneck of fixed weights in perceiving sudden abnormalities, and enhance robustness against interference. Furthermore, it innovatively incorporates weighted T... 2 The statistic (Hotling's statistic) solves the problems of traditional T... 2 Clinical limitations of statistical measures, such as the different indicative values ​​of different parameters for pathological states, and the corresponding weights should also be different, such as weighted T. 2 Statistics can reduce false alarm rates, increase anomaly detection rates, and enhance adaptability and anti-interference capabilities. Furthermore, the innovative introduction of an adaptive weighted contribution decomposition algorithm decomposes the contributions of each variable according to adaptive weights. Simultaneously, a dynamic control limit based on quantiles is employed to adapt to the physiological baseline differences of different patient groups or individuals, reducing false alarm and false negative rates. This enables automatic identification of dominant variables and personalized anomaly discrimination, improving system intelligence and clinical usability. Abnormal parameters are obtained based on contribution values, supporting automatic location of dominant abnormal variables and enhancing anomaly interpretability and source tracing efficiency.

[0014] Furthermore, the first calculation formula for obtaining the first weight is:

[0015] ;

[0016] in, Indicates the first The first feature is The first weight of time, Indicates the first The first feature is The first weight of time, Represents the smoothing coefficient. Indicates the first The first feature is Time residuals Indicates the first The first feature is Time residuals Indicates the number of the first feature. and Both represent integers greater than or equal to 1. Indicates the time.

[0017] Furthermore, obtain the T 2 The second formula for calculating the statistic is:

[0018] ;

[0019] in, T represents 2 Statistic, Indicates a new observation sample. This represents the average value of a sample of historical medical data. This represents the adaptive weight diagonal matrix. Represents the covariance matrix. This indicates transpose.

[0020] Furthermore, the third calculation formula for obtaining the reconstructed value is:

[0021] ;

[0022] The fourth formula for obtaining the Q statistic is:

[0023] ;

[0024] in, Indicates the reconstructed value. Represents the eigenvector matrix, This represents the average value of a standard medical data matrix. This represents the standard deviation of a standard medical data matrix. This represents the Q statistic.

[0025] Furthermore, the fifth calculation formula for obtaining the residual vector is:

[0026] ;

[0027] The sixth formula for obtaining the contribution value is:

[0028] ;

[0029] in, Indicates the first Each residual vector Indicates the first The contribution value of the first feature, Indicates the first The first weight of the first feature.

[0030] Furthermore, obtain the T 2 The seventh formula for calculating the dynamic control threshold is:

[0031] ;

[0032] The eighth calculation formula for obtaining the Q dynamic control threshold is:

[0033] ;

[0034] in, Indicates the first T in a patient group 2 Dynamically control thresholds, Represents the quantile function. Indicates the confidence level. Indicates the first The first patient group The history of each feature T 2 Observation sequence, Indicates the first Q dynamic control threshold for each patient group Indicates the first The first patient group The historical Q-observation sequence of each feature and Both represent integers greater than or equal to 1.

[0035] Furthermore, the method also includes:

[0036] Obtain the new observation sample that is abnormal, and obtain the abnormal sample;

[0037] Pre-set patient metadata, cluster the historical medical data sample based on the patient metadata to obtain several state clusters, obtain several state models based on the state clusters, obtain state data of the state models, the state data includes benchmark mean vector, covariance matrix and principal component loading matrix, and obtain a model library based on all the state models.

[0038] The abnormal samples are matched with the model library to obtain a reference model;

[0039] Based on the abnormal samples and the reference model, the comprehensive deviation and residual abnormality are obtained, and the exceedance index is obtained based on the comprehensive deviation and the residual abnormality.

[0040] Obtain the contribution of each second feature in the abnormal sample to the exceeding index;

[0041] Abnormal information is obtained based on the aforementioned contribution level and the clinical event database;

[0042] The specific steps for obtaining the clinical event database include:

[0043] The process involves acquiring structured medical data and patient medical record data, whereby the structured medical data includes standard medical ontology data, clinical guideline data, and medical textbook data; obtaining structured data based on the structured medical data, and obtaining triples based on the patient medical record data, wherein the structured data includes clinical entities, structured pathways, and key relationships between the clinical entities, and the triples include medical parameters, direction of change, and clinical events;

[0044] A medical knowledge graph is constructed based on the structured data and the triples. The medical knowledge graph includes several nodes and edges. The nodes include parameter nodes, change direction nodes, and clinical event nodes. Based on the medical knowledge graph, the path from the parameter node to the clinical event node is traversed, and seed rules are obtained based on the path.

[0045] Time-series data is obtained based on the historical medical data sample, and data-driven rules are obtained based on the time-series data. The time-series data includes parameter contribution increase sequence patterns, high contribution parameters, and association rules between the high contribution parameters.

[0046] Based on the seed rules, the data-driven rules, and the preset knowledge base, event rules are obtained, and the clinical event database is obtained based on the event rules.

[0047] Traditional bedside monitors set independent static thresholds (alarm upper and lower limits) for each physiological parameter. This method fails to capture the synergistic relationships between multiple parameters. For example, a slight decrease in blood pressure and a significant increase in heart rate may not exceed their respective thresholds from a univariate perspective, but their synergistic pattern is a crucial indicator of early shock. This method, by capturing synergistic changes in multiple parameters, can identify early, hidden risks that univariate methods cannot detect. Furthermore, it reduces false alarms through a personalized model and transforms abstract multivariate abnormalities into concrete parameter changes understandable to clinicians through contribution analysis, significantly enhancing the clinical value and reliability of alarms and achieving accurate detection of multivariate synergistic abnormalities.

