Whole-process quality control management system based on multi-source data fusion of medical units

Abstract

The application relates to the technical field of medical equipment quality control, in particular to a whole-process quality control management system based on medical unit multi-source data fusion. The system comprises a clinical multi-modal data module: obtaining an electronic medical record diagnosis and treatment text set, a medical equipment log set and an image lesion slice set, and outputting a clinical feature matrix set; a medical quality control network construction module: extracting device node state parameters and medical topology correlation parameters contained in the clinical feature matrix set, connecting through topology projection, and outputting a medical quality control directed graph; a reconstruction deviation module: carrying out neighborhood aggregation and inverse projection on the medical quality control directed graph, outputting a reconstruction deviation set through deviation derivation; a clinical abnormality identification and blocking module: executing distribution fitting identification on the reconstruction deviation set, outputting a medical abnormality identification parameter, and executing port blocking. Through the strong coupling architecture of medical data sensing, medical atlas deviation derivation and bottom-layer terminal physical blocking, the application reduces the probability of false triggering isolation and shortens the link delay of abnormal terminal disablement.

Description

Technical Field

[0001] This invention relates to the field of medical equipment quality control technology, specifically to a full-process quality control management system based on the fusion of multi-source data from medical institutions. Background Technology

[0002] In complex clinical scenarios, feature sequence aggregation and full-cycle data flow quality control of multimodal healthcare data across multiple terminals are important technological applications in the field of medical informatics. Current technologies generally employ a system management approach that combines single-device data flow review with static clinical data rule comparison. This involves monitoring changes in physiological indicators or test data collected from a single type of medical terminal. When a specific risk threshold is reached, the system automatically identifies and determines the status of all data anomalies or operational violations within a pre-defined standardized clinical information flow chain. This is currently the mainstream technological approach for controlling data quality and operational reliability in medical information systems.

[0003] However, existing technologies have significant shortcomings in the depth of cross-medical terminal information aggregation and the tracking of continuous clinical data, with data silos being a common phenomenon within medical systems. Clinical electronic medical records, medical equipment operating status, and medical images are often managed separately by independent subsystems. Existing quality control assessments heavily rely on data slices from single business lines, making it difficult to effectively integrate clinical information from different dimensions. This results in the management system lacking a global perspective for assessing complex medical events. Existing anomaly identification mechanisms are ill-suited to dynamic and continuous clinical operations, easily leading to frequent false alarms due to fluctuations in single indicators and failing to track continuous status changes during business processes. When faced with complex concurrent medical situations, this linear rule-based comparison method struggles to capture cascading risks between upstream and downstream processes, resulting in a persistently high false negative rate in overall quality control. Furthermore, existing systems lack the ability to perform deep, structured mapping of heterogeneous medical data. When faced with massive amounts of heterogeneous data, they cannot identify unknown abnormal state deviations, and most alerts remain at the software business level, leading to significant execution lag in the quality control system.

[0004] To address this, a full-process quality control management system based on the fusion of multi-source data from medical institutions is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a full-process quality control management system based on the fusion of multi-source data from medical units, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a full-process quality control management system based on multi-source data fusion in medical institutions, comprising: Clinical multimodal data module: acquires electronic medical record diagnosis and treatment text sets, medical device log sets, and image lesion slice sets, inputs them into a multi-source parsing model to perform cross-modal attention allocation and weighted concatenation, and outputs a set of clinical feature matrices; The medical quality control network construction module extracts the device node status parameters and medical topology association parameters contained in the clinical feature matrix set; the device node status parameters are vectorized through a node mapper to output the quality control node set; the medical topology association parameters are input into a Gaussian kernel function to perform similarity conversion, outputting a marginal weight matrix, which is then connected to the quality control node set via topological projection to output a directed medical quality control graph. The reconstructed deviation module: inputs the medical quality control directed graph into the feature aggregation network for neighborhood aggregation and outputs a high-order node matrix; inputs the high-order node matrix into the decoding network layer for inverse projection and outputs a reconstructed node matrix; performs deviation derivation with the high-order node matrix and outputs a reconstructed deviation set. Clinical anomaly identification and blocking module: Performs distribution fitting identification on the reconstructed deviation set and outputs medical anomaly identification parameters; generates a network blocking order based on the medical anomaly identification parameters and transmits it to the corresponding target terminal to perform port blocking.

