Medical data classification management system based on big data

Abstract

The invention relates to the technical field of medical data management, and discloses a medical data classification management system based on big data. A data integration and classification module of the system receives a historical medical data set and a real-time medical data stream, performs feature fusion and classification prediction through a deep learning model, and generates a theoretical classification result. And the deviation quantitative analysis module performs multi-dimensional deviation calculation on the theoretical classification result and the clinical actual classification result to generate a patient-level multi-dimensional deviation tensor. And the graph network anomaly positioning module inputs the multi-dimensional deviation tensor into a graph neural network for anomaly propagation simulation based on a medical institution topological structure, and generates an anomaly probability distribution diagram to identify a physical region where an anomaly occurs. And the adaptive decision control module dynamically allocates monitoring resources according to the abnormal probability distribution diagram, including starting high-frequency data acquisition for a high-probability abnormal region and executing an intervention test on adjacent nodes.

Description

Technical Field

[0001] This invention relates to the field of medical data management technology, specifically a medical data classification and management system based on big data. Background Technology

[0002] Current medical data classification management systems primarily employ rule-based engines or traditional machine learning algorithms, which have significant limitations when handling large-scale, heterogeneous medical data. Existing technologies typically separate the data classification process from actual clinical applications, lacking an effective mechanism for evaluating the consistency between classification results and clinical practice. The systems cannot quantify the degree of deviation between theoretical classification results and actual clinical diagnoses, making it difficult to identify systemic biases in classification models during practical applications. Medical data anomaly detection is often limited to individual patients or isolated cases, failing to fully consider the interrelationships between departments and processes within medical institutions, and unable to identify the propagation paths and scope of impact of anomalies within the medical system. Resource allocation strategies are often based on fixed rules or empirical arrangements, unable to dynamically adjust monitoring intensity and intervention scope according to real-time anomaly risks. Existing systems need to address the optimization of the entire process from data classification to anomaly localization to resource scheduling, particularly by establishing quantitative analysis of classification deviations and an anomaly propagation mechanism based on system topology. This would enable precise localization from data anomalies to physical anomalies, and on this basis, the development of adaptive resource allocation strategies.

[0003] Existing technologies for processing medical big data still suffer from the following problems: classification models focus too much on global performance indicators, neglecting individual patient-level deviation analysis; anomaly detection is disconnected from the institution's topology, making it difficult to accurately locate the source of anomalies; and resource allocation strategies lack a dynamic adjustment mechanism linked to real-time risks. These limitations severely affect the practicality and effectiveness of medical data management systems, necessitating an innovative approach that can span the entire process of data classification, deviation analysis, anomaly localization, and resource regulation. Summary of the Invention

[0004] The purpose of this invention is to provide a medical data classification and management system based on big data to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a medical data classification and management system based on big data, the system comprising: The module includes a data integration and classification module, a deviation quantification and analysis module, a graph network anomaly localization module, and an adaptive decision control module. The data integration and classification module is used to receive historical medical data sets and real-time medical data streams, and to perform feature fusion and classification prediction through a deep learning model to generate theoretical classification results. The deviation quantification analysis module is used to perform multi-dimensional deviation calculation between theoretical classification results and actual clinical classification results, including time dimension deviation, feature dimension deviation and event sequence deviation, to generate a patient-level multi-dimensional deviation tensor. The graph network anomaly localization module is used to simulate anomaly propagation by inputting a multidimensional deviation tensor into a graph neural network based on the topology of the medical institution, and to generate an anomaly probability distribution map to identify the physical area where the anomaly occurs. The adaptive decision control module is used to dynamically allocate monitoring resources according to the anomaly probability distribution map, including starting a high-frequency data acquisition mode for high-probability anomaly areas and performing intervention tests on adjacent nodes.

[0006] Preferably, the data integration and classification module includes a data preprocessing unit, a feature extraction unit, and a classification model unit; The data preprocessing unit is used to perform noise filtering and format standardization on historical medical data sets and real-time medical data streams to obtain purified medical data. The feature extraction unit is used to extract multimodal features from the purified medical data, including using a convolutional neural network to extract the spatiotemporal features of physiological signal waveforms, using a recurrent neural network to extract the time-dependent features of laboratory test data, and using an attention mechanism to extract key event features of diagnostic event sequences. The classification model unit is used to input multimodal features into a deep belief network for joint training to output theoretical classification results.

[0007] Preferably, in the feature extraction unit, multimodal feature extraction includes: Wavelet transform decomposition of physiological signal waveforms is performed to extract energy entropy and zero-crossing rate features of different frequency bands; The laboratory test data were segmented into time series using a sliding window, and the statistical characteristics of each window were calculated, including mean, variance, and skewness. The sequence-to-sequence model is used to encode the diagnostic event sequence to capture long-term dependencies between events.

[0008] Preferably, the deviation quantification analysis module includes a time deviation calculation unit, a feature deviation calculation unit, and an event deviation calculation unit; The time deviation calculation unit is used to align the theoretical classification results and the actual clinical classification results at fixed time intervals, and uses a dynamic time warping algorithm to calculate the cumulative time deviation vector. The feature deviation calculation unit is used to estimate the kernel density of the feature distribution of theoretical classification results and actual clinical classification results, calculate the Jason-Shannon divergence in the feature space, and generate a feature deviation vector. The event bias calculation unit is used to compare the event sequences of theoretical classification results and actual clinical classification results using the edit distance algorithm, and generate an event bias vector; The deviation quantization analysis module fuses the time deviation vector, feature deviation vector, and event deviation vector into a tensor to generate a multidimensional deviation tensor.