[0048] Currently, the rule bases (i.e., clinical event bases) of existing clinical alarm systems are mostly built based on the following two methods:

[0049] Manual summaries based on medical textbooks and clinical guidelines: IF-THEN rules are manually written by experts. While authoritative, this method is static, slowly updated, and unable to adapt to new medical discoveries or the characteristics of specific hospital patient groups, and it struggles to cover all complex multi-parameter collaborative scenarios. Simple combinations based on fixed thresholds: e.g., heart rate > 100 & blood pressure < 90. However, this method is too rigid, unable to capture subtle, dynamically changing collaborative patterns, and the subjective nature of threshold setting can easily lead to alarm fatigue.

[0050] This method constructs a medical knowledge graph, making explicit and structured the tacit knowledge of human experts. It traverses the knowledge graph to generate seed rules; mines data-driven patterns, utilizing association rules and sequence patterns to discover new patterns hidden in the data beyond textbooks, uncovering complex and dynamic pathophysiological patterns that human experts have not yet summarized or are difficult to quantify precisely; and integrates both methods with a continuous learning mechanism to construct a hybrid framework for real-world data mining guided by the knowledge graph. This framework combines the structuring of authoritative medical knowledge with the discovery of new patterns driven by data, generating highly reliable clinical event rules in stages and automatically. This transforms the process of generating a clinical event database from a potentially observable... The concept of a simple, artificial rule base is elevated to a dynamic, self-learning, data-driven intelligent knowledge discovery system, thereby constructing a dynamically evolving, evidence-driven, and personalized clinical event database. By combining symbolic AI (knowledge graph) with connectionist / statistical learning (data mining) and introducing a feedback loop, the event database is no longer static but can be self-updated and optimized through continuous data injection and clinical feedback. This enables it to autonomously discover and learn medical knowledge from data, and each rule clearly indicates its knowledge source (guidelines, literature, in-house data), greatly enhancing clinicians' trust and providing strong interpretability.

[0051] Furthermore, the specific steps to obtain the reference model include:

[0052] Obtain the Mahalanobis distance between the abnormal samples and each state model, and obtain the reference model based on the minimum Mahalanobis distance;

[0053] The ninth formula for obtaining the overall deviation is:

[0054] ;

[0055] in, Indicates the overall deviation. Indicates an abnormal sample. Represents the baseline mean vector. Indicates transpose. This represents the contribution weight of the reference model. Represent the covariance matrix;

[0056] The tenth formula for obtaining the residual abnormality is:

[0057] ;

[0058] in, Indicates the degree of residual abnormality. Represents the identity matrix. Represents the principal component loading matrix;

[0059] The eleventh formula for calculating the contribution is as follows:

[0060] ;

[0061] in, Indicates contribution level. Indicates partial derivative, Indicates the index exceeding the standard. Indicates the first A second characteristic, Represents an integer greater than or equal to 1.

[0062] Furthermore, the specific steps for obtaining the contribution weight include:

[0063] Based on the state model, several principal components are obtained, the variance explained rate of the principal components is obtained, the absolute value of the loading of each principal component is obtained based on the principal component loading matrix, and the basic weight of the second feature is obtained based on the absolute value of the loading and the variance explained rate.

[0064] Obtain the state value of the principal component, obtain the adjustment factor based on the state value, and obtain the contribution weight based on the adjustment factor and the basic weight;

[0065] The twelfth formula for obtaining the basic weights is:

[0066] ;

[0067] in, Indicates the first The basic weights of the second feature, Indicates the number of principal components. Represents the principal component loading matrix of the th The second feature in the first The absolute value of the loading of each principal component. Indicates the first The variance explained by each principal component;

[0068] The thirteenth formula for obtaining the state value is:

[0069] ;

[0070] in, Indicates the first The second feature is The state value at time t, Indicates the attenuation factor. Indicates the first The second feature is The state value at time t, Indicates the first The second feature is The time-based exception indicator function;

[0071] The fourteenth formula for obtaining the adjustment factor is:

[0072] ;

[0073] in, Indicates the first The second feature is Adjustment factor for time, Indicates the magnification factor;

[0074] The fifteenth formula for obtaining the contribution weight is:

[0075] ;

[0076] in, Indicates the first The contribution weight of each second feature.

[0077] Introducing a dynamic adjustment factor enables the model to adaptively adjust the importance of variables based on recent anomalies, thereby more sensitively detecting co-occurring anomalies related to recent anomalies. It also allows for adaptive adjustment of monitoring priorities based on recent anomaly patterns, improving sensitivity to persistent anomaly patterns while maintaining model interpretability. Furthermore, because the dynamic adjustment factor decays, the model is not perpetually dominated by past anomalies, exhibiting self-recovery capabilities.

[0078] This invention also provides a multivariate adaptive weighting medical data anomaly monitoring system, the system comprising:

[0079] Data Unit: Used to acquire historical medical data samples of different patient groups and new observation samples of individual patients, construct a historical medical data matrix based on the historical medical data samples, standardize the historical medical data matrix to obtain a standard medical data matrix, the standard medical data matrix includes several first features;

[0080] Calculation unit: used to update the initial weight of the first feature based on the residual of the first feature, and obtain the first weight;

[0081] And for obtaining T based on the first weight and the new observation sample 2 Statistic;

[0082] And for obtaining a feature vector matrix based on the standard medical data matrix, obtaining a reconstructed value based on the new observation sample and the feature vector matrix, and obtaining a Q statistic based on the new observation sample and the reconstructed value, wherein the Q statistic is used to quantify the severity of the new observation sample deviating from the normal state;

[0083] And to obtain the contribution value of the Q statistic based on the residual vector of the Q statistic and the first weight;

[0084] Threshold unit: used to obtain historical T based on the historical medical data sample. 2 The observed value sequence and the historical Q observed value sequence, based on the historical T 2 The observed value sequence and the historical Q observed value sequence are used to obtain T respectively. 2 Dynamic control threshold and Q-dynamic control threshold;

[0085] Analysis unit: used if the T 2 The statistic is greater than the stated T 2 If the dynamic control threshold is reached, or if the Q statistic is greater than the Q dynamic control threshold, the new observation sample is determined to be abnormal, and abnormal data is obtained based on the contribution value.