[0007] Preferably, the specific generation process of the clinical feature matrix set includes: the multi-source parsing model includes a medical record semantic parsing branch, a device temporal parsing branch, and an image spatial parsing branch; inputting the electronic medical record diagnosis and treatment text set into the medical record semantic parsing branch to perform semantic mapping operation, and outputting a medical record semantic feature matrix; inputting the medical device log set into the device temporal parsing branch to perform feature aggregation operation, and outputting a device vital signs temporal matrix; inputting the medical image lesion slice set into the image spatial parsing branch to perform visual extraction operation, and outputting an image spatial feature matrix; inputting the medical record semantic feature matrix, the device vital signs temporal matrix, and the image spatial feature matrix into a cross-attention layer to perform dot product conversion, and outputting a modal attention score; performing a multiplication weighting operation on the modal attention score and the image spatial feature matrix, and outputting an image weighted feature set; performing channel dimension concatenation on the medical record semantic feature matrix, the device vital signs temporal matrix, and the image weighted feature set, and outputting a clinical feature matrix set.

[0008] Preferably, the specific generation process of the quality control node set includes: performing principal component analysis orthogonal dimensionality reduction on the clinical feature matrix set to output device node state parameters and medical topology association parameters; inputting the device node state parameters into a node mapper containing a linear projection layer and a nonlinear activation layer, performing historical mean interpolation on the device node state parameters, and outputting a complete state matrix; inputting the complete state matrix into a denoising autoencoder to perform background white noise filtering, and outputting a clean state matrix; extracting and calculating the high-frequency component parameters corresponding to the clean state matrix through Fourier transform; inputting the high-frequency component parameters and the clean state matrix into the nonlinear activation layer to perform high-dimensional manifold transformation, and outputting a feature manifold matrix; inputting the feature manifold matrix into an attention pool to perform local feature compression, and outputting the quality control node set.

[0009] Preferably, the specific construction process of the medical quality control directed graph includes: extracting the diagnosis and treatment time sequence occurrence stamp parameters and physical spatial location parameters included in the medical topology association parameters; performing absolute difference conversion on the diagnosis and treatment time sequence occurrence stamp parameters to output a collaborative delay matrix; calculating the displacement deviation square parameter corresponding to the physical spatial location parameter; performing square root derivation on the displacement deviation square parameter to output a spatial distance matrix; performing linear weighted summation on the collaborative delay matrix and the spatial distance matrix to output a joint distance matrix; combining the joint distance matrix with natural exponential negative decay to output a spatiotemporal decay matrix; calculating the feature covariance parameter corresponding to the quality control node set and the respective feature standard deviation parameters of the source node and the target node; dividing the feature covariance parameter by the product of the feature standard deviation parameters of the source node and the target node to output a feature correlation matrix; performing a Hadamard product operation on the feature correlation matrix and the spatiotemporal decay matrix to output a marginal weight matrix; performing topological mapping and marginal connection operations on the quality control node set and the marginal weight matrix to output a medical quality control directed graph.

[0010] Preferably, the specific generation process of the higher-order node matrix includes: a feature aggregation network extracting the source node feature matrix and the target node feature matrix contained in the directed graph of medical quality control; inputting the source node feature matrix and the target node feature matrix into a gated attention kernel to perform channel saliency discrimination operation, and outputting an attention weight matrix; obtaining the cascade fault threshold of medical equipment; inputting the cascade fault threshold of medical equipment and the attention weight matrix into a sparse truncate unit to perform low-relevance edge filtering operation, and outputting a sparse weight matrix; performing neighborhood feature aggregation with the sparse weight matrix and the directed graph of medical quality control, and outputting an aggregated feature matrix; inputting the aggregated feature matrix into a nonlinear activation kernel using a modified linear unit with leakage to perform gradient diffusion suppression, and outputting a higher-order node matrix.