[0009] Preferably, in the feature deviation calculation unit, kernel density estimation includes: The probability density of each feature dimension of the theoretical classification results and the actual clinical classification results is estimated using the Gaussian kernel function. Calculate the Jason-Shannon divergence between the two probability density functions as the feature bias value; The deviation values ​​of all feature dimensions are normalized to form a feature deviation vector.

[0010] Preferably, in the event deviation calculation unit, the edit distance algorithm uses the Levenstein distance to calculate the minimum number of edit operations between two event sequences, including insertion, deletion, and replacement operations, in order to quantify the event sequence deviation.

[0011] Preferably, the graph network anomaly localization module includes a topology modeling unit, an anomaly propagation simulation unit, and a probabilistic graph generation unit; The topology modeling unit is used to construct a graph structure based on the departmental connection relationships of a medical institution, where nodes represent departments or equipment, edges represent data flows or patient flow paths, and nodes are assigned load attributes and edges are assigned weight attributes. The anomaly propagation simulation unit is used to map the multidimensional deviation tensor to the corresponding node of the graph structure, perform feature propagation through the graph convolutional network, and use the random walk algorithm to simulate the diffusion path of anomalies in the graph structure. The probability map generation unit is used to count the frequency of abnormal propagation paths, calculate the abnormal probability value of each node in combination with node attributes, and generate an abnormal probability distribution map.

[0012] Preferably, in the anomaly propagation simulation unit, the graph convolutional network includes multiple layers of graph convolutional operations, each layer aggregating the features of adjacent nodes and performing nonlinear transformations through activation functions; The random walk algorithm starts from each node and performs multiple walks, recording the accumulation of outliers in the walk path to simulate anomaly propagation.

[0013] Preferably, the adaptive decision control module includes a resource scheduling unit and an intervention execution unit; The resource scheduling unit is used to analyze the anomaly probability distribution map. When the anomaly probability value of a certain area exceeds a preset threshold, it sends a control command to the monitoring equipment in that area to increase the data acquisition frequency and enable real-time feature tracking. The intervention execution unit is used to inject test signals into neighboring nodes with significant gradient changes in the anomaly probability distribution map and monitor the response to verify the anomaly correlation.

[0014] Preferably, the present invention also includes a medical data classification and management method based on big data, which includes all the modules and method processes of the above-mentioned medical data classification and management system based on big data.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This method calculates multi-dimensional biases between theoretical and clinical classification results, including time-dimensional bias, feature-dimensional bias, and event sequence bias, generating a patient-level multi-dimensional bias tensor. Time-dimensional bias reflects the stability of classification results over time, feature-dimensional bias reveals differences in the model's judgment of the importance of different features, and event sequence bias captures logical anomalies in the clinical diagnosis and treatment process. This multi-dimensional bias quantification method can comprehensively evaluate the performance of classification models in practical applications and discover systematic bias patterns that traditional single indicators cannot identify. Based on patient-level bias tensor analysis, classification anomalies can be accurately identified at the individual level, providing data support for subsequent targeted interventions.

[0016] Based on the topology of medical institutions, a multidimensional bias tensor is input into a graph neural network to simulate anomaly propagation. Graph neural networks can effectively model the complex relationships between departments and processes within a medical institution, simulating the propagation of anomalies along these networks. Anomaly propagation simulation generates an anomaly probability distribution map, identifying high-probability physical areas where anomalies occur, such as specific departments, treatment units, or equipment areas. This anomaly localization method based on system topology associates anomalies at the data level with their actual physical locations, achieving anomaly mapping from digital space to physical space. Monitoring resources are dynamically allocated according to the anomaly probability distribution map, and high-frequency data acquisition is initiated for high-probability anomaly areas to enhance monitoring of these areas. Simultaneously, intervention tests are performed on adjacent nodes of the anomaly area to prevent the spread of anomalies through network relationships. This dynamic resource allocation strategy based on anomaly probability distribution optimizes the allocation of monitoring resources, improving the timeliness of anomaly detection and the effectiveness of intervention. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the big data-based medical data classification management system described in this invention. Figure 2 A flowchart illustrating the working principle of the data integration and classification module; Figure 3 This is a flowchart illustrating the working principle of the deviation quantification analysis module. Figure 4A frequency distribution diagram of abnormal propagation paths in a medical data classification and management system; Figure 5 This is a comparison chart of resource scheduling collection frequencies. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a big data-based medical data classification management system. The system includes: a data integration and classification module that receives historical medical data sets and real-time medical data streams, uses a deep learning model for feature fusion and classification prediction, and generates theoretical classification results; a deviation quantification and analysis module that compares the theoretical classification results with actual clinical classification results, calculates time dimension deviation, feature dimension deviation, and event sequence deviation, and forms a patient-level multidimensional deviation tensor; a graph network anomaly localization module that, based on the topology of the medical institution, inputs the multidimensional deviation tensor into a graph neural network to simulate anomaly propagation, generates an anomaly probability distribution map, and thus identifies the physical areas where anomalies occur; and an adaptive decision control module that dynamically allocates monitoring resources according to the anomaly probability distribution map, including initiating high-frequency data acquisition modes for high-probability anomaly areas and performing intervention tests on adjacent nodes.