[0086] The principle and effect of this system are similar to those of this method, so no further details will be provided for this system.

[0087] One or more technical solutions provided by this invention have at least the following technical effects or advantages:

[0088] 1. This method introduces adaptive weights for features, reflecting their dynamic contribution to overall abnormalities and the dynamic changes in the importance of each physiological parameter under different patient conditions or stages of disease. This can improve the identification rate of critical states, overcome the bottleneck of fixed weights in perceiving sudden abnormalities, and enhance robustness against interference; and it also innovatively incorporates weighted T... 2 Statistics to solve the problems of traditional T 2 Clinical limitations of statistical measures, such as the different indicative values ​​of different parameters for pathological states, and the corresponding weights should also be different, such as weighted T. 2Statistics can reduce false alarm rates, increase anomaly detection rates, and enhance adaptability and anti-interference capabilities. Furthermore, the innovative introduction of an adaptive weighted contribution decomposition algorithm decomposes the contributions of each variable according to adaptive weights. Simultaneously, a dynamic control limit based on quantiles is employed to adapt to the physiological baseline differences of different patient groups or individuals, reducing false alarm and false negative rates. This enables automatic identification of dominant variables and personalized anomaly discrimination, improving system intelligence and clinical usability. Abnormal parameters are obtained based on contribution values, supporting automatic location of dominant abnormal variables and enhancing anomaly interpretability and source tracing efficiency.

[0089] 2. This method can identify early and hidden risks that cannot be detected by univariate methods by capturing the synergistic changes of multiple parameters. At the same time, it reduces false alarms through personalized models and transforms abstract multivariate anomalies into specific parameter changes that clinicians can understand through contribution analysis, which greatly improves the clinical value and credibility of alarms and achieves accurate detection of multivariate synergistic anomalies.

[0090] 3. This method constructs a medical knowledge graph, making explicit and structured the tacit knowledge of human experts. It traverses the knowledge graph to generate seed rules; mines data-driven patterns, utilizing association rules and sequence patterns to discover new patterns beyond textbooks and hidden in the data, uncovering complex and dynamic pathophysiological patterns that human experts have not yet summarized or are difficult to quantify precisely; and integrates both methods with a continuous learning mechanism to construct a hybrid framework for real-world data mining guided by the knowledge graph. This framework combines the structuring of authoritative medical knowledge with the discovery of new patterns driven by data, generating highly reliable clinical event rules in stages and automatically, thus realizing the transformation of the clinical event database generation process from a possible... The concept, previously considered a simple artificial rule base, has been elevated to a dynamic, self-learning, data-driven intelligent knowledge discovery system, thereby constructing a dynamically evolving, evidence-driven, and personalized clinical event database. By combining symbolic AI (knowledge graph) with connectionist / statistical learning (data mining) and introducing a feedback loop, the event database is no longer static but can be self-updated and optimized through continuous data injection and clinical feedback. This enables it to autonomously discover and learn medical knowledge from data, and each rule clearly indicates its knowledge source (guidelines, literature, in-house data), greatly enhancing clinicians' trust and providing strong interpretability.

[0091] 4. Introducing a dynamic adjustment factor enables the model to adaptively adjust the importance of each variable based on recent anomalies, thereby more sensitively detecting co-occurring anomalies related to recent anomalies. It also allows for adaptive adjustment of monitoring priorities based on recent anomaly patterns, improving sensitivity to persistent anomaly patterns while maintaining model interpretability. Furthermore, because the dynamic adjustment factor decays, the model is not perpetually dominated by past anomalies, exhibiting self-recovery capabilities. Attached Figure Description

[0092] The accompanying drawings, which are provided to further illustrate embodiments of the invention and constitute a part of this invention, are not intended to limit the scope of the invention.

[0093] Figure 1 This is a flowchart illustrating the medical data anomaly monitoring method with multivariate adaptive weights in this invention. Detailed Implementation

[0094] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0095] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0096] Example 1

[0097] refer to Figure 1 This embodiment provides a method for monitoring medical data anomalies using multivariate adaptive weights, the method comprising:

[0098] Obtain historical medical data samples from different patient groups and new observation samples from individual patients. Construct a historical medical data matrix based on the historical medical data samples. Let the historical medical data matrix be... , For the sample size, The characteristic number;

[0099] The historical medical data matrix is ​​standardized to obtain a standard medical data matrix. Let the standard medical data matrix be... The standardized processing formula can be:

[0100] ;

[0101] in, and The first In the nth sample The original feature values ​​and the standardized feature values, The first The mean and standard deviation of each feature; in this embodiment, existing sliding window methods or recursive updates can also be used to dynamically estimate. To adapt to changes in data distribution.

[0102] The standard medical data matrix includes several first features, which may include vital signs, laboratory indicators, monitoring data, such as RR interval, QRS amplitude, ST segment slope, systolic blood pressure (SBP), diastolic blood pressure (DBP), blood oxygen saturation, body temperature, hemoglobin (Hb), white blood cell count (WBC), platelet count (PLT), blood glucose (GLU), creatinine (Cr), blood urea nitrogen (BUN), etc.

[0103] The initial weights of the first feature are updated based on the residuals of the first feature to obtain the first weights; for example, the initial weights can be:

[0104] ;

[0105] in, Indicates the initial weights;

[0106] The initial weights are then recursively updated, and the first calculation formula for obtaining the first weight is:

[0107] ;

[0108] in, Indicates the first The first feature is The first weight of time, Indicates the first The first feature is The first weight of time, Represents the smoothing coefficient (0 < <1), Indicates the first The first feature is Time residuals Indicates the first The first feature is Time residuals Indicates the number of the first feature. and Both represent integers greater than or equal to 1. Indicates time, , .

