Anomaly detection system for high-dimensional medical time series data

Abstract

The application discloses an abnormality detection system for high-dimensional medical time series data, which first proposes a hierarchical time series causal SHAP value method. Through the overall architecture of dimension grouping pruning and time hierarchy decomposition, the method can greatly reduce the feature contribution degree calculation complexity of high-dimensional medical time series data. Through a time series causal constraint mechanism, the method realizes a causal explanation conforming to physiological logic, provides more accurate decision support information for clinics, helps doctors more accurately understand the causes of abnormalities, and further reduces the feature contribution degree calculation complexity of high-dimensional medical time series data through frequency domain sparse sampling, avoids high overhead of time domain point-by-point calculation, and further reduces the feature contribution degree calculation complexity of high-dimensional medical time series data. In addition, a cross-cycle deviation degree matrix can be outputted, so that the contribution degree of each time step relative to the deviation degree of the historical same period normal level is analyzed, and more rich decision support information is provided for clinical decision making.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, and in particular, to an anomaly detection system for high-dimensional medical time-series data. Background Technology

[0002] Traditional anomaly detection systems typically use fixed thresholds for anomaly detection, relying solely on whether a single data point deviates from the normal range to determine the presence of anomalies. This approach can easily misjudge normal periodic fluctuations in data as anomalies and is also unable to effectively handle complex anomalies caused by long-term trends or short-term sudden events, resulting in poor detection accuracy. For medical time-series data, anomalies are not merely single data points deviating from the normal range; they often involve complex anomaly patterns, such as long-term trend anomalies and periodic fluctuation anomalies. For example, blood glucose data in diabetic patients exhibits not only short-term fluctuations but also long-term trend changes. Periodic changes in blood glucose levels are influenced by diet, exercise, and medication, and are dynamic. Traditional anomaly detection methods often focus on deviations at specific data points, ignoring these long-term trend changes, leading to decreased accuracy in predicting disease progression. Similarly, for diseases with periodic fluctuations, such as sleep apnea syndrome, respiratory rate exhibits a clear diurnal cycle. Healthy individuals typically experience more stable breathing during deep sleep at night, while during the day, breathing varies depending on activity levels. Traditional anomaly detection systems may overlook these periodic fluctuations, incorrectly identifying nighttime breathing changes as abnormal. However, this periodic fluctuation is precisely one of the early signs of sleep apnea, and accurately capturing this information is crucial for disease prevention. Therefore, anomaly detection systems in medical settings need to be able to identify more complex data anomalies to more accurately reflect patients' health conditions and provide doctors with more accurate references or suggestions for disease diagnosis.

[0003] With the continuous development of machine learning technology, Hybrid Expert Models (MoE) have been proposed. They improve model performance by combining the advantages of multiple expert models, especially when dealing with large-scale data. MoE can flexibly allocate computational resources to the most relevant subtasks, improving efficiency and accuracy, and is well-suited for anomaly detection in multidimensional medical time-series data. However, because machine learning models are "black box models" lack sufficient interpretability, and medical scenarios have high requirements for the interpretability of model output, existing technologies have introduced the SHAP value method to calculate the feature contribution of each dimension, thereby performing interpretability analysis on multimodal data in medical time-series anomaly detection systems. However, existing SHAP value methods require full permutation calculation of all features, with a complexity of O(2^3). (M×T)In this context, M represents the dimension and T represents the time step. However, for high-dimensional medical monitoring data containing multiple physiological indicators and with time steps of thousands of points, the computational complexity explodes exponentially, resulting in excessive computational load that cannot meet the real-time interpretation requirements of clinical anomaly detection. Furthermore, existing SHAP value methods assume that features are independent of each other and randomly remove features to calculate marginal contributions, ignoring temporal causal relationships. However, some related physiological indicators have strict temporal causality; for example, elevated blood sugar usually affects heart rate 15-30 minutes later. Random removal disrupts physiological logic, leading to interpretation results containing a large number of "reverse temporal" pseudo-interactions that can mislead clinical decisions. In addition, existing SHAP value methods calculate SHAP values ​​point-by-point in the time domain, failing to utilize the sparse distribution characteristics of physiological signals in the frequency domain. For example, heart rate energy is concentrated in the 1-2Hz fundamental frequency, resulting in wasted computational resources in low signal-to-noise ratio frequency bands, which cannot meet the latency requirements of real-time monitoring. Summary of the Invention

[0004] This invention provides an anomaly detection system for high-dimensional medical time-series data, which can greatly reduce the computational complexity of feature contribution of high-dimensional medical time-series data, and can also realize causal explanations that conform to physiological logic, providing more accurate and richer decision support information for clinical decision-making.