[0020] Example 1: See Figure 2In specific implementation, the data preprocessing unit of the data integration and classification module performs noise filtering on historical medical data sets and real-time medical data streams. Noise filtering employs a wavelet threshold-based denoising algorithm to remove high-frequency interference in physiological signals. Format standardization converts data from different medical devices into a standardized time-series format required for internal processing, aligning timestamps to millisecond precision and normalizing numerical ranges to the [0,1] interval through linear transformation. In specific implementation, the convolutional neural network of the feature extraction unit is configured as a multi-layer one-dimensional convolutional structure. The input is standardized physiological signal waveform data, such as electrocardiograms or electroencephalograms. The convolutional layers use the ReLU activation function followed by a max-pooling layer to progressively extract local spatiotemporal features from the waveforms. The recurrent neural network uses a gated recurrent unit structure to process laboratory test data sequences, such as time series of blood routine indicators. The memory unit of the gated recurrent unit retains long-term dependencies of historical information. The attention mechanism module encodes diagnostic event sequences, represented by International Classification of Disease codes. The attention mechanism calculates the attention weight of each event in the sequence, and the weight values ​​are used for weighted summation to generate key event feature vectors.

[0021] In practical implementation, the multimodal feature extraction process performs wavelet transform decomposition on the physiological signal waveform. The wavelet transform decomposes the signal into multiple frequency band sub-signals, and calculates the energy entropy feature for each frequency band sub-signal. The energy entropy characterizes the energy distribution complexity of the signal. Simultaneously, it calculates the zero-crossing rate feature, which reflects the frequency at which the signal waveform crosses zero points. For laboratory test data, a sliding window is applied for time series segmentation. The sliding window size is dynamically adjusted according to the data sampling frequency. Within each window, statistical features of the data are calculated, including the arithmetic mean, variance, and skewness. The mean describes the central tendency of the data, the variance measures the dispersion of the data, and the skewness indicates the direction and degree of asymmetry in the data distribution. Diagnostic event sequences are encoded using a sequence-to-sequence model, which includes an encoder network and a decoder network. The encoder network converts the input event sequence into a fixed-dimensional context vector, capturing long-term dependencies between events. The decoder network generates feature representations based on the context vector.

[0022] In practical implementation, the deep belief network of the classification model unit consists of multiple stacked restricted Boltzmann machine layers. The deep belief network receives multimodal feature vectors from the feature extraction unit as input. These multimodal feature vectors are the concatenation of features from convolutional neural networks, recurrent neural networks, and attention mechanisms. The training process of the deep belief network is divided into a pre-training phase and a fine-tuning phase. In the pre-training phase, each restricted Boltzmann machine layer undergoes unsupervised, layer-by-layer greedy training to learn the probability distribution of the input data. In the fine-tuning phase, the entire deep belief network is trained under supervision using backpropagation and labeled data to optimize the network weight parameters. Finally, the output layer uses the Softmax function to generate the probability distribution of the theoretical classification result. The theoretical classification result corresponds to a specific disease diagnosis category or patient health status classification. The format standardization operation of the data preprocessing unit also includes handling missing values. Missing values ​​are filled using a multiple imputation method. This multiple imputation method creates multiple complete datasets based on chain equations and estimates the missing values ​​through an iterative regression model. In some embodiments, the convolutional neural network of the feature extraction unit can be replaced with a temporal convolutional network, which uses causal convolution to ensure temporal order. The recurrent neural network can be replaced with a long short-term memory network, which enhances long-term memory capacity by introducing cell states and gating mechanisms.

[0023] Optionally, the wavelet transform decomposition of physiological signal waveforms during multimodal feature extraction can adjust the number of decomposition layers. The number of decomposition layers is determined based on the frequency components of the signal, and the energy entropy calculation uses the Shannon entropy formula applied to the wavelet coefficients. Optionally, the sliding window segmentation of laboratory test data can employ an overlapping window strategy. The overlapping window strategy sets the window step size to be smaller than the window length to capture more continuous temporal evolution patterns. The deep belief network of the classification model unit uses Dropout technology during training to prevent overfitting. Dropout technology randomly ignores some neuronal connections during training. It can be understood that the output of the theoretical classification result can be configured as a multi-label classification format, which allows an input instance to belong to multiple disease categories simultaneously.

[0024] Example 2: See Figure 3In practical implementation, the time deviation calculation unit of the deviation quantification analysis module aligns the theoretical classification results and the actual clinical classification results at fixed time intervals. The fixed time interval is set according to the data acquisition frequency, such as one data point per minute or every five minutes. The theoretical classification results come from the output of the data integration and classification module, while the actual clinical classification results come from the diagnostic results confirmed by doctors and recorded in the hospital information system. The dynamic time warping algorithm is applied to calculate the cumulative time deviation vector. This algorithm constructs a cost matrix to find the optimal curved path between two time series, minimizing the cumulative alignment cost. The cumulative time deviation vector records the sum of the absolute or squared differences between the theoretical and actual values ​​at each alignment point along this optimal path, thus forming a deviation sequence arranged in chronological order. When processing sequences with different lengths or local time scaling, the dynamic time warping algorithm effectively overcomes the limitations of simple Euclidean distance.

[0025] In practical implementation, the feature bias calculation unit performs kernel density estimation on the feature distributions of the theoretical classification results and the actual clinical classification results. Kernel density estimation is a non-parametric probability density function estimation method. Kernel density estimation uses a Gaussian kernel function to estimate the probability density of each feature dimension of the theoretical and actual clinical classification results. The bandwidth parameter of the Gaussian kernel function is determined through the Silverman rule or cross-validation to obtain a smooth and unbiased probability density curve. The Jason-Shannon divergence between the two probability density functions is calculated as the feature bias value. The Jason-Shannon divergence is a symmetric variant of the KL divergence, with a bounded range, making it more suitable for comparing two distributions. The feature bias calculation unit normalizes the bias values ​​of all feature dimensions to form a feature bias vector. The normalization process typically uses a min-max scaling method to linearly transform the Jason-Shannon divergence value of each dimension to the [0,1] interval, making the biases of different feature dimensions comparable.