[0109] And the first weight can be normalized:

[0110] ;

[0111] T is obtained based on the first weight and the new observation sample. 2 The statistic (i.e., the Hotelling statistic); wherein, the T is obtained. 2 The second formula for calculating the statistic is:

[0112] ;

[0113] in, T represents 2 Statistic, Indicates a new observation sample. This represents the average value of a sample of historical medical data. , represents the adaptive weight diagonal matrix, Represents a diagonal matrix. This represents the diagonal element, i.e., the first weight. Represents the covariance matrix. This indicates transpose.

[0114] A feature vector matrix is ​​obtained based on the standard medical data matrix, and a reconstructed value is obtained based on the new observation sample and the feature vector matrix. The third calculation formula for obtaining the reconstructed value is as follows:

[0115] ;

[0116] The Q statistic is obtained based on the new observation sample and the reconstructed value. The Q statistic is used to quantify the severity of the deviation of the new observation sample from the normal state. The fourth formula for obtaining the Q statistic is:

[0117] ;

[0118] in, Indicates the reconstructed value. Represents the eigenvector matrix, To pass principal component analysis The matrix formed by the first d eigenvectors, where d is an integer greater than or equal to 1. This represents the average value of a standard medical data matrix. This represents the standard deviation of a standard medical data matrix. This represents the Q statistic.

[0119] Based on the residual vector of the Q statistic and the first weight, the contribution value of the Q statistic is obtained; wherein, the fifth calculation formula for obtaining the residual vector is:

[0120] ;

[0121] The sixth formula for calculating the contribution value using adaptive weights is as follows:

[0122] ;

[0123] in, Indicates the first Each residual vector Indicates the first The contribution value of the first feature, Indicates the first The first weight of the first feature.

[0124] Historical T was obtained based on the aforementioned historical medical data samples. 2 Observational series and historical Q observational series, such as for different patient groups Establish historical normal datasets respectively Within a sliding window or periodic batch, obtain T. 2 Historical observations of the T statistic and Q statistic are used to obtain historical T values. 2 Observational sequence and historical Q-observational sequence;

[0125] Based on the aforementioned historical T 2 The observed value sequence and the historical Q observed value sequence are used to obtain T respectively. 2 Dynamic control threshold and Q-dynamic control threshold;

[0126] Wherein, the T is obtained 2 The seventh formula for calculating the dynamic control threshold is:

[0127] ;

[0128] The eighth calculation formula for obtaining the Q dynamic control threshold is:

[0129] ;

[0130] in, Indicates the first T in a patient group 2 Dynamically control thresholds, Represents the quantile function. Indicates the confidence level. Indicates the first The first patient group The history of each feature T 2 Observation sequence, Indicates the first Q dynamic control threshold for each patient group Indicates the first The first patient group The historical Q-observation sequence of each feature and Both represent integers greater than or equal to 1.

[0131] If the T 2 The statistic is greater than the stated T 2If the dynamic control threshold is reached, or if the Q statistic is greater than the Q dynamic control threshold, the new observation sample is determined to be abnormal. Abnormal data is obtained based on the contribution value, such as taking the feature with a contribution value greater than a certain threshold as abnormal data.

[0132] Example 2

[0133] Based on Embodiment 1, in this embodiment, the method further includes:

[0134] Obtain the new observation sample that is abnormal, and obtain the abnormal sample;

[0135] Using machine learning methods and clustering algorithms (such as K-means, DBSCAN), pre-defined patient metadata (which may include department, disease, and patient group characteristics such as age and underlying diseases) is used. Based on the patient metadata, the historical medical data samples are clustered to obtain several state clusters. Based on the state clusters, several state models are obtained. For example, a state model is trained for each typical clinical state (such as stable period, infection risk period, pre-circulatory failure period, etc.) (e.g., using principal component analysis to establish a PCA model). Each state model not only includes the baseline mean and covariance matrix of physiological parameters in that clinical state, but also dynamically quantifies the contribution weight of each parameter to the overall state of the model. The state data of the state models is obtained, including the baseline mean vector, covariance matrix, and principal component loading matrix. A model library is obtained based on all the state models.

[0136] The abnormal samples are matched with the model library to obtain a reference model; the multidimensional data of the current patient are quickly matched with the model library, and the best matching benchmark model is selected as the reference model for the current state.

[0137] Based on the abnormal samples and the reference model, the comprehensive deviation and residual abnormality are obtained. Based on the comprehensive deviation and residual abnormality, the exceedance index is obtained. If either the comprehensive deviation or the residual abnormality exceeds a preset threshold, it is considered an exceedance index. When the exceedance index exceeds the control limit, the contribution decomposition algorithm is activated to accurately calculate which one or more physiological parameters and their direction of change contribute the most to this abnormality.

[0138] The contribution of each second feature in the abnormal sample to the exceedance index is obtained. In this embodiment, the second feature may include vital signs, test indicators, monitoring data, such as RR interval, QRS amplitude, ST segment slope, systolic blood pressure (SBP), diastolic blood pressure (DBP), blood oxygen saturation, body temperature, hemoglobin (Hb), white blood cell count (WBC), platelet count (PLT), blood glucose (GLU), creatinine (Cr), blood urea nitrogen (BUN), etc.

[0139] Abnormal information is obtained based on the contribution level and the clinical event database; the contribution levels are sorted and mapped to a predefined clinical event database to generate abnormal information, such as an increase in heart rate and a decrease in blood pressure, which may indicate a risk of shock.