[0005] According to one aspect of the present invention, an anomaly detection system for high-dimensional medical time-series data is provided, comprising: The data preprocessing module is used to input high-dimensional medical time series data sequences, divide the high-dimensional medical time series data sequences into time windows, and extract high-dimensional feature vectors based on the divided time windows; The anomaly detection module is used to input the extracted high-dimensional feature vectors into a pre-trained hybrid expert model to obtain anomaly detection results; The causal explanation and analysis module is used to calculate the contribution of each dimension's features to the anomaly detection results using the hierarchical temporal causal SHAP value method, and outputs a causal SHAP value matrix and a cross-period deviation matrix. The hierarchical temporal causal SHAP value method reduces the computational complexity of feature contributions through four mechanisms: dimension grouping and pruning, temporal hierarchical decomposition, temporal causal constraints, and frequency domain sparse sampling. The causal SHAP value matrix is ​​used to characterize the contribution of each dimension's features to the anomaly detection results at each time step, and the cross-period deviation matrix is ​​used to characterize the degree of deviation of the contribution at each time step from the normal level of the same historical period. The detection report generation module is used to generate an anomaly detection report based on the anomaly detection results, the causal SHAP value matrix, and the cross-period deviation matrix.

[0006] Furthermore, the causal explanation analysis module includes: The dimensional grouping and pruning unit is used to divide the multidimensional physiological indicators in high-dimensional medical time series data into multiple functional groups. Each functional group corresponds to multiple physiological indicators that are clinically related in a physiological system, and the SHAP value of each functional group is calculated independently. The time-level decomposition unit is used to divide the time dimension into an intra-period layer and an inter-period layer. In the intra-period layer, the SHAP value of each dimension feature at each time step is calculated. In the inter-period layer, the cross-period baseline SHAP value is approximately obtained by using the aggregation function based on the SHAP value calculation results of multiple intra-period layers. The temporal causal constraint unit is used to establish a feature alliance set under intra-group temporal causal constraints among multiple physiological indicators in each functional group, and to calculate the SHAP value of the intra-cycle layer based on the feature alliance set. The frequency domain sparse sampling unit is used to select only the main frequency set for SHAP value calculation when calculating the SHAP value of the layer within the calculation period; The output integration unit is used to generate a causal SHAP value matrix based on the SHAP value calculation results of the inner layer of the cycle. It also calculates the cross-cycle deviation of each time step based on the SHAP values ​​of each dimension feature at each time step and the cross-cycle benchmark SHAP value, and generates a cross-cycle deviation matrix.

[0007] Furthermore, the cross-cycle benchmark SHAP value is approximated based on the following formula: ; in, This represents the cross-cycle reference SHAP value at the τ-th phase position. Indicates the first p The time step with phase position τ in the inner layer of each period. P Indicates the number of layers within a period. Indicates the first d 3D features at time step SHAP value, This represents an aggregate function.

[0008] Furthermore, the feature union set under the intra-group temporal causal constraint is represented as: ; in, Indicates the calculation of the first k Within the first functional group i When considering the SHAP value of a feature, the constructed feature union set under intra-group temporal causal constraints is... Indicates the first k Functional groups S Represents a feature subset. Indicates the first j The time step of the dimensional feature Indicates the first i The time step of the dimensional feature Indicates the first k The first functional group j Dimensional features for the 1st i The minimum lag time for the effect of a dimensional feature.

[0009] Furthermore, the cross-cycle deviation for each time step is calculated based on the following formula: ; in, Indicates the first d 3D features at time step t Cross-cycle deviation, Indicates the first d 3D features at time step t SHAP value, This represents the cross-cycle reference SHAP value corresponding to the phase position. Represents the phase mapping function, , Indicates the length of the time window.

[0010] Furthermore, the set of dominant frequencies is determined based on the following formula: ; in, Indicates the first i The set of principal frequencies of dimensional features Indicates the first i The 3D features are obtained by Fourier transform. k Each frequency component amplitude, This indicates the amplitude threshold.

[0011] Furthermore, the causal explanation analysis module also includes a cross-group causal inference engine unit, which is used to determine whether there is a causal transmission path between different functional groups, and when a causal transmission path is determined to exist, to enhance the correlation of the weight coefficients when the hybrid expert model is weighted and fused.

[0012] Furthermore, the weighting coefficients are further enhanced based on the following formula: ; in, Indicates functional group v Enhanced weighting coefficients Indicates functional group v Weighting coefficients before enhancement Represents the causal coupling coefficient. Indicates the indicator function, when the function group u and functional groups v The value is 1 if a causal transmission path exists, otherwise the value is 0.

[0013] Furthermore, if the following conditions are met: Then it is determined that there is a causal transmission path between different functional groups, where, , Indicates functional group u The set of abnormal time steps Indicates functional group u Abnormal time step, Indicates functional group u , Represents the causal SHAP value matrix. d Representing feature dimension, t Indicates a time step. Indicates functional group u The threshold for abnormal SHAP values, Indicates functional group v Abnormal time step, Indicates functional group u For functional groups v The minimum lag time that produces an effect. This indicates the allowable physiological fluctuation tolerance.

[0014] Furthermore, the data preprocessing module first performs a Fourier transform on the medical time series data of each dimension to extract the principal frequency component of each dimension, then determines the window size of each dimension based on the corresponding principal frequency component, filters the largest among all windows, takes an integer multiple of the largest window as the unified time window, and zero-padding is performed on the dimension data whose time series length is less than the time window.