[0026] In its implementation, the event bias calculation unit uses an edit distance algorithm to compare the event sequences of theoretical classification results with those of actual clinical classification results. An event sequence refers to diagnostic, treatment, or examination event codes arranged chronologically. The edit distance algorithm employs the Levinstein distance, defined as the minimum number of single-character edit operations required to transform one event sequence into another. Single-character edit operations include inserting, deleting, or replacing an event. The event bias calculation unit calculates the Levinstein distance using dynamic programming. Dynamic programming constructs a two-dimensional matrix, where each element stores the minimum edit distance between sub-sequences. The value in the lower right corner of the final matrix is ​​the Levinstein distance for the entire sequence. This distance value is directly used as a quantitative indicator of event sequence bias and constitutes the core element of the event bias vector.

[0027] In its implementation, the deviation quantification analysis module fuses the time deviation vector, feature deviation vector, and event deviation vector into a multidimensional deviation tensor. This tensor fusion operation involves vector concatenation, dimensionality enhancement, and standardization. First, the time deviation vector, feature deviation vector, and event deviation vector are concatenated into a high-dimensional composite vector, the dimension of which is the sum of the dimensions of the three vectors. Then, this composite vector is reshaped into a three-dimensional tensor structure, with its three dimensions corresponding to the time index, feature index, and event sequence index, respectively. Finally, each element in the three-dimensional tensor is Z-score standardized. Z-score standardization ensures that deviation values ​​from different sources have zero mean and unit variance, thus generating the final patient-level multidimensional deviation tensor, which serves as input data for the graph network anomaly localization module.

[0028] In some embodiments, the fixed time interval used by the time bias calculation unit can be dynamically adjusted according to the data update frequency of the actual clinical classification results. For example, if the actual clinical classification results are updated once per hour, the time interval can be set to the hourly level. Optionally, when calculating the Levinstein distance, the event bias calculation unit can assign different operation weights to insertion, deletion, and replacement operations. For example, replacement operations can be assigned a higher cost weight than insertion or deletion operations to reflect the differences in the importance of different event changes in clinical practice. Optionally, in the tensor fusion process, in addition to Z-score normalization, a decimal scaling normalization method can also be used to scale the composite vector. Decimal scaling normalization achieves normalization by shifting the decimal point position of the data.

[0029] It is understandable that the cumulative time deviation vector calculated by the time deviation calculation unit not only reflects the overall deviation trend, but its local extrema may also indicate significant classification discrepancies at specific time points. Similarly, the feature deviation vector generated by the feature deviation calculation unit helps identify which feature dimensions show the greatest difference between predictions and actual situations, providing a clear direction for model optimization. The event deviation vector generated by the event deviation calculation unit reveals the differences between theoretical prediction processes and actual treatment processes from a clinical pathway perspective.

[0030] Example 3: In specific implementation, the kernel density estimation process of the feature bias calculation unit uses a Gaussian kernel function to estimate the probability density of each feature dimension of the theoretical classification results and the actual clinical classification results. The feature values ​​of the theoretical classification results constitute one sample set, and the feature values ​​of the actual clinical classification results constitute another sample set. For continuous feature dimensions, the Gaussian kernel function constructs a normal distribution centered on each data point. The entire probability density function is obtained by weighted averaging of the Gaussian distributions corresponding to all data points. The value of the probability density function at any point reflects the density of the data distribution near that point. The bandwidth parameter of the Gaussian kernel function controls the smoothness of the probability density estimation curve. If the bandwidth parameter is too small, the probability density function will overfit and exhibit a multi-peak sawtooth shape. If the bandwidth parameter is too large, the probability density function will be overly smoothed, masking the true data distribution characteristics. The optimal bandwidth parameter is usually determined based on empirical rules or through cross-validation. After completing the probability density estimation, the Jason-Shannon divergence between the two probability density functions is calculated as the feature bias value. The Jason-Shannon divergence is a symmetric KL divergence, and its calculation is based on the arithmetic mean of the two probability distributions. The Jason-Shannon divergence value is always a non-negative real number. It is zero when two probability distributions are identical, and increases accordingly as the difference between the two distributions increases. The feature deviation calculation unit normalizes the deviation values ​​for all feature dimensions, forming a feature deviation vector. The normalization process uses a min-max scaling method to linearly transform the Jason-Shannon divergence value of each feature dimension to the interval between zero and one, making feature deviations of different dimensions or orders of magnitude comparable.

[0031] In its implementation, the edit distance algorithm of the event deviation calculation unit uses the Levinstein distance to calculate the minimum number of edit operations between two event sequences. The event sequence consists of a series of medical event codes arranged chronologically. The Levinstein distance is defined as the minimum number of single-step edit operations required to transform one event sequence into another. Single-step edit operations are limited to three basic types: inserting a new medical event into the sequence, deleting an existing medical event from the sequence, and replacing one medical event in the sequence with another different medical event. The edit distance algorithm is implemented through dynamic programming. Dynamic programming constructs a two-dimensional distance matrix. The number of rows in the distance matrix equals the length of the source event sequence plus one, and the number of columns equals the length of the target event sequence plus one. The initialization rule for the distance matrix is ​​that the first row and first column are filled with integer values ​​from zero to the sequence length. The value of each remaining element in the distance matrix is ​​recursively determined by the values ​​of its three adjacent elements above, to the left, and to the upper left. The specific calculation formula is as follows: ; in: This indicates the beginning of the source event sequence. The events are transformed into a target event sequence. The minimum editing cost required for each event Represents the first in the source event sequence One event, Represents the first in the target event sequence One event. , and These represent the fixed cost weights assigned to performing one insertion, deletion, and replacement operation, respectively. It is an indicator function, when and The function value is one when they are not both, and zero when they are both. The minimum number of edit operations is ultimately given by the value of the bottom right element of the distance matrix, which is directly used as the quantification result of the event sequence bias.