[0011] Preferably, the specific generation process of the reconstructed deviation set includes: the decoding network layer inputs the higher-order node matrix into the inverse convolution kernel to perform a high-dimensional manifold unrolling operation, and outputs a manifold transition matrix; the manifold transition matrix is ​​input into the physical mapper to perform a clinical feature alignment operation using a dynamic time warping algorithm, and outputs the reconstructed node matrix; the reconstructed node matrix and the higher-order node matrix are subjected to an absolute difference conversion operation, and output a node deviation matrix; the reconstructed node matrix and the higher-order node matrix are input into the inner product derivation layer to perform a topology restoration operation, and output a comparison topology matrix; the comparison topology matrix is ​​subjected to an association break detection operation, and output a structural deviation matrix; the node deviation matrix and the structural deviation matrix are input into an adaptive penalizer to perform a joint weighted aggregation operation, and output a reconstructed deviation set.

[0012] Preferably, the specific generation process of the medical anomaly identification parameters includes: acquiring a clinical historical baseline sample set and medical equipment service life parameters; inputting the clinical historical baseline sample set into an isolated forest distribution evaluator to perform a hyperplane random cutting operation and outputting a baseline contour matrix; performing local outlier factor conversion on the reconstructed deviation set and the baseline contour matrix to output a density deviation parameter; performing aging offset conversion on the medical equipment service life parameters to output a time aging compensation parameter; inputting the time aging compensation parameter and the density deviation parameter into an adaptive comparator to perform a dynamic baseline correction operation and output a dynamic out-of-bounds parameter; performing a hard threshold binarization truncation operation on the dynamic out-of-bounds parameter and outputting the medical anomaly identification parameters.

[0013] Preferably, the specific process of executing port blocking includes: obtaining a medical IoT intranet topology library; performing routing resolution and media control addressing on the medical anomaly identifier parameter and the medical IoT intranet topology library, and outputting the physical address parameter of the target medical terminal; performing a security message encapsulation operation on the medical anomaly identifier parameter, and outputting the network blocking order; inputting the network blocking order and the physical address parameter into a bus scheduler, and outputting a directed isolation control frame; transmitting the directed isolation control frame to the target medical terminal corresponding to the medical anomaly identifier parameter; and inputting the directed isolation control frame into a preset isolation control interface to perform a local area network link stripping operation.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Through the clinical multimodal data module, electronic medical record diagnostic text sets, medical device log sets, and image lesion slice sets are acquired. These are input into a multi-source parsing model to perform cross-modal attention allocation and weighted concatenation, outputting a clinical feature matrix set. This allows the system to move beyond mechanically combining data and adjust the attention weights of multidimensional data according to the internal logic of different clinical scenarios. This processing method effectively eliminates semantic and dimensional barriers between unstructured text, time-series logs, and spatial images. The output features possess clinical representativeness and data alignment accuracy, providing a better high-quality data foundation for subsequent global quality control.

[0015] 2. By employing a graph autoencoder architecture combining feature aggregation and inverse projection, the directed graph of medical quality control is input into a feature aggregation network for neighborhood aggregation, outputting a high-order node matrix. This high-order node matrix is ​​then input into a decoding network layer for inverse projection, outputting a reconstructed node matrix. Deviation is then performed between this reconstructed and high-order node matrices, outputting a reconstructed deviation set. This allows the system to capture latent deviations in upstream and downstream collaborative data in an unsupervised manner by comparing the deviations between high-order and reconstructed nodes. This approach enables the system not only to identify known violations but also to autonomously discover unknown marginal anomalies that transcend historical experience, improving the risk detection rate in complex and cross-cutting clinical scenarios.