[0015] The present invention has the following beneficial effects: This invention presents an anomaly detection system for high-dimensional medical time-series data, which for the first time proposes a hierarchical temporal causal SHAP value method. This method, through an overall architecture of dimensional grouping pruning and time-level decomposition, significantly reduces the computational complexity of feature contribution calculations for high-dimensional medical time-series data. Furthermore, it achieves a causal explanation consistent with physiological logic through a temporal causal constraint mechanism, providing more accurate decision support information for clinicians and helping doctors understand the causes of anomalies more accurately. Moreover, it accelerates computation through frequency domain sparse sampling, avoiding the high overhead of point-by-point time-domain calculations, further reducing the computational complexity of feature contribution calculations for high-dimensional medical time-series data. In addition, it can output a cross-period deviation matrix to analyze the degree of deviation of the contribution at each time step from the historical normal level, providing richer decision support information for clinical decision-making.

[0016] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the module structure of an anomaly detection system for high-dimensional medical time-series data according to a preferred embodiment of this application. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] Reference Figure 1 A preferred embodiment of this application provides an anomaly detection system for high-dimensional medical time-series data, comprising: The data preprocessing module is used to input high-dimensional medical time series data sequences, divide the high-dimensional medical time series data sequences into time windows, and extract high-dimensional feature vectors based on the divided time windows; The anomaly detection module is used to input the extracted high-dimensional feature vectors into a pre-trained hybrid expert model to obtain anomaly detection results; The causal explanation and analysis module is used to calculate the contribution of each dimension's features to the anomaly detection results using the hierarchical temporal causal SHAP value method, and outputs a causal SHAP value matrix and a cross-period deviation matrix. The hierarchical temporal causal SHAP value method reduces the computational complexity of feature contributions through four mechanisms: dimension grouping and pruning, temporal hierarchical decomposition, temporal causal constraints, and frequency domain sparse sampling. The causal SHAP value matrix is ​​used to characterize the contribution of each dimension's features to the anomaly detection results at each time step, and the cross-period deviation matrix is ​​used to characterize the degree of deviation of the contribution at each time step from the normal level of the same historical period. The detection report generation module is used to generate an anomaly detection report based on the anomaly detection results, the causal SHAP value matrix, and the cross-period deviation matrix.

[0020] It is understood that the anomaly detection system for high-dimensional medical time-series data in this embodiment proposes, for the first time, a hierarchical temporal causal SHAP value method. This method, through its overall architecture of dimensional grouping pruning and time-level decomposition, can significantly reduce the computational complexity of feature contribution calculations for high-dimensional medical time-series data. Furthermore, it achieves a causal explanation consistent with physiological logic through a temporal causal constraint mechanism, providing more accurate decision support information for clinicians and helping doctors understand the causes of anomalies more accurately. Moreover, it accelerates computation through frequency domain sparse sampling, avoiding the high overhead of point-by-point time-domain calculations, further reducing the computational complexity of feature contribution calculations for high-dimensional medical time-series data. In addition, it can output a cross-period deviation matrix to analyze the degree of deviation of the contribution at each time step from the historical normal level, providing richer decision support information for clinical decision-making.

[0021] The input high-dimensional medical time-series data sequence specifically includes electrocardiogram data, respiratory rate data, body temperature data, blood pressure data, blood glucose data, and other physiological indicator data. The data preprocessing module first cleans, standardizes, or performs other preprocessing on the high-dimensional medical time-series data sequence to ensure data quality. Then, it divides the high-dimensional medical time-series data sequence into a unified time window. Specifically, the data preprocessing module first performs a Fourier transform on the medical time-series data of each dimension to extract the principal frequency component of each dimension. The principal frequency is used as the periodic feature of that dimension. Then, based on the corresponding principal frequency component, the window size of each dimension is determined, specifically based on the following formula: ,in, Indicates the first d The window size corresponding to the physiological index data. Indicates the first d The dominant frequency of the physiological indicator data is determined. Then, the largest value in all windows is selected, and integer multiples of the largest window are used as the unified time window. Preferably, the largest window is used as the time window. Furthermore, to ensure consistency across the data windows for each dimension, data padding is performed to ensure that the time series length for each dimension matches the window size, avoiding inconsistencies caused by data length mismatches. Therefore, zero-padding is applied to dimension data with time series lengths shorter than the time window. Finally, high-dimensional feature vectors are extracted based on the unified time window division. These high-dimensional feature vectors include the periodic features of the data in each dimension within the time window.