[0032] In some embodiments, when the feature bias calculation unit performs kernel density estimation, the bandwidth parameter of the Gaussian kernel function can be automatically calculated using the Scott rule. The Scott rule assumes that the data follows a normal distribution and derives the bandwidth value based on the sample size and standard deviation. In some embodiments, when the event bias calculation unit calculates the Levinstein distance, the cost weights of insertion, deletion, and replacement operations can be set to equal values, for example, all set to one. In this case, the Levinstein distance simplifies to the minimum number of single-step operations. In some embodiments, for sequences containing specific clinically significant events (such as critical value events), the event bias calculation unit can assign higher cost weights to replacement operations involving such events to amplify their bias contribution.

[0033] Optionally, before calculating the Jason-Shannon divergence, the feature bias calculation unit can perform a smoothness test on the estimated probability density function. The smoothness test determines whether the curve is excessively volatile by calculating the second derivative of the probability density function. If the smoothness is insufficient, the bandwidth parameter is automatically increased and the kernel density estimation is performed again. Through kernel density estimation and Jason-Shannon divergence calculation, the feature bias calculation unit can capture subtle differences in the feature distribution patterns between theoretical classification results and actual clinical classification results, such as distribution skewness and multimodal characteristics. It can be understood that the Levenstein distance used by the event bias calculation unit can effectively measure the overall difference between two event sequences in terms of order and content, rather than simply comparing overlap or common subsequences. The concept of the number of editing operations intuitively reflects the degree of "effort" required to adjust the predicted event sequence to the actual event sequence. The feature bias vector, event bias vector, and time bias vector together constitute the output basis of the bias quantification analysis module. The multidimensional bias tensor generated after tensor fusion provides multi-faceted bias information input for subsequent graph network analysis.

[0034] Suppose a hospital admits a patient suspected of having heart failure. The data integration and classification module generates a theoretical classification result based on the patient's vital sign monitoring data, laboratory test results, and preliminary diagnostic records within 24 hours of admission, predicting that the patient belongs to the "high-risk heart failure" category. However, the clinical team, based on a more comprehensive clinical assessment, ultimately classifies the patient into the "intermediate-risk heart failure" category. The feature bias calculation unit begins processing this case. The feature vector corresponding to the theoretical classification result includes ten key feature dimensions, such as BNP hormone level, left ventricular ejection fraction, and respiratory rate variability. The actual clinical classification result corresponds to another set of feature values ​​obtained from actual measurements. For the continuous feature dimension of BNP hormone level, the feature value distribution of the theoretical classification result is derived from the model training data, while the feature values ​​of the actual clinical classification result come from the actual values ​​of multiple blood tests conducted on the patient. The kernel density estimation process uses a Gaussian kernel function to estimate the probability density of both sets of feature values. The bandwidth parameter of the Gaussian kernel function is automatically calculated and determined using the Silverman rule, generating two smooth probability density curves. Jason-Shannon divergence calculations revealed significant differences between the two probability density curves along the BNP feature dimension. These differences primarily stemmed from the fact that theoretical predictions were generally higher than actual measurements. This divergence value, after normalization, became the deviation value corresponding to the BNP dimension in the feature deviation vector. The feature deviation calculation unit repeatedly performed the kernel density estimation and divergence calculation process on all ten feature dimensions, ultimately forming a ten-dimensional feature deviation vector.

[0035] The event bias calculation unit simultaneously processes the patient's event sequence data. The event sequence corresponding to the theoretical classification result is generated by model prediction, and the sequence is: ["Hospital Admission", "Basic Vital Signs Monitoring", "Venous Blood Test", "Chest X-ray", "Echocardiography", "Start Intravenous Diuretic Treatment"]. The event sequence corresponding to the actual clinical classification result comes from the real records of the hospital information system, and the sequence is: ["Hospital Admission", "Basic Vital Signs Monitoring", "Venous Blood Test", "Start Oral Diuretic Treatment"]. The event bias calculation unit uses the Levenstein distance to calculate the minimum number of edit operations between two event sequences. Edit operations include insertion, deletion, and replacement. A dynamic programming algorithm constructs a 7x5 two-dimensional distance matrix. The rows of the matrix correspond to the 6 events of the theoretical sequence plus the initial empty state, and the columns correspond to the 4 events of the actual sequence plus the initial empty state. The distance matrix is ​​filled starting from the top left corner. After initializing the first row and the first column, the algorithm progressively calculates the minimum edit cost for each position. For example, when comparing the "chest X-ray examination" event in the theoretical sequence with the "start oral diuretic treatment" event in the actual sequence, the cost of the replacement operation needs to be calculated due to the different event types. The final calculated Levenstein distance value is 3, meaning that at least three edit operations are required to convert the theoretical event sequence into the actual event sequence. This distance value is directly used as the quantification result of the event sequence deviation, forming the core element of the event deviation vector.

[0036] In practical implementation, when the feature bias calculation unit estimates the kernel density for each feature dimension, it needs to ensure the integrity of the input data. For any missing test values ​​that may exist in the actual clinical classification results, they are imputed using previous valid measurements from the same patient to ensure the accuracy of the probability density estimation. After the Jason-Shannon divergence calculation is completed, the feature bias calculation unit uses a min-max scaling method to normalize the divergence values ​​of all feature dimensions to the range of zero to one, making the feature biases of different medical dimensions comparable. For example, the bias of BNP level is placed on the same scale as the bias of respiratory rate. When processing event sequences, the event bias calculation unit uses a unified standardized code for all medical events to ensure the accuracy of event comparisons.