[0016] 3. By performing distribution fitting and identification on the reconstructed deviation set, medical anomaly identification parameters are output. A network blocking order is generated based on these parameters and transmitted to the corresponding target terminal for port blocking. This feature changes the passive situation of traditional medical quality control systems that rely solely on application-layer software alarms. It directly transforms the anomaly features discovered by the high-dimensional graph network into concrete IoT control commands. Furthermore, the medical anomaly identification parameters are routed and addressed using the medical IoT intranet topology library. The generated directional isolation control frame is then input into a preset isolation control interface to perform a local area network link stripping operation, directly cutting off the communication of the abnormal device at the data link layer. This not only constructs a robust end-to-end closed loop from intelligent anomaly early warning to hardware physical isolation, but also achieves low-latency automated safety blocking in the face of sudden medical equipment failures, enhancing the underlying risk resistance capability of the medical quality control system.

[0017] 4. The feature matrix output by the clinical multimodal data module provides a strictly aligned computational foundation for high-order node aggregation, effectively filtering out interference from single dirty data on graph topology evolution. Meanwhile, the unsupervised reconstruction bias set output by the graph network directly serves as the quantitative basis for generating directional isolation control frames. This strongly coupled architecture of multimodal perception, graph bias derivation, and underlying physical blocking ensures that each step is interdependent and errors converge step by step. The fusion accuracy of the front-end module determines the addressing accuracy of the back-end port blocking, which not only completely eliminates the fragmentation of business evaluation caused by information silos but also fundamentally avoids the erroneous disconnection of medical equipment due to fluctuations in a single sensor, achieving low false alarms, high accuracy, and fully automated quality control and physical-level security management throughout the entire process. Attached Figure Description

[0018] Figure 1 This is a structural diagram of the end-to-end quality control management system based on multi-source data fusion in medical units, as proposed in an embodiment of this invention application. Figure 2 This is a flowchart illustrating the process of generating a clinical multimodal feature matrix according to an embodiment of this invention. Figure 3 This is a flowchart illustrating the generation process of medical anomaly identifiers based on dynamic benchmark correction, as proposed in an embodiment of this invention. Detailed Implementation

[0019] 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.

[0020] Please see Figures 1-3 The present invention provides a full-process quality control management system based on multi-source data fusion in medical units, with the following specific modules: Clinical multimodal data module: acquires electronic medical record diagnosis and treatment text sets, medical device log sets, and image lesion slice sets, inputs them into a multi-source parsing model to perform cross-modal attention allocation and weighted concatenation, and outputs a set of clinical feature matrices; The medical quality control network construction module extracts the device node status parameters and medical topology association parameters contained in the clinical feature matrix set; the device node status parameters are vectorized through a node mapper to output the quality control node set; the medical topology association parameters are input into a Gaussian kernel function to perform similarity conversion, outputting a marginal weight matrix, which is then connected to the quality control node set via topological projection to output a directed medical quality control graph. The reconstructed deviation module: inputs the medical quality control directed graph into the feature aggregation network for neighborhood aggregation and outputs a high-order node matrix; inputs the high-order node matrix into the decoding network layer for inverse projection and outputs a reconstructed node matrix; performs deviation derivation with the high-order node matrix and outputs a reconstructed deviation set. Clinical anomaly identification and blocking module: Performs distribution fitting identification on the reconstructed deviation set and outputs medical anomaly identification parameters; generates a network blocking order based on the medical anomaly identification parameters and transmits it to the corresponding target terminal to perform port blocking.