[0140] The specific steps for obtaining the reference model include:

[0141] Obtain the Mahalanobis distance between the abnormal samples and each state model, and obtain the reference model based on the minimum Mahalanobis distance;

[0142] The ninth formula for obtaining the overall deviation is:

[0143] ;

[0144] in, Indicates the overall deviation. Indicates an abnormal sample. Represents the baseline mean vector. Indicates transpose. The contribution weights of the reference model are represented by a diagonal weight matrix, which amplifies the sensitivity to key parameters. Represent the covariance matrix;

[0145] The tenth formula for obtaining the residual abnormality is:

[0146] ;

[0147] in, Indicates the degree of residual abnormality. Represents the identity matrix. Represents the principal component loading matrix;

[0148] The eleventh formula for calculating the contribution is as follows:

[0149] ;

[0150] in, Indicates contribution level. Indicates partial derivative, Indicates the index exceeding the standard. Indicates the first A second characteristic, Represents an integer greater than or equal to 1.

[0151] The specific steps for obtaining the contribution weight include:

[0152] Using principal component analysis, several principal components are obtained based on the state model. The variance explained by each principal component (i.e., the proportion of the eigenvalues ​​of that principal component to the total eigenvalues) is obtained. The absolute value of the loading of each principal component is obtained based on the principal component loading matrix. Let the principal component loading matrix be... , Indicates the first The second feature in the first The load of each principal component is the absolute value of the load. The number of principal components is represented, and the basic weights of the second feature are obtained based on the absolute value of the loading and the variance explained rate.

[0153] The twelfth formula for obtaining the basic weights is:

[0154] ;

[0155] in, Indicates the first The basic weights of the second feature, Indicates the number of principal components. Represents the principal component loading matrix of the th The second feature in the first The absolute value of the loading of each principal component. Indicates the first The variance explained by each principal component;

[0156] The baseline weights are entirely data-driven, objectively reflecting the inherent importance of each physiological parameter in defining normal variation patterns under specific clinical conditions (such as the stable phase). Furthermore, the differential importance of parameters is embedded in the model initialization phase, laying the foundation for subsequent dynamic adjustments.

[0157] Obtain the state value of the principal component, obtain the adjustment factor based on the state value, and obtain the contribution weight based on the adjustment factor and the basic weight;

[0158] Maintain a state variable for each patient and each physiological parameter, i.e. First, initialize: ;

[0159] At each monitoring time, for the parameters Calculate its standardized univariate residuals:

[0160] ;

[0161] in, Indicates the first The second feature is Standardized univariate residuals at time 10:00 Indicates the first The second feature is The actual observed value at time [time]. Indicates the first The second feature is The model's predicted (or reconstructed) value at time 1. Indicates the first The historical standard deviation of the residuals of the second characteristic;

[0162] like If it exceeds the preset threshold, then = ,otherwise =0;

[0163] Introducing a decay mechanism, the thirteenth formula for calculating the state value is:

[0164] ;

[0165] in, Indicates the first The second feature is The state value at time t, Indicates the attenuation factor. Indicates the first The second feature is The state value at time t, Indicates the first The second feature is The time-based exception indicator function;

[0166] This constitutes an abnormal memory; once a certain parameter becomes abnormal, its memory state will be... It will rise immediately and remain at a high level for a period of time before gradually declining, mimicking the thinking of clinicians: a parameter that has just fluctuated, even if it is temporarily normal, still requires high vigilance; using standardized residuals instead of just Boolean values ​​allows the severity of the abnormality to be quantified and incorporated into memory, with different memory strengths for slight fluctuations and severe abnormalities; and The recursive update formula has a very small computational cost, making it suitable for real-time streaming data processing.

[0167] The fourteenth formula for obtaining the adjustment factor is:

[0168] ;

[0169] in, Indicates the first The second feature is Adjustment factor for time, This represents the amplification factor, used to control the magnitude of dynamic adjustment;

[0170] The fifteenth formula for obtaining the contribution weight is:

[0171] ;

[0172] in, Indicates the first The contribution weight of each second feature.

[0173] Creatively integrate long-term static importance With short-term abnormal memory Multiplication, integrating long-term patterns with short-term memory, significantly amplifies the dynamic weight of a parameter that is inherently important and has recently exhibited abnormalities, enabling proactive monitoring. For example, during the development of shock, the heart rate may initially increase abnormally before blood pressure begins to decrease. When the heart rate first becomes abnormal, its dynamic weight is increased. This is then applied in subsequent calculations of the overall deviation. At this time, the system becomes more sensitive to any subsequent minute changes in heart rate. Therefore, when blood pressure begins to show a slight downward trend, the weak synergistic abnormal signal between the two can be captured earlier and more reliably because the focus is on heart rate, thus enabling earlier warning of shock; and the entire dynamic weight adjustment process is fully automated, without the need for clinicians to manually set or adjust thresholds, adapting to the personalized monitoring needs of different patients and different stages of disease.

[0174] The specific steps for obtaining the clinical event database include:

[0175] The process involves acquiring structured medical data and patient medical record data, including standard medical ontology data (such as SNOMED CT), clinical guideline data, and medical textbook data. Based on this structured medical data, structured data is obtained, comprising clinical entities, structured paths, and key relationships between these entities. For example, plain text extraction from clinical guidelines is performed using a PDF parsing library (such as PyMuPDF), and relationship extraction is achieved using pre-trained biomedical language models (such as BioBERT and ClinicalBERT) for named entity recognition. Computer vision libraries (such as OpenCV) are used to detect flowcharts, tables, and lists in the documents. For detected flowcharts, graphic element segmentation is performed, identifying rhombuses (decision boxes), rectangles (process / action boxes), arrows (flow directions), and OCR technology (such as PaddleOCR) is used to extract text within the graphic elements. This paper aligns and complements the flowchart structure extracted from the visual layer with the logical relationships extracted from the text layer to form a preliminary logical diagram with a geometric layout. Based on keywords (such as yes / no, presence / absence, evaluation, diagnosis) and graphical elements (diamonds), all decision points are identified, and based on keywords (such as give, monitor, suggest, refer) and graphical elements (rectangles), all action nodes are identified. Using graph theory algorithms, fragmented decision elements are automatically assembled into complete decision paths. For example, using decision nodes and action nodes as vertices, directed edges are established between nodes based on the flow direction (arrows) extracted from the text and flowchart. The direction of the edges represents the flow of decisions, and a directed graph is constructed. Starting from the entry point of the graph, a depth-first search algorithm is used to traverse all possible paths to obtain a structured path.