[0037] Example 4: In a specific implementation, the topology modeling unit of the graph network anomaly localization module constructs a graph structure based on the departmental connections of the medical institution. These connections are derived from the hospital's administrative organizational chart and actual patient flow records. In the graph structure, nodes represent specific medical departments or key medical equipment, such as the emergency department, radiology department, intensive care unit, or CT scanner. Edges represent data flow paths between departments or actual patient flow paths. The topology modeling unit assigns a load attribute to each node, which characterizes the amount of data processed or the number of patients treated by the department represented by that node per unit time. Simultaneously, it assigns a weight attribute to each edge, which reflects the data transmission frequency or patient flow intensity of the path between two nodes. The completed graph structure is stored in the system in the form of an adjacency matrix and a node attribute matrix. The adjacency matrix is ​​a sparse matrix used to record the connections between nodes, while the node attribute matrix stores multi-dimensional attribute information for each node row-wise, including but not limited to load attributes, department type, and geographic location code.

[0038] In its implementation, the anomaly propagation simulation unit maps multidimensional bias tensors to corresponding nodes in the graph structure. The mapping is based on matching the departmental identifiers of the patient data sources. After mapping, the temporal bias vector, feature bias vector, and event bias vector contained in the multidimensional bias tensor are concatenated into the initial feature vector of each node. The anomaly propagation simulation unit propagates node features through a graph convolutional network (GCNN). This GCNN contains multiple layers of graph convolution operations. Each layer aggregates the feature information of the target node's first-order neighbors. The aggregation method typically involves a weighted summation of the neighbor node features, with the weights determined by the adjacency matrix and degree matrix. The aggregated feature vector is then combined with the target node's own feature vector and transformed using a non-linear activation function, such as ReLU. By stacking multiple layers of graph convolution operations, each node can aggregate indirect feature information from its multi-hop neighbors, thus simulating the propagation effect of anomalous features on the graph structure. The anomaly propagation simulation unit also uses a random walk algorithm to simulate the diffusion path of anomalies in a graph structure. The random walk algorithm starts from each node containing a non-zero deviation value and performs a random walk with a fixed step size based on the transition probability determined by the edge weight attribute. The higher the edge weight, the greater the probability that the walk will choose that path. Each walk records all the nodes passed on the path and accumulates the deviation values ​​corresponding to these nodes, thereby simulating the potential trajectory of anomalies spreading along the network path.

[0039] In practical implementation, referring to Table 1, the probabilistic graph generation unit counts the frequency of anomaly propagation paths, which are derived from a large number of path instances generated by the random walk algorithm. The probabilistic graph generation unit calculates the anomaly probability value for each node based on its node attributes. The calculation process considers two main factors: first, the total frequency of the node's appearance in all random walk paths; and second, the node's own load attribute value. Nodes with higher loads are considered to have a greater influence in anomaly propagation. The anomaly probability value can be calculated based on Bayes' theorem or simple frequency statistics. For example, the anomaly probability of a node can be proportional to the ratio of the number of anomaly propagation paths passing through that node to the total number of paths, further weighted by its load attribute. Finally, an anomaly probability distribution map is generated. This map is a heatmap overlaid on the original organizational topology map. Each node in the map is assigned a different color depth or size based on its calculated anomaly probability value. Darker colors or larger nodes indicate a higher anomaly probability in the area where the node is located, thus visually identifying the physical area where anomalies occur, such as identifying a nursing station on the third floor of the inpatient department where anomalies are likely to be prevalent. The anomaly probability distribution map is output as a digital image and associated data file for use by the adaptive decision control module.

[0040] Table 1: Example of Node Attributes and Anomaly Probability Calculation

[0041] In some embodiments, during feature propagation in a graph convolutional network, a batch normalization layer can be added after each graph convolutional operation to accelerate the convergence of the training process and improve stability. The batch normalization layer standardizes the feature vector of each node in the same batch of data, making its mean zero and variance one. The random walk algorithm can employ a restarted random walk strategy. This strategy jumps back to the starting node with a preset probability at each step, helping to maintain focus on the local area surrounding the starting node while exploring the global network structure. Optionally, when calculating the anomaly probability value, the probabilistic graph generation unit can introduce the edge weight attribute as an adjustment factor, in addition to the node load attribute. For example, two nodes connected by a high-weight edge may have a stronger mutual influence on their anomaly probabilities.

[0042] It is understandable that the graph network anomaly localization module, by mapping abstract deviation data onto a concrete institutional topology and simulating the propagation process of anomalies, can transform the patient-level classification bias problem into a spatial anomaly region localization problem, providing a clear decision-making basis for subsequent precise resource allocation. The anomaly probability distribution map not only identifies high-incidence points of anomalies but also reveals potential anomaly propagation chains through the connections between nodes.

[0043] See Figure 4This diagram is a key visualization result of the anomaly propagation simulation and analysis in the graph network anomaly localization module. The horizontal axis displays typical departmental propagation paths within a medical institution, covering six categories: Emergency Department → Intensive Care Unit, Cardiology Department → Central Laboratory, Emergency Department → Radiology Department, Intensive Care Unit → Operating Room, Inpatient Department → Pharmacy, and Outpatient Department → Radiology Department. The vertical axis represents the propagation path frequency, quantifying the number of times anomalies propagate along each path. By quantifying the propagation path frequency, this diagram provides the graph network anomaly localization module with an intuitive reference for anomaly diffusion hot paths, helping the system identify key links in anomaly propagation and clarifying key areas for subsequent resource scheduling and intervention decisions. It fully demonstrates the practical value of topology-based anomaly propagation simulation in medical data management.