[0021] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0022] Example 1 This application discloses a full-process quality control management system based on multi-source data fusion in medical institutions. (See attached document.) Figure 1 The specific modules proposed in this invention include: a clinical multimodal data module: acquiring electronic medical record diagnosis and treatment text sets, medical device log sets, and image lesion slice sets; inputting these into a multi-source analysis model to perform cross-modal attention allocation and weighted concatenation; and outputting a clinical feature matrix set; a medical quality control network construction module: extracting device node state parameters and medical topology association parameters contained in the clinical feature matrix set; performing vector entityization of the device node state parameters through a node mapper; outputting a quality control node set; inputting the medical topology association parameters into a Gaussian kernel function to perform similarity conversion; outputting a marginal weight matrix; performing topological projection connection with the quality control node set; and outputting a medical quality control directed graph; a reconstruction deviation module: inputting the medical quality control directed graph into a feature aggregation network for neighborhood aggregation; outputting a high-order node matrix; inputting the high-order node matrix into a decoding network layer for inverse projection; outputting a reconstructed node matrix; performing deviation derivation with the high-order node matrix; and outputting a reconstructed deviation set; and a clinical anomaly identification and blocking module: performing distribution fitting identification on the reconstructed deviation set; outputting medical anomaly identification parameters; generating a network blocking order based on the medical anomaly identification parameters; and transmitting it to the corresponding target terminal for port blocking.

[0023] Furthermore, the system acquires electronic medical record diagnostic text sets, medical device log sets, and image lesion slice sets, inputs them into a multi-source analysis model to perform cross-modal attention allocation and weighted concatenation, and outputs a clinical feature matrix set; corresponding to the clinical multimodal data module; see reference. Figure 2 The specific implementation process includes: The multi-source parsing model includes a medical record semantic parsing branch, a device temporal parsing branch, and an image spatial parsing branch. The electronic medical record diagnostic text set is input into the medical record semantic parsing branch to perform semantic mapping operations, outputting a medical record semantic feature matrix. The medical device log set is input into the device temporal parsing branch to perform feature aggregation operations, outputting a device vital signs temporal matrix. The medical image lesion slice set is input into the image spatial parsing branch to perform visual extraction operations, outputting an image spatial feature matrix. The medical record semantic feature matrix, the device vital signs temporal matrix, and the image spatial feature matrix are input into a cross-attention layer to perform dot product conversion, outputting a modal attention score. The modal attention score and the image spatial feature matrix are multiplied and weighted, outputting an image weighted feature set. The medical record semantic feature matrix, the device vital signs temporal matrix, and the image weighted feature set are concatenated along the channel dimension, outputting a clinical feature matrix set.

[0024] Specifically, the process of generating the clinical feature matrix set is as follows: In this embodiment, the electronic medical record diagnosis and treatment text set, medical device log set, and image lesion slice set are aligned at the underlying index level based on the same target subject and slices in the same continuous diagnosis and treatment time.

[0025] First, an electronic medical record (EMR) text set is retrieved from the medical information system database. This set contains a large amount of unstructured clinical text records, disease co-occurrence descriptions, and drug treatment association information. The EMR text set is then input into the semantic parsing branch of the multi-source parsing model to perform semantic mapping operations. In this branch, the system employs a large-scale language model (e.g., the Transformer architecture under the MedTok mechanism) that incorporates domain medical knowledge as the core text encoder. To ensure the contextual integrity of long-term longitudinal medical records and effectively avoid the loss of clinical semantics due to context truncation, the maximum context length of the input text sequence is set to 32,000 tokens. After multi-head self-attention transformation by the text encoder, each input text consultation record is ultimately projected and mapped into a 768-dimensional dense floating-point vector space, outputting a medical record semantic feature matrix.

[0026] The medical device log set acquired by the system covers continuous multi-channel time stream data generated by common devices such as electrocardiogram (ECG) monitors. Taking a 12-lead ECG monitor as an example, its generated records contain waveform rhythms lasting for 5 minutes, with sampling frequency parameters typically set to 500Hz or 1000Hz, and the signal dynamic range covering AC differential fluctuations of ±10mV and waveform resolution of 1.22 microvolts.