[0022] It is understandable that traditional methods for anomaly detection in multidimensional time-series data typically employ a uniform window partitioning mechanism. This involves using a fixed-size sliding window to segment the time-series data and analyzing each window. However, for physiological signals with long or short periods, this fixed-window method often fails to adapt to the different periodic characteristics of the data, leading to the loss of important information or the inability to effectively capture periodic changes. For example, long-term periodic changes, such as circadian rhythms and hormonal cycles, affect the body's basic health status and long-term physiological trends. For instance, circadian rhythms regulate multiple physiological activities such as sleep, body temperature, and heart rate. Any irregular circadian rhythm can adversely affect the heart, endocrine system, and other systems, thereby impacting health. Therefore, accurately capturing long-term periodic data is crucial for predicting the risk of chronic diseases (such as diabetes, depression, and cardiovascular disease). For such data, if the time window is set too short, the periodic change patterns cannot be fully captured. Short-term periodic data (such as heart rate and respiratory rate) are typically caused by rapid fluctuations in physiological activity or emotional state. This type of data is particularly important for monitoring short-term health conditions. For example, heart rate variability (HRV) is an important indicator of cardiac health, reflecting the autonomic nervous system's ability to regulate heart rhythm. Heart rate has very short periodic fluctuations, and traditional fixed-window partitioning methods may not be able to accurately capture the minute changes in each heartbeat, thus missing valuable health signals. Furthermore, traditional unified window partitioning mechanisms are often designed only for single-dimensional time series and cannot be directly applied to high-dimensional medical time series data where different dimensions have different periodic characteristics. They usually require iterative calculations, resulting in high algorithmic complexity, making them unsuitable for medical monitoring scenarios with high real-time requirements.

[0023] This invention directly uses the reciprocal of the dominant frequency obtained from the Fourier transform as the window size. It utilizes the Fourier transform to analyze the periodic characteristics of each dimension's data, thereby selecting the most suitable window size for each dimension. This approach ensures that the time-series data of each dimension is segmented according to its periodic characteristics, thus better preserving the periodic information of the data and improving the accuracy of anomaly detection. Furthermore, it eliminates the need for complex iterative calculations and signal reconstruction verification, with a computational complexity of only O(n log n), meeting the latency requirements of real-time medical monitoring. Data completion ensures that the data length of all dimensions reaches a uniform time window length, effectively solving the problem of large periodic differences in different physiological indicators in medical data. This guarantees the consistency of multi-dimensional data in subsequent processing. The window size is dynamically determined based on the actual periodic characteristics of the data, rather than a preset fixed value. The window boundary allows for fine-tuning according to data characteristics to better adapt to the periodic fluctuations of physiological signals. It can flexibly handle individual differences among different patients and physiological changes of the same patient at different times, significantly improving the accuracy of window segmentation and the system's generalization ability.

[0024] Furthermore, the anomaly detection module employs a pre-trained hybrid expert system (MoE model) for anomaly detection. After the extracted high-dimensional feature vector is input into the pre-trained hybrid expert model, the MoE model evaluates the matching degree between the input data and each expert model through a routing mechanism. Specifically, a gating network is used to calculate the matching score of each expert model, which can be expressed as: , Representation of expert models With the input high-dimensional feature vector Match scores between them The matching score calculation function is described in detail here, as its specific expression is existing technology. The softmax function then transforms the matching score of each expert model into a selection probability, representing the priority of each expert model in processing this high-dimensional feature vector. This can be expressed as: , Indicates the first i The selection probability of an expert model N To represent the number of expert models, the gating network will generate a length of... N probability vector ,satisfy Based on the probability vector, the routing mechanism determines which expert models to activate, and then performs a weighted fusion of the activated expert models to form a comprehensive anomaly detection result, which can be represented as: .

[0025] In the MoE model, each expert model employs an agent specific to a particular knowledge domain, such as cardiology or diabetes, to focus on detecting anomalous data within that domain, thus facilitating accurate disease prediction. Each agent is trained on an autoencoder within a domain-specific dataset, learning normal behavior patterns within that domain. Each domain's expert model is optimized based on specific features and anomalous patterns to ensure a sensitive response to anomalies. Expert models with high matching scores receive input data, which is then processed by the autoencoder to calculate the reconstruction error. A smaller reconstruction error indicates the input data closely approximates the normal pattern, while a larger error indicates deviation from the normal pattern, potentially indicating anomalous data. For multiple dimensions of a patient's health data (e.g., ECG, blood glucose, blood pressure), each domain expert performs an independent anomaly detection. The final detection results are then fused using ensemble learning to generate a comprehensive anomaly detection report. This invention employs a Bayesian method to weightedly fuse the outputs of multiple expert models, outputting the final anomaly detection result. Specifically, each expert model has a prior probability. That is, the first iThe probability that an expert model considers data to be anomalous without any prior data is typically defined based on factors such as the expert model's historical performance, domain knowledge, and accuracy. Subsequently, each expert model analyzes the input data and outputs a likelihood probability value. , indicating that in the given data D Under the conditions, the first i Each expert model determines the probability of data anomalies. Next, the posterior probability of each expert model needs to be calculated by combining the prior probability and the likelihood probability. , can be represented as: , The evidence probability represents the total probability of all expert models judging the data, and is usually a normalized constant. Finally, the posterior probabilities of each expert model are weighted and fused to obtain the final data anomaly probability. .