[0044] Example 5: In specific implementation, the resource scheduling unit of the adaptive decision control module continuously parses the anomaly probability distribution map from the graph network anomaly localization module. The anomaly probability distribution map is input in the form of a set of nodes with probability values. The resource scheduling unit internally sets a configurable preset probability threshold, which is determined based on historical anomaly data analysis or domain expert experience; for example, the threshold is set to 0.7. The resource scheduling unit traverses all nodes in the anomaly probability distribution map. When it detects that the anomaly probability value of a node exceeds the preset probability threshold, the resource scheduling unit determines that the physical area represented by that node is a high-probability anomaly area. The resource scheduling unit generates control commands, which are sent via the hospital's internal network protocol to the monitoring equipment deployed in the high-probability anomaly area. The monitoring equipment includes intelligent bedside monitors, continuous vital signs sensors, or laboratory automated production line control terminals. The control commands contain specific parameter adjustment content. After the command is issued, the monitoring equipment in the target area will switch from the conventional intermittent data acquisition mode to a high-frequency data acquisition mode, for example, increasing the sampling frequency of electrocardiogram monitoring from 125Hz to 500Hz, and shortening the upload interval of blood oxygen saturation data from 5 minutes to 30 seconds. Meanwhile, the resource scheduling unit enables real-time feature tracking. This function requires the monitoring equipment to perform real-time calculations and streaming uploads of preset key features while collecting high-frequency data, ensuring that the data integration and classification module can obtain more timely data for classification and prediction.

[0045] In practice, the intervention execution unit performs intervention tests on adjacent nodes with significant gradient changes in the anomaly probability distribution map. The significance of the gradient change is determined by calculating the spatial derivative of the anomaly probability values ​​between nodes. The intervention execution unit first identifies node pairs that are directly connected topologically, and whose probability difference on the connecting edges exceeds a set threshold. For example, if node A has an anomaly probability of 0.75 and its neighbor node B has an anomaly probability of 0.25, a difference of 0.5 is considered a significant gradient change. The intervention execution unit then injects test signals into nodes with lower anomaly probabilities (such as node B). These test signals simulate normal or typical abnormal data patterns. For example, injecting a set of simulated test results conforming to quality control standards into a laboratory information system node, or injecting a simulated CT image labeled with specific imaging features into a PACS node. After injecting the test signal, the intervention execution unit simultaneously initiates response monitoring of the relevant nodes and their connected paths. Response monitoring includes observing whether the data flow is transmitted normally, monitoring whether abnormal records appear in the monitoring equipment's operational status logs, and checking for temporary deviations between theoretical classification results and actual clinical classification results after this test. By analyzing the differences in system response before and after the injection of test signals, the intervention execution unit can verify whether there is a real abnormal correlation between high-probability abnormal nodes and adjacent nodes, thereby distinguishing between local independent anomalies and network propagation anomalies.

[0046] In some embodiments, the control commands sent by the resource scheduling unit may include device resource wake-up commands. When the anomaly probability value of a certain area drops from a high level and stabilizes below the threshold for a period of time, the resource scheduling unit will send a command to switch the data acquisition frequency of the relevant monitoring equipment from high-frequency mode back to normal mode, so as to achieve dynamic energy saving of computing and communication resources. In some embodiments, the intervention execution unit may inject test signals using a gradual strategy. The gradual strategy first injects low-intensity test signals, which are close to the normal data range. If the system does not produce the expected response or the response is weak, the intensity of the test signal or the obviousness of the abnormal characteristics is gradually increased until a significant change in response is observed or a safety threshold is reached.

[0047] Optionally, when increasing the data collection frequency, the resource scheduling unit can adjust the frequency in stages based on the specific anomaly probability value. For example, when the anomaly probability value is between 0.7 and 0.85, the data collection frequency can be increased to twice the normal mode; when the anomaly probability value exceeds 0.85, it can be increased to four times the normal mode, achieving refined control of resource allocation. Optionally, when monitoring responses, the intervention execution unit can introduce third-party system logs as auxiliary verification information. For example, after injecting a test signal into the pharmacy node, in addition to monitoring the data flow of its own system, it can also query whether the drug inventory update records in the hospital ERP system are synchronized, thereby enhancing the reliability of the verification. Optionally, the preset probability threshold can be designed as a dynamic threshold. The dynamic threshold adaptively adjusts according to the overall operational load of the hospital, appropriately increasing the threshold during high-load periods to avoid excessive false alarms leading to resource strain.

[0048] Understandably, the adaptive decision control module, through proactive monitoring and instruction issuance by the resource scheduling unit, achieves dynamic, on-demand allocation of monitored resources based on anomaly risks, avoiding the average allocation and waste of resources. The testing and verification mechanism of the intervention execution unit provides a feedback loop for anomaly localization results, helping to confirm the root cause of the anomaly and reduce the risk of misjudgment. This enables the entire system not only to discover problems but also to proactively take measures for investigation and verification, improving the system's intelligence and decision reliability.

[0049] See Figure 5 This diagram provides a clear and quantitative representation of the resource scheduling strategy of the adaptive decision control module. Through dual-dimensional visualization, it clearly explains how the system dynamically adjusts monitoring resources based on anomaly probabilities. The horizontal axis represents monitoring modes, covering five categories: routine, low-frequency, medium-frequency, high-frequency, and ultra-high-frequency. The left vertical axis represents sampling frequency, displayed as a gray bar chart; the right vertical axis represents upload intervals, presented as a black broken line. This dual-indicator linkage design satisfies both the timeliness and resource optimization requirements of medical data management. It also visually demonstrates the core mechanism of the adaptive decision control module's on-demand scheduling and dynamic balancing, enhancing data density and real-time performance in high-risk areas to provide timely support for anomaly intervention, while maintaining routine monitoring in low-risk areas to avoid resource redundancy.