[0027] The system inputs the acquired medical device log set into the device time-series parsing branch to perform feature aggregation. For the time-series features of dynamic physiological signals, the core architecture of this branch employs a sequence feature extractor with a two-layer bidirectional long short-term memory network. During the feature aggregation process, the multi-layer long short-term memory network utilizes its internally constructed forgetting and input gate mechanisms to perform nonlinear recursive extraction on the passive respiratory mechanics assessment interval or ECG RR interval time series with a time step divided into 200-millisecond intervals. After multi-time-step feature superposition, the one-dimensional device log time-series signal is compressed and projected into a high-dimensional hidden-state representation space with strong time dimension dependence. Furthermore, the output dimension of this hidden-state feature is rigidly standardized to 256 dimensions, outputting the device vital signs time-series matrix.

[0028] In this embodiment, all medical image slice data follow the unified DICOM 3.0 international standard to ensure consistent physical spatial attribute definitions and pixel dimension alignment. For example, the physical specifications of the acquired tomographic or magnetic resonance imaging image matrix are often normalized and resampled to 512 pixels wide × 512 pixels high × 100 slice depths, and the DICOM header file strictly preserves the slice thickness parameters (labels: 0018, 0050), pixel pitch parameters (labels: 0028, 0030), and image position parameters based on the LPH+ (left-back-head) right-handed coordinate system.

[0029] The structured medical image lesion slice set is input into the image spatial analysis branch to perform visual extraction operations. The underlying architecture of this branch is a stacked convolutional topology structure of a deep residual network (such as the ResNet-50 architecture) with 50 convolutional computation layers. By using a preset 3×3 two-dimensional sliding convolutional kernel, combined with multiple spatial downsampling with a stride of 1 and nonlinear activation operations, the network extracts low-level grayscale texture features of lesion tissue layer by layer, while continuously expanding the image receptive field to construct its high-level spatial semantic feature representation. After a global average pooling operation at the end, the output is a dimensionality-reduced image spatial feature matrix with a unified feature dimension mapping to 256 dimensions.

[0030] The system inputs the medical record semantic feature matrix, the device vital signs time series matrix, and the image spatial feature matrix into the cross-attention layer to perform a dot product transformation operation. Specifically, an orthogonal multi-head bilinear attention network is used. This network model is configured with four parallel independent attention heads, and the projection vector dimension of each query, key, and value is set to 32 dimensions. The computational logic is as follows: using the image spatial feature matrix as the query term, the medical record semantic feature matrix and the device vital signs time series matrix are concatenated to form the key and value terms; the mathematical inner product between the query vector and the key vector is calculated and divided by the square root of the vector dimension to prevent excessive gradients; finally, the data is processed by the Softmax exponential normalization function to quantify and assign cross-modal relevance weights for different modal feature vectors within the same diagnostic time slice. After the transformation operation, the modal attention score is output.

[0031] In this embodiment, a confidence check mechanism based on range distribution is introduced. The difference between the maximum value and the average value within the attention vector is calculated as the range deviation value. When this deviation value is lower than a set threshold, the distribution is determined to be divergent, triggering a historical moving average filter for weighted correction, and outputting an updated modal attention score. After the initial modal attention score is output in the cross-attention layer, the system adds a confidence check step based on range distribution. The specific calculation steps are as follows: The system first extracts the modal attention vector output in the current time slice and calculates the arithmetic mean of all values ​​within the vector; then, it selects the maximum value in the vector. The system subtracts the arithmetic mean from the maximum value to obtain the range deviation value, which reflects the concentration of the distribution. The system configures a fixed judgment threshold in the background, usually set to 0.2. When the calculated range deviation is greater than or equal to 0.2, the system determines that the current multimodal feature alignment is clear and the confidence level meets the standard, and directly uses the initial modal attention score. When the range deviation is strictly less than 0.2, it indicates that due to the lack of data in a certain clinical modality, the attention score exhibits an disordered and evenly distributed divergent state in multiple dimensions. At this time, the system calls the preset moving average filter to perform a weighted summation of the three most recent stable attention scores and the current divergent attention score with a fixed weight of 4:1, forcibly shrinking and correcting the current excessively discrete distribution towards the historical steady state. After the above verification and weighted correction calculation, the system outputs the updated modal attention score, avoiding weight divergence caused by the lack of a single modality, and improving the anti-interference stability of the multimodal fusion process with low computational overhead.