[0026] It is understandable that autoencoder-based expert models are specifically designed for anomaly detection in medical time-series data. By learning normal behavior patterns within the domain and judging anomalies based on reconstruction errors, they better align with the essential characteristics of medical anomaly detection. Furthermore, the Bayesian fusion mechanism dynamically weights prior probabilities and likelihood probabilities, comprehensively considering both historical expert performance and current data evidence to generate more reliable diagnostic results. Simultaneously, the cross-domain expert collaboration mechanism effectively addresses complex medical scenarios with multiple coexisting diseases, achieving accurate anomaly detection across multiple domains of medical data. This effectively solves the problem of single models being unable to handle complex and diverse anomaly types and the collaborative analysis of cross-domain data in existing technologies. Additionally, by calculating the matching scores of each expert model through a gating network, then using the Softmax function to convert the matching scores into selection probabilities, and finally dynamically determining which expert models to activate based on the selection probabilities, this dynamic routing mechanism ensures that each data input selects the most suitable expert for analysis, avoiding resource waste.

[0027] Furthermore, addressing the computational infeasibility of the standard SHAP value method in high-dimensional medical time-series data, this invention proposes for the first time a Hierarchical Temporal Causal SHAP (HTC-SHAP) method. This algorithm, through a triple mechanism of dimensional grouping pruning, temporal hierarchical decomposition, and frequency domain sparse sampling, reduces the computational complexity of feature contribution from exponential to polynomial levels, significantly lowering computational complexity. Moreover, it achieves a causal explanation consistent with physiological logic through a temporal causal constraint mechanism. Specifically, the causal explanation analysis module includes: The dimensional grouping and pruning unit is used to divide the multidimensional physiological indicators in high-dimensional medical time series data into multiple functional groups. Each functional group corresponds to multiple physiological indicators that are clinically related in a physiological system, and the SHAP value of each functional group is calculated independently. The time-level decomposition unit is used to divide the time dimension into an intra-period layer and an inter-period layer. In the intra-period layer, the SHAP value of each dimension feature at each time step is calculated. In the inter-period layer, the cross-period baseline SHAP value is approximately obtained by using the aggregation function based on the SHAP value calculation results of multiple intra-period layers. The temporal causal constraint unit is used to establish a feature alliance set under intra-group temporal causal constraints among multiple physiological indicators in each functional group, and to calculate the SHAP value of the intra-cycle layer based on the feature alliance set. The frequency domain sparse sampling unit is used to select only the main frequency set for SHAP value calculation when calculating the SHAP value of the layer within the calculation period; The output integration unit is used to generate a causal SHAP value matrix based on the SHAP value calculation results of the inner layer of the cycle. It also calculates the cross-cycle deviation of each time step based on the SHAP values ​​of each dimension feature at each time step and the cross-cycle benchmark SHAP value, and generates a cross-cycle deviation matrix.

[0028] The dimensional grouping and pruning unit divides the multidimensional physiological indicators in high-dimensional medical time-series data into multiple functional groups based on prior medical knowledge. Each functional group corresponds to multiple clinically relevant physiological indicators within a physiological system; for example, they can be divided into cardiovascular, metabolic, and respiratory groups, etc., denoted as... , K This indicates the number of functional groups. Each functional group has its own SHAP value calculated separately. Only the temporal interactions between dimensions within that functional group are analyzed; associations between different functional groups are not included in the single-group SHAP value calculation. This avoids computational redundancy caused by cross-group feature combinations without clinical relevance. By calculating SHAP values ​​in groups, invalid computations can be significantly reduced, lowering the algorithm complexity to [value missing]. ,in, This indicates that the maximum number of dimensions in each functional group is much smaller than the total number of dimensions.

[0029] Furthermore, the time-level decomposition unit divides the time dimension into an intra-period layer and an inter-period layer based on the optimal window obtained by the Fourier transform in the window partitioning mechanism. For example, for the first... d 3D time series data, according to the corresponding optimal window Can be divided into P Non-overlapping periodic blocks Each periodic block contains A series of consecutive time steps, each periodic block As an inner layer within a cycle, within each inner layer, precise SHAP value calculation is performed under temporal causality constraints, calculating the SHAP value of each dimension feature at each time step. ,in, , Indicates the first d 3D features at time step t The SHAP value, which directly reflects the first d Dimensional features in the time step of the inner layer of the corresponding period t The precise marginal contribution at each time step. Therefore, after calculating the SHAP values ​​of all dimensional features at each time step, a causal SHAP value matrix can be constructed. For the periodic layers, this invention is based on the periodic stability assumption of physiological signals: at the same phase position (such as the same time step within each periodic layer), there should be similar contribution patterns under normal physiological conditions. Therefore, this invention uses an aggregation function to aggregate the local contribution values ​​of all periodic blocks at the same phase position, which can approximate the cross-period baseline SHAP value without re-performing the highly complex full permutation calculation for each period. This can be expressed as: ; in, This represents the cross-cycle reference SHAP value at the τ-th phase position. Indicates the first p The time step with phase position τ in the inner layer of each period. P Indicates the number of layers within a period. Indicates the first d 3D features at time step SHAP value, This represents an aggregate function, and the specific aggregate function depends on the first... d The characteristics of physiological indicators are adaptively selected. For basic rhythm indicators (such as body temperature and resting heart rate), the arithmetic mean aggregation function is used, which can be expressed as: To smooth short-term fluctuations and establish typical normal benchmarks, for peak-sensitive indicators (such as systolic blood pressure and postprandial blood glucose), the maximum aggregation function is used, which can be expressed as: The extreme contribution levels of the same historical period are retained as the anomaly detection threshold.