[0176] By using natural language processing to extract entities and key relationships from medical textbook data, clinical entities and key relationships between them are obtained. Using NLP techniques (such as BERT, RE), triples are obtained based on the patient medical record data. The triples include medical parameters, direction of change, and clinical events. For example, from the text: "The patient experienced an increased heart rate and decreased blood pressure, suggesting possible shock," two triples are extracted: (heart rate, increased, shock) and (blood pressure, decreased, shock).

[0177] A medical knowledge graph is constructed based on the structured data and the triples. The medical knowledge graph includes several nodes and edges. The nodes include parameter nodes (such as physiological parameters, drugs, diseases, examination and test indicators, etc.), change direction nodes (such as increase, decrease, and no change, etc.), and clinical event nodes (such as clinical symptoms such as shock and asphyxia). The edges are used to define the relationships between nodes, such as causing, accompanying, being a symptom of, being a cause of, antagonizing, and synergistic.

[0178] Caused: indicates that one entity (such as severe infection) is a sufficient or necessary condition for another entity (such as septic shock).

[0179] Symptoms of...: Connecting signs / parameters (such as shortness of breath) with pathological states (such as acute respiratory distress syndrome). This relationship includes specificity and sensitivity attributes used to assess the value of the symptom in diagnosing the state.

[0180] Accompanying: This indicates that two entities often occur simultaneously in clinical practice, but there is not necessarily a clear causal relationship (such as hypokalemia and arrhythmia).

[0181] Antagonism: This refers to the effect of one entity (such as the use of vasopressors) counteracting the effect of another entity (such as hypotension). This is crucial in multiparameter monitoring to interpret paradoxical inconsistencies under therapeutic interventions.

[0182] Synergy: This means that when two entities (such as increased heart rate and decreased myocardial contractility) act together, their combined effect on a clinical event (such as decreased cardiac output) is greater than the sum of their individual effects.

[0183] Based on the medical knowledge graph, the path from the parameter node to the clinical event node is traversed, that is, all paths from the parameter combination node to the clinical event node are traversed. Seed rules are obtained based on the path. For example, starting with heart rate, the nodes connected to it are obtained, and mean arterial pressure and shock are obtained. The relationship between heart rate and the two is obtained according to the edge. Heart rate and mean arterial pressure are associated, and heart rate and shock are causative. Then the direction of change of heart rate and mean arterial pressure is obtained, thereby generating seed rule 1: IF (heart rate: significantly increased) AND (mean arterial pressure: continuously decreased) THEN suspected: shock (knowledge source: clinical guidelines, confidence: high).

[0184] Similarly, we can derive Seed Rule 2: IF (lactic acid: increased) AND (capillary refill time: prolonged) THEN suspected: shock (Knowledge source: literature mining, confidence level: medium).

[0185] Time-series data is obtained based on the historical medical data samples, and data-driven rules are derived from this time-series data. The time-series data includes increasing parameter contribution sequence patterns, high-contribution parameters, and association rules between these high-contribution parameters. For example, algorithms (such as PrefixSpan) are used to find frequently occurring increasing parameter contribution sequence patterns in patient data experiencing adverse events (such as shock). For instance, it was found that in the two hours before shock, a common pattern emerged: a sustained high heart rate contribution, followed by a rise in blood pressure contribution, and then a sudden spike in lactate contribution. Each sequence is an ordered list consisting of parameters, contribution rankings, and trends. Using algorithms such as FP-Growth, frequently co-occurring high-contribution parameter combinations are mined within the time window before the abnormal event. For example, a strong association rule was found: {Top 1 contribution: increased heart rate, Top 2 contribution: decreased blood pressure, Top 3 contribution: decreased urine output} – shock.

[0186] Transform time-series data into data-driven rules, such as pattern element parsing:

[0187] Each element is parsed into the following structure: physiological parameter name, direction of change, contribution, and time constraint.

[0188] Physiological parameter names, such as heart rate and blood pressure;

[0189] Direction of change, such as rising, falling, or no change;

[0190] Contribution can be a specific numerical value or a ranking, such as Top 1 or Top 2.

[0191] For time constraints, there is a specific order between elements in sequence patterns; for association rules, the elements must be within the same time window.

[0192] Rule condition construction:

[0193] For sequence patterns, each element in the sequence is transformed into a condition while preserving its order. For example, a sequence pattern of [(heart rate, high contribution), (blood pressure, high contribution)] can be transformed into two conditions and required to occur in sequence, such as the heart rate contribution entering the top 3 occurring within a certain number of minutes before the blood pressure contribution enters the top 3.

[0194] For association rules, each item in the itemset is transformed into a condition, and these conditions occur simultaneously.

[0195] Rule-based conclusion mapping:

[0196] The conclusion part of the data-driven rule (i.e. the mined clinical event) is mapped to the clinical event node in the knowledge graph. If the data mining uses labeled data (such as the known patient has experienced shock), then the conclusion is the clinical event. If it is unsupervised mining, the pattern can be classified into the known clinical event or marked as a new clinical event pattern through clustering or other methods.

[0197] Rule quantification and threshold setting:

[0198] Set a threshold for each parameter in the rule conditions to define the direction of change.