[0032] Subsequently, a multiplication weighting operation is performed on the modal attention score and the image spatial feature matrix. The modal attention score, which is a scalar or weighted map, is then multiplied by matrix dot product with each corresponding feature channel element in the image spatial feature matrix, which is a tensor. This amplifies the spatial pixel regions in the medical image that are highly correlated with the semantic description of high-risk medical records and vital signs of equipment alarms, and outputs the image weighted feature set after target enhancement processing.

[0033] Finally, the system performs a channel-dimensional concatenation operation along the channel dimension of the data structure, combining the medical record semantic feature matrix, the device vital signs temporal matrix, and the image weighted feature set. By changing the number of channels in the feature tensor, forced alignment and integration of multi-source features within the same vector space are achieved, ultimately outputting a clinical feature matrix set with a global fusion receptive field.

[0034] By employing three dedicated branches—medical record semantics, device temporal sequence, and image spatial analysis—modal-specific features are extracted. Modal attention scores are generated using dot product transformations through cross-attention layers, and image spatial features are multiplicatively weighted to highlight key lesions and abnormal indicators. Finally, a clinical feature matrix set is generated through channel-level concatenation, effectively preventing feature dilution and information loss during heterogeneous data fusion.

[0035] Further, the device node state parameters and medical topology association parameters contained in the clinical feature matrix set are extracted; the device node state parameters are vectorized through a node mapper to output a quality control node set; the medical topology association parameters are input into a Gaussian kernel function to perform similarity conversion, outputting a marginal weight matrix, which is then topologically projected and connected with the quality control node set to output a directed medical quality control graph; this corresponds to the medical quality control network construction module; the specific implementation process includes: The clinical feature matrix set is subjected to principal component analysis and orthogonal dimensionality reduction to output device node state parameters and medical topology association parameters. The device node state parameters are input into a node mapper containing a linear projection layer and a nonlinear activation layer, and historical mean interpolation is performed on the device node state parameters to output a complete state matrix. The complete state matrix is ​​input into a denoising autoencoder to filter out background white noise and output a clean state matrix. The high-frequency component parameters corresponding to the clean state matrix are extracted and calculated through Fourier transform. The high-frequency component parameters and the clean state matrix are input into the nonlinear activation layer to perform high-dimensional manifold transformation and output a feature manifold matrix. The feature manifold matrix is ​​input into an attention pool to perform local feature compression and output the quality control node set.

[0036] Specifically, the process of generating the quality control node set is as follows: The system performs principal component analysis (PCA) orthogonal dimensionality reduction on the clinical feature matrix set. The specific logic is as follows: First, the mean of the clinical feature matrix set along the column dimensions is calculated and data-centric. Next, the eigencovariance matrix of the centered matrix is ​​calculated. Then, the eigenvalues ​​and their orthogonal eigenvectors corresponding to this covariance matrix are solved using singular value decomposition (SVD). Finally, the eigenvalues ​​are arranged in descending order, and the number of principal components required to reach a set threshold (e.g., 95%) is calculated and defined as a positive integer k. The first k principal component orthogonal eigenvectors are then projected onto the transformation matrix, and the original data is projected into this low-dimensional orthogonal space. Subsequently, the dimensionality-reduced feature vector matrix is ​​hard-segmented according to a preset feature channel dimension index: the first 50% of the feature channel dimensions are forcibly extracted, and their physical semantics are designated as device node state parameters; simultaneously, the remaining 50% of the feature channel dimensions are extracted, and their physical semantics are designated as medical topology association parameters.

[0037] Through the aforementioned dimensionality reduction, the system removes a large amount of noise and collinearity-redundant features, separating and outputting two core intermediate parameters: device node status parameters and medical topology association parameters. The former carries the inherent operating conditions of individual medical terminal devices, while the latter retains the external association clues of the device cluster in physical space and time.