[0030] It is understood that this invention performs hierarchical decomposition of the time dimension, accurately calculates the SHAP value within the periodic layer, and approximates the calculation through aggregation functions at the inter-period layer. This not only preserves the local dynamic characteristics within the period, but also realizes cross-period trend analysis, and reduces computational complexity.

[0031] Furthermore, this invention considers that the influence between different physiological indicators within the same functional group often exhibits a time lag rather than instantaneous occurrence. For example, within the metabolome, the peak insulin secretion after oral glucose administration typically requires approximately 15-30 minutes to trigger a significant change in blood glucose concentration; similarly, within the cardiovascular group, there is a clear lag in the transmission relationship between blood pressure fluctuations caused by sympathetic nerve excitation and subsequent compensatory changes in heart rate. Therefore, this invention establishes a feature alliance set under intra-group temporal causal constraints among multiple physiological indicators within each functional group through a temporal causal constraint unit. Under the constraints of this feature alliance set, precise SHAP values ​​are calculated for the intra-cycle layer, achieving a causal explanation consistent with physiological logic. Specifically, this invention pre-establishes an intra-group temporal dependency matrix. , where the elements in the matrix Indicates the first k The first functional group j Dimensional features (corresponding to the 1st dimension) j (4 physiological indicators) for the first i The minimum lag time for the effect of a dimensional feature, if A value of 0 indicates that there is no lag effect or the effect is negligible between the two indicators. In calculating the... i When selecting the SHAP value of a feature, only a subset of features belonging to the same functional group and satisfying the temporal sequence constraint are allowed to be selected. For subsets S Any feature in j Must meet Therefore, this invention defines a feature alliance set under intra-group temporal causal constraints, which can be represented as: ; in, Indicates the calculation of the first k Within the first functional group i When considering the SHAP value of a feature, the constructed feature union set under intra-group temporal causal constraints is... Indicates the first k Functional groups S Represents a feature subset. Indicates the first j The time step of the dimensional feature Indicates the first i The time step of the dimensional feature Indicates the first k The first functional group j Dimensional features for the 1st i The minimum lag time for the influence of a feature. The formula for calculating the SHAP value of the inner layer of a cycle based on this feature consortium set can be expressed as: , Representing a feature subset S The predicted value, Representing a feature subset S Add features i Subsequent predicted values, Indicates the first k The number of physiological indicators (i.e., the number of dimensions) within each functional group. This represents the dimension of the feature subset.

[0032] Furthermore, this invention considers that physiological indicator signals typically exhibit sparse distribution characteristics in the frequency domain. For example, taking heart rate signals as an example, their energy is mainly concentrated in the fundamental frequency of 1-2Hz and a small number of harmonics, rather than being uniformly distributed across all frequencies. The standard SHAP value method calculates the contribution of each time step point-by-point in the time domain, which is equivalent to dense sampling across the entire frequency band, wasting a significant amount of computation on frequency components with extremely low signal-to-noise ratios. Therefore, this invention uses a frequency domain sparse sampling unit to select only the main frequency set for SHAP value calculation. For example, for the first... i The 24 physiological indicators, after Fourier transform, are the 1st and 2nd physiological indicators. k Each frequency component is The frequency domain marginal contribution of this component can be expressed as: , This represents the model's prediction loss function. This indicates the loss of sensitivity to that frequency component. Due to the amplitude of the high-frequency noise component in the physiological signal... Approaching zero, its corresponding contribution value It also naturally decays, thus achieving energy adaptive sparsity. The frequency domain sparse sampling unit determines the main frequency set based on the following formula: ; in, Indicates the first i The set of dominant frequencies with significant eigenenergy. Indicates the first i The 3D features are obtained by Fourier transform. k Each frequency component amplitude, Indicates the amplitude threshold. , T This represents the length of the high-dimensional medical time-series data sequence. After calculating the SHAP value of the principal frequency set, the time-domain SHAP value is reconstructed through discrete Fourier transform, and can be expressed as: .

[0033] It is understandable that this invention, by establishing a joint frequency-time domain computation framework, selects only the main frequency set for SHAP value calculation, and then maps it back to the time domain in one go through a fast inverse Fourier transform, thus reducing the computational complexity from that in the time domain to that in the time domain. Reduced to the frequency domain This avoids the high overhead of point-by-point computation in the time domain.

[0034] In addition, the output integration unit can calculate the result based on the SHAP value of the inner layer of the cycle. Generate a causal SHAP value matrix Furthermore, based on the SHAP values ​​of each dimension's features at each time step and the cross-cycle baseline SHAP value, the cross-cycle deviation at each time step is calculated, and a cross-cycle deviation matrix is ​​generated. Specifically, the cross-cycle deviation at each time step is calculated based on the following formula: ; in, Indicates the first d 3D features at time step t Cross-cycle deviation, Indicates the first d 3D features at time step t SHAP value, This represents the cross-cycle reference SHAP value corresponding to the phase position. Represents the phase mapping function, , Indicates the length of the time window. If... A value significantly greater than zero indicates that the contribution at that time step exceeds the normal level for the same period in history, suggesting a cross-cycle anomaly. Furthermore, after calculating the cross-cycle deviation for each time step, a cross-cycle deviation matrix can be generated. .