[0199] Formal representation of rules:

[0200] Finally, the rules are represented in a machine-readable standard format, such as SWRL (Semantic Web Rule Language) rules, or a custom JSON format, to generate data-driven rules, such as the rule: IF (contribution pattern: heart rate & blood pressure, synergistic abnormality) AND (urine volume: contribution enters the top three) THEN warning: shock (data source: ICU data of our hospital, confidence level: 92%).

[0201] Based on the seed rules, the data-driven rules, and the preset knowledge base, event rules are obtained, and the clinical event database is obtained based on the event rules. If the seed rules and data-driven rules are input into a unified knowledge base, when rules conflict (e.g., the seed rules consider A important, while the data rules consider B important), a weighted scoring mechanism can be established. Weighting factors include: data source authority, confidence level, and time freshness. This mechanism determines whether A or B is more important, and a new rule with higher overall confidence is automatically generated.

[0202] Example 3

[0203] Based on the above embodiments, this embodiment also provides a medical data anomaly monitoring system with multivariate adaptive weights, the system comprising:

[0204] Data Unit: Used to acquire historical medical data samples of different patient groups and new observation samples of individual patients, construct a historical medical data matrix based on the historical medical data samples, standardize the historical medical data matrix to obtain a standard medical data matrix, the standard medical data matrix includes several first features;

[0205] Calculation unit: used to update the initial weight of the first feature based on the residual of the first feature, and obtain the first weight;

[0206] And for obtaining T based on the first weight and the new observation sample 2 Statistic;

[0207] And for obtaining a feature vector matrix based on the standard medical data matrix, obtaining a reconstructed value based on the new observation sample and the feature vector matrix, and obtaining a Q statistic based on the new observation sample and the reconstructed value, wherein the Q statistic is used to quantify the severity of the new observation sample deviating from the normal state;

[0208] And to obtain the contribution value of the Q statistic based on the residual vector of the Q statistic and the first weight;

[0209] Threshold unit: used to obtain historical T based on the historical medical data sample. 2 The observed value sequence and the historical Q observed value sequence, based on the historical T 2 The observed value sequence and the historical Q observed value sequence are used to obtain T respectively. 2 Dynamic control threshold and Q-dynamic control threshold;

[0210] Analysis unit: used if the T 2 The statistic is greater than the stated T 2 If the dynamic control threshold is reached, or if the Q statistic is greater than the Q dynamic control threshold, the new observation sample is determined to be abnormal, and abnormal data is obtained based on the contribution value.

[0211] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0212] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A multivariate adaptive weighting method for monitoring medical data anomalies, characterized in that, The method includes: Acquire historical medical data samples from different patient groups and new observation samples from individual patients. Construct a historical medical data matrix based on the historical medical data samples. Standardize the historical medical data matrix to obtain a standard medical data matrix. The standard medical data matrix includes several first features. The initial weights of the first feature are updated based on the residuals of the first feature to obtain the first weights; obtaining T based on the first weight and the new observation sample 2 statistic A feature vector matrix is ​​obtained based on the standard medical data matrix. A reconstructed value is obtained based on the new observation sample and the feature vector matrix. A Q statistic is obtained based on the new observation sample and the reconstructed value. The Q statistic is used to quantify the severity of the new observation sample deviating from the normal state. Based on the residual vector of the Q statistic and the first weight, the contribution value of the Q statistic is obtained; Historical T was obtained based on the aforementioned historical medical data samples. 2 The observed value sequence and the historical Q observed value sequence, based on the historical T 2 The observed value sequence and the historical Q observed value sequence are used to obtain T respectively. 2 Dynamic control threshold and Q-dynamic control threshold; If the T 2 The statistic is greater than the stated T 2 If the dynamic control threshold is reached, or if the Q statistic is greater than the Q dynamic control threshold, the new observation sample is determined to be abnormal, and abnormal data is obtained based on the contribution value. The first calculation formula for obtaining the first weight is: ； in, Indicates the first The first feature is The first weight of time, Indicates the first The first feature is The first weight of time, Represents the smoothing coefficient. Indicates the first The first feature is Time residuals Indicates the first The first feature is Time residuals Indicates the number of the first feature. and Both represent integers greater than or equal to 1. Indicates the time.

2. The method for monitoring medical data anomalies using multivariate adaptive weights according to claim 1, characterized in that, Obtain the T 2 The second formula for calculating the statistic is: ； in, T represents 2 Statistic, Indicates a new observation sample. This represents the average value of a sample of historical medical data. This represents the adaptive weight diagonal matrix. Represents the covariance matrix. This indicates transpose.

3. The method for monitoring medical data anomalies using multivariate adaptive weights according to claim 2, characterized in that, The third calculation formula for obtaining the reconstructed value is: ； The fourth formula for obtaining the Q statistic is: ； in, Indicates the reconstructed value. Represents the eigenvector matrix, This represents the average value of a standard medical data matrix. This represents the standard deviation of a standard medical data matrix. This represents the Q statistic.

4. The method for monitoring medical data anomalies using multivariate adaptive weights according to claim 3, characterized in that, The fifth formula for obtaining the residual vector is: ； The sixth formula for obtaining the contribution value is: ； in, Indicates the first Each residual vector Indicates the first The contribution value of the first feature, Indicates the first The first weight of the first feature.

5. The method for monitoring medical data anomalies using multivariate adaptive weights according to claim 1, characterized in that, Obtain the T 2 The seventh formula for calculating the dynamic control threshold is: ； The eighth calculation formula for obtaining the Q dynamic control threshold is: ； in, Indicates the first T in a patient group 2 Dynamically control thresholds, Represents the quantile function. Indicates the confidence level. Indicates the first The first patient group The history of each feature T 2 Observation sequence, Indicates the first Q dynamic control threshold for each patient group Indicates the first The first patient group The historical Q-observation sequence of each feature and Both represent integers greater than or equal to 1.