[0050] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A big data-based medical data classification management system, characterized by, The system comprises a data integration and classification module, a bias quantification analysis module, a graph network anomaly positioning module, and an adaptive decision control module. The data integration and classification module is configured to receive a historical medical data set and a real-time medical data stream, and perform feature fusion and classification prediction through a deep learning model to generate a theoretical classification result. The bias quantification analysis module is configured to perform multi-dimensional bias calculation on the theoretical classification result and the clinical actual classification result, including time dimension bias, feature dimension bias, and event sequence bias, to generate a multi-dimensional bias tensor at the patient level. The graph network anomaly positioning module is configured to input the multi-dimensional bias tensor into a graph neural network based on the topological structure of the medical institution to perform anomaly propagation simulation and generate an anomaly probability distribution map to identify the physical area where the anomaly occurs. The adaptive decision control module is configured to dynamically allocate monitoring resources according to the anomaly probability distribution map, including starting a high-frequency data acquisition mode for high-probability anomaly areas and performing intervention tests on adjacent nodes.

2. The big data based medical data classification management system according to claim 1, wherein, The data integration and classification module comprises a data preprocessing unit, a feature extraction unit, and a classification model unit. The data preprocessing unit is configured to perform noise filtering and format standardization on the historical medical data set and the real-time medical data stream to obtain purified medical data. The feature extraction unit is configured to perform multi-modal feature extraction on the purified medical data, including using a convolutional neural network to extract the spatio-temporal features of physiological signal waveforms, using a recurrent neural network to extract the time-dependent features of laboratory test data, and using an attention mechanism to extract the key event features of diagnosis event sequences. The classification model unit is configured to input the multi-modal features into a deep belief network for joint training to output the theoretical classification result. 3.The big data-based medical data classification management system of claim 2, wherein, In the feature extraction unit, the multi-modal feature extraction includes: Performing wavelet transform decomposition on physiological signal waveforms to extract energy entropy and zero-crossing rate features of different frequency bands; Using a sliding window to segment the time series of laboratory test data to calculate statistical features including mean, variance, and skewness for each window; Using a sequence-to-sequence model to encode the diagnosis event sequence to capture the long-term dependence between events.

4. The big data based medical data classification management system according to claim 1, wherein, The bias quantification analysis module comprises a time bias calculation unit, a feature bias calculation unit, and an event bias calculation unit. The time bias calculation unit is configured to align the theoretical classification result and the clinical actual classification result at fixed time intervals, and calculate the cumulative time bias vector using the dynamic time warping algorithm. The feature bias calculation unit is configured to perform kernel density estimation on the feature distribution of the theoretical classification result and the clinical actual classification result, calculate the Jensen-Shannon divergence in the feature space, and generate a feature bias vector. The event bias calculation unit is configured to compare the event sequences of the theoretical classification result and the clinical actual classification result using the edit distance algorithm to generate an event bias vector. The bias quantification analysis module tensor-fuses the time bias vector, the feature bias vector, and the event bias vector to generate a multi-dimensional bias tensor.

5. The big data based medical data classification management system as claimed in claim 4, wherein, In the feature bias calculation unit, the kernel density estimation includes: Using a Gaussian kernel function to estimate the probability density of each feature dimension of the theoretical classification result and the clinical actual classification result. calculating the Jensen-Shannon divergence between two probability density functions as the feature bias value; normalizing the bias values of all feature dimensions to form a feature bias vector.

6. The big data based medical data classification management system according to claim 4, wherein, In the event bias calculation unit, the edit distance algorithm adopts Levenshtein distance to calculate the minimum number of editing operations between two event sequences, including insertion, deletion and replacement operations, to quantify the event sequence bias.

7. The big data based medical data classification management system according to claim 1, wherein, The graph network anomaly positioning module includes a topology modeling unit, an anomaly propagation simulation unit and a probability graph generation unit. The topology modeling unit is used to construct a graph structure according to the department connection relationship of the medical institution, wherein the node represents the department or device, the edge represents the data flow or patient flow path, and the node is assigned with load attribute and edge weight attribute. The anomaly propagation simulation unit is used to map the multi-dimensional bias tensor to the corresponding node of the graph structure, propagate the features through the graph convolution network, and simulate the diffusion path of the anomaly in the graph structure using the random walk algorithm. The probability graph generation unit is used to count the frequency of the anomaly diffusion path, calculate the anomaly probability value of each node combined with the node attribute, and generate an anomaly probability distribution graph. 8.The big data-based medical data classification management system of claim 7, wherein, In the anomaly propagation simulation unit, the graph convolution network includes multiple layers of graph convolution operations, each layer aggregates the features of adjacent nodes and performs nonlinear transformation through the activation function. The random walk algorithm performs multiple walks from each node and records the accumulation of abnormal values in the walk path to simulate the propagation of anomalies. 9.The big data-based medical data classification management system of claim 1, wherein, The adaptive decision control module includes a resource scheduling unit and an intervention execution unit. The resource scheduling unit is used to analyze the anomaly probability distribution graph, and when the anomaly probability value of a certain area exceeds the preset threshold, sends a control instruction to the monitoring device in the area to increase the data acquisition frequency and enable real-time feature tracking. The intervention execution unit is used to inject test signals into adjacent nodes with significant gradient changes in the anomaly probability distribution graph and monitor the response to verify the anomaly association. 10.A method for managing medical data classification based on big data, characterized in that, All modules and method processes of the medical data classification management system based on big data according to any one of claims 1-9 are included.

Images