[0035] It is understandable that for physiological causal influences between different functional groups, such as the association between abnormal blood glucose in the metabologroup and changes in heart rate in the cardiovascular group, this invention does not directly solve for the SHAP value during calculation. Instead, it uses the independent detection results of each expert model in the hybrid expert system and combines the posterior probabilities of each expert model in the Bayesian fusion stage for comprehensive inference. This avoids computational explosion caused by cross-group high-dimensional interactions while retaining the ability to analyze clinical associations across systems. Optionally, the causal explanation analysis module also includes a cross-group causal inference engine unit, used to determine whether there is a causal transmission path between different functional groups, and to enhance the association of the weight coefficients when the weighted fusion of the hybrid expert models is performed when a causal transmission path is determined to exist. Specifically, to quantify the physiological causal influences between different functional groups, the cross-group causal inference engine unit introduces a cross-group temporal dependency matrix in the Bayesian fusion layer of the hybrid expert system. Elements in the matrix Indicates functional group For functional groups The minimum lag time for an effect to occur, such as the time from abnormal blood glucose in the metabolomics group to heart rate variability in the cardiovascular group, corresponding to The time-dependent matrix is ​​30 minutes long and is obtained based on prior medical knowledge or statistical learning of historical data, and is related to the intra-group temporal dependency matrix. Together, they form a two-layer time-dependency architecture. This is true if the following conditions are met: Then, the cross-group causal inference engine unit determines that there is a causal transmission path between different functional groups and records the cross-group causal transmission event as follows: ,in, , Indicates functional group u The set of abnormal time steps Indicates functional group u Abnormal time step, Indicates functional group u , Represents the causal SHAP value matrix. d Representing feature dimension, t Indicates a time step. Indicates functional group u The threshold for abnormal SHAP values, Indicates functional group v Abnormal time step, Indicates functional group u For functional groups v The minimum lag time that produces an effect. This represents the allowable tolerance for physiological fluctuations, typically taken as 1-2 sampling intervals. During the Bayesian fusion phase, cross-group causal transmission events are used to dynamically adjust the weights of the expert model; for example, if the expert model... and If a verified causal transmission path exists, then for the result group... The correlation is enhanced based on the following formula for its weight coefficients: ; in, Indicates functional group v Enhanced weighting coefficients Indicates functional group v Weighting coefficients before enhancement This represents the causal coupling coefficient. If a strong causal relationship is determined based on clinical literature, the value is 0.3; if a weak causal relationship is determined based on statistical correlation, the value is 0.1. Indicates the indicator function, when the function group u and functional groups v The value is 1 when a causal transmission path exists; otherwise, it is 0. For example, when a causal transmission path consistent with physiological lag exists between abnormal blood glucose in the metabolomics group and heart rate variability in the cardiovascular group, the weight coefficient of the cardiovascular expert model will be enhanced by causal evidence, thereby improving the accuracy of cross-functional group anomaly detection in scenarios with multiple coexisting diseases. In addition, while outputting the comprehensive anomaly detection results, the system can also generate a cross-group causal transmission path report, which can be represented as: This report can provide clinicians with cross-functional group interpretations with clear timestamps and physiological logic, such as "the abnormal blood glucose occurred at 09:15, and after 30 minutes of metabolic transmission, the heart rate variability reached its peak at 09:45." This is decision support information that a single intragroup SHAP value analysis cannot provide.

[0036] Furthermore, after the anomaly detection module outputs the anomaly detection results and the causal interpretation analysis module outputs the causal SHAP value matrix and the cross-period deviation matrix, the detection report generation module can generate an anomaly detection report based on the anomaly detection results, the causal SHAP value matrix, and the cross-period deviation matrix. This provides comprehensive and accurate decision support information for clinical diagnosis, helping doctors to more accurately understand the causes of anomalies. The specific report generation process is existing technology and will not be elaborated upon here.

[0037] Although preferred embodiments of this application 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 the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0038] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

[0039] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An anomaly detection system for high-dimensional medical time-series data, characterized in that, include: The data preprocessing module is used to input high-dimensional medical time series data sequences, divide the high-dimensional medical time series data sequences into time windows, and extract high-dimensional feature vectors based on the divided time windows; The anomaly detection module is used to input the extracted high-dimensional feature vectors into a pre-trained hybrid expert model to obtain anomaly detection results; The causal explanation and analysis module is used to calculate the contribution of each dimension's features to the anomaly detection results using the hierarchical temporal causal SHAP value method, and outputs a causal SHAP value matrix and a cross-period deviation matrix. The hierarchical temporal causal SHAP value method reduces the computational complexity of feature contributions through four mechanisms: dimension grouping and pruning, temporal hierarchical decomposition, temporal causal constraints, and frequency domain sparse sampling. The causal SHAP value matrix is ​​used to characterize the contribution of each dimension's features to the anomaly detection results at each time step, and the cross-period deviation matrix is ​​used to characterize the degree of deviation of the contribution at each time step from the normal level of the same historical period. The detection report generation module is used to generate an anomaly detection report based on the anomaly detection results, the causal SHAP value matrix, and the cross-period deviation matrix.