6. The method for monitoring medical data anomalies using multivariate adaptive weights according to claim 1, characterized in that, The method further includes: Obtain the new observation sample that is abnormal, and obtain the abnormal sample; Pre-set patient metadata, cluster the historical medical data sample based on the patient metadata to obtain several state clusters, obtain several state models based on the state clusters, obtain state data of the state models, the state data includes benchmark mean vector, covariance matrix and principal component loading matrix, and obtain a model library based on all the state models. The abnormal samples are matched with the model library to obtain a reference model; Based on the abnormal samples and the reference model, the comprehensive deviation and residual abnormality are obtained, and the exceedance index is obtained based on the comprehensive deviation and the residual abnormality. Obtain the contribution of each second feature in the abnormal sample to the exceeding index; Abnormal information is obtained based on the aforementioned contribution level and the clinical event database; The specific steps for obtaining the clinical event database include: The process involves acquiring structured medical data and patient medical record data, whereby the structured medical data includes standard medical ontology data, clinical guideline data, and medical textbook data; obtaining structured data based on the structured medical data, and obtaining triples based on the patient medical record data, wherein the structured data includes clinical entities, structured pathways, and key relationships between the clinical entities, and the triples include medical parameters, direction of change, and clinical events; A medical knowledge graph is constructed based on the structured data and the triples. The medical knowledge graph includes several nodes and edges. The nodes include parameter nodes, change direction nodes, and clinical event nodes. Based on the medical knowledge graph, the path from the parameter node to the clinical event node is traversed, and seed rules are obtained based on the path. Time-series data is obtained based on the historical medical data sample, and data-driven rules are obtained based on the time-series data. The time-series data includes parameter contribution increase sequence patterns, high contribution parameters, and association rules between the high contribution parameters. Based on the seed rules, the data-driven rules, and the preset knowledge base, event rules are obtained, and the clinical event database is obtained based on the event rules.

7. The method for monitoring medical data anomalies using multivariate adaptive weights according to claim 6, characterized in that, The specific steps to obtain the reference model include: Obtain the Mahalanobis distance between the abnormal samples and each state model, and obtain the reference model based on the minimum Mahalanobis distance; The ninth formula for obtaining the overall deviation is: ； in, Indicates the overall deviation. Indicates an abnormal sample. Represents the baseline mean vector. Indicates transpose. This represents the contribution weight of the reference model. Represent the covariance matrix; The tenth formula for obtaining the residual abnormality is: ； in, Indicates the degree of residual abnormality. Represents the identity matrix. Represents the principal component loading matrix; The eleventh formula for calculating the contribution is as follows: ； in, Indicates contribution level. Indicates partial derivative, Indicates the index exceeding the standard. Indicates the first A second characteristic, Represents an integer greater than or equal to 1.

8. The method for monitoring medical data anomalies using multivariate adaptive weights according to claim 7, characterized in that, The specific steps for obtaining the contribution weight include: Based on the state model, several principal components are obtained, the variance explained rate of the principal components is obtained, the absolute value of the loading of each principal component is obtained based on the principal component loading matrix, and the basic weight of the second feature is obtained based on the absolute value of the loading and the variance explained rate. Obtain the state value of the principal component, obtain the adjustment factor based on the state value, and obtain the contribution weight based on the adjustment factor and the basic weight; The twelfth formula for obtaining the basic weights is: ； in, Indicates the first The basic weights of the second feature, Indicates the number of principal components. Represents the principal component loading matrix of the th The second feature in the first The absolute values ​​of the loadings of each principal component. Indicates the first The variance explained by each principal component; The thirteenth formula for obtaining the state value is: ； in, Indicates the first The second feature is The state value at time t, Indicates the attenuation factor. Indicates the first The second feature is The state value at time t, Indicates the first The second feature is The time-based exception indicator function; The fourteenth formula for obtaining the adjustment factor is: ； in, Indicates the first The second feature is Adjustment factor for time, Indicates the magnification factor; The fifteenth formula for obtaining the contribution weight is: ； in, Indicates the first The contribution weight of each second feature.

9. A multivariate adaptive weighted medical data anomaly monitoring system, characterized in that, The system includes: Data Unit: Used to acquire historical medical data samples of different patient groups and new observation samples of individual patients, construct a historical medical data matrix based on the historical medical data samples, standardize the historical medical data matrix to obtain a standard medical data matrix, the standard medical data matrix includes several first features; Calculation unit: used to update the initial weight of the first feature based on the residual of the first feature, and obtain the first weight; And for obtaining T based on the first weight and the new observation sample 2 Statistic; And for obtaining a feature vector matrix based on the standard medical data matrix, obtaining a reconstructed value based on the new observation sample and the feature vector matrix, and obtaining a Q statistic based on the new observation sample and the reconstructed value, wherein the Q statistic is used to quantify the severity of the new observation sample deviating from the normal state; And to obtain the contribution value of the Q statistic based on the residual vector of the Q statistic and the first weight; Threshold unit: used to obtain historical T based on the historical medical data sample. 2 The observed value sequence and the historical Q observed value sequence, based on the historical T 2 The observed value sequence and the historical Q observed value sequence are used to obtain T respectively. 2 Dynamic control threshold and Q-dynamic control threshold; Analysis unit: used if the T 2 The statistic is greater than the stated T 2 If the dynamic control threshold is reached, or if the Q statistic is greater than the Q dynamic control threshold, the new observation sample is determined to be abnormal, and abnormal data is obtained based on the contribution value. The first calculation formula for obtaining the first weight is: ； in, Indicates the first The first feature is The first weight of time, Indicates the first The first feature is The first weight of time, Represents the smoothing coefficient. Indicates the first The first feature is Time residuals Indicates the first The first feature is Time residuals Indicates the number of the first feature. and Both represent integers greater than or equal to 1. Indicates the time.

Images