[0038] The system inputs device node state parameters into a node mapper containing a linear projection layer and a nonlinear activation layer structure. In the first processing stage, the device node state parameters undergo historical mean interpolation. The system calls a dictionary of historical state mean values ​​accumulated during the past stable and fault-free operation of a specific medical device in the background. When a missing value is detected in the current state parameter input vector due to network packet loss or sensor detachment, the system uses the historical mean to smoothly fill in the missing dimension. After interpolation, the data restores multidimensional integrity and outputs a complete state matrix. Input to the linear projection layer, the system first performs a spatial organization rearrangement operation on the matrix. Specifically, the system extracts the complete state matrix, which was originally arranged in a one-dimensional feature sequence. According to the time sliding window length parameter preset by the system background (specifically set to 64 consecutive sampling steps in this embodiment) and the fixed number of device feature dimensions (specifically set to 64 dimensions), it folds and splices it into a two-dimensional feature grid tensor. If the feature dimensions of the state matrix extracted by the principal component analysis in the early stage are less than 64 dimensions, zero-padding is performed at the end of the tensor to strictly align the dimensions.

[0039] Subsequently, the complete state matrix is ​​input into the pre-trained denoising autoencoder for background white noise filtering. In this embodiment, the denoising autoencoder is configured as a lightweight symmetric encoder-decoder architecture consisting of three alternating convolutional layers and three pooling layers. Specifically, the encoder's first layer is configured with a 5×5 kernel, a padding parameter of 2, a stride of 1, and ReLU activation; the second layer is a max-pooling layer with a 2×2 kernel; and the third layer is configured with a 3×3 kernel, padding of 1, and a stride of 1. Following the third layer of the encoder, a strictly symmetric decoder structure is configured to complete feature reconstruction. The decoder sequentially includes: a first deconvolutional layer with a 3×3 kernel and padding of 1; a second non-linear upsampling interpolation layer with a 2×2 kernel to restore spatial resolution; and a third deconvolutional layer with a 5×5 kernel. Finally, a sigmoid activation function is connected to the output of the decoder to remap and stretch the high-dimensional compressed latent variable tensor back to the feature dimension that is completely consistent with the original complete state matrix, thereby achieving high-fidelity unsupervised signal reconstruction.

[0040] During the training phase, a Gaussian white noise matrix with a noise factor of 0.5 is injected into the input of the autoencoder. This 0.5 factor is derived from the measurement of the maximum background electromagnetic interference peak value of a conventional medical sensor under normal operating conditions. The autoencoder reconstructs the original, uncontaminated real state signal by forcibly constraining the decoding end, thus effectively removing jitter noise from the sensor's high-frequency electrical signal during device transmission. After filtering and reconstruction, a clean state matrix with a high signal-to-noise ratio is output (the peak signal-to-noise ratio PSNR of this matrix is ​​generally greater than 30 dB).

[0041] To further capture the dynamic abrupt changes in the operating conditions of the equipment in the frequency domain, the system calculates the high-frequency component parameters corresponding to the pure state matrix using Discrete Fourier Transform (DFT). Specifically, the system performs a DFT on the discretely acquired pure state matrix to obtain the corresponding frequency domain spectral distribution. By applying a high-pass filter threshold, it separates and intercepts the high-frequency segmented energy distributions representing drastic signal fluctuations and rate changes in the spectrum, outputting the high-frequency component parameters in the frequency domain.

[0042] In this embodiment, an adaptive high-pass cutoff frequency optimization mechanism based on the total energy percentage is introduced. The total spectral energy is integrated, and 10% of the total energy is used as the target cutoff reference value. Energy is accumulated from the highest frequency downwards, and the critical frequency reaching this reference value is set as the dynamic high-pass filter threshold. Specifically, the system first performs full-band energy integration on the complete spectral density distribution function within the current analysis period, accumulating the squared amplitude values ​​of all discrete frequency points to obtain the total spectral energy value of