2. The anomaly detection system for high-dimensional medical time-series data as described in claim 1, characterized in that, The causal explanation and analysis module includes: The dimensional grouping and pruning unit is used to divide the multidimensional physiological indicators in high-dimensional medical time series data into multiple functional groups. Each functional group corresponds to multiple physiological indicators that are clinically related in a physiological system, and the SHAP value of each functional group is calculated independently. The time-level decomposition unit is used to divide the time dimension into an intra-period layer and an inter-period layer. In the intra-period layer, the SHAP value of each dimension feature at each time step is calculated. In the inter-period layer, the cross-period baseline SHAP value is approximately obtained by using the aggregation function based on the SHAP value calculation results of multiple intra-period layers. The temporal causal constraint unit is used to establish a feature alliance set under intra-group temporal causal constraints among multiple physiological indicators in each functional group, and to calculate the SHAP value of the intra-cycle layer based on the feature alliance set. The frequency domain sparse sampling unit is used to select only the main frequency set for SHAP value calculation when calculating the SHAP value of the layer within the calculation period; The output integration unit is used to generate a causal SHAP value matrix based on the SHAP value calculation results of the inner layer of the cycle. It also calculates the cross-cycle deviation of each time step based on the SHAP values ​​of each dimension feature at each time step and the cross-cycle benchmark SHAP value, and generates a cross-cycle deviation matrix.

3. The anomaly detection system for high-dimensional medical time-series data as described in claim 2, characterized in that, The cross-cycle baseline SHAP value is approximated based on the following formula: ； in, This represents the cross-cycle reference SHAP value at the τ-th phase position. Indicates the first p The time step with phase position τ in the inner layer of each period. P Indicates the number of layers within a period. Indicates the first d 3D features at time step SHAP value, This represents an aggregate function.

4. The anomaly detection system for high-dimensional medical time-series data as described in claim 2, characterized in that, The feature union set under the intra-group temporal causal constraint is represented as follows: ； in, Indicates the calculation of the first k Within the first functional group i When considering the SHAP value of a feature, the constructed feature union set under intra-group temporal causal constraints is... Indicates the first k Functional groups S Represents a feature subset. Indicates the first j The time step of the dimensional feature Indicates the first i The time step of the dimensional feature Indicates the first k The first functional group j dimensional features for the 1st i The minimum lag time for the effect of a dimensional feature.

5. The anomaly detection system for high-dimensional medical time-series data as described in claim 2, characterized in that, The cross-cycle deviation at each time step is calculated using the following formula: ； in, Indicates the first d 3D features at time step t Cross-cycle deviation, Indicates the first d 3D features at time step t SHAP value, This represents the cross-cycle reference SHAP value corresponding to the phase position. Represents the phase mapping function, , Indicates the length of the time window.

6. The anomaly detection system for high-dimensional medical time-series data as described in claim 2, characterized in that, The set of dominant frequencies is determined based on the following formula: ； in, Indicates the first i The set of principal frequencies of dimensional features Indicates the first i The 3D features are obtained by Fourier transform. k Each frequency component amplitude, This indicates the amplitude threshold.

7. The anomaly detection system for high-dimensional medical time-series data as described in claim 2, characterized in that, The causal explanation analysis module also includes a cross-group causal inference engine unit, which is used to determine whether there is a causal transmission path between different functional groups, and when a causal transmission path is determined to exist, to enhance the correlation of the weight coefficients when the hybrid expert model is weighted and fused.

8. The anomaly detection system for high-dimensional medical time-series data as described in claim 7, characterized in that, The weight coefficients are correlated and enhanced based on the following formula: ； in, Indicates functional group v Enhanced weighting coefficients Indicates functional group v Weighting coefficients before enhancement Represents the causal coupling coefficient. Indicates the indicator function, when the function group u and functional groups v The value is 1 if a causal transmission path exists, otherwise the value is 0.

9. The anomaly detection system for high-dimensional medical time-series data as described in claim 7, characterized in that, If the following conditions are met: Then it is determined that there is a causal transmission path between different functional groups, where, , Indicates functional group u The set of abnormal time steps Indicates functional group u Abnormal time step, Indicates functional group u , Represents the causal SHAP value matrix. d Representing feature dimension, t Indicates a time step. Indicates functional group u The threshold for abnormal SHAP values, Indicates functional group v Abnormal time step, Indicates functional group u For functional groups v The minimum lag time that produces an effect. This indicates the allowable physiological fluctuation tolerance.

10. The anomaly detection system for high-dimensional medical time-series data as described in claim 1, characterized in that, The data preprocessing module first performs a Fourier transform on the medical time series data for each dimension to extract the principal frequency component of each dimension. Then, based on the corresponding principal frequency component, it determines the window size for each dimension, filters out the largest among all windows, uses an integer multiple of the largest window as the unified time window, and zero-padding is performed on the dimension data whose time series length is less than the time window.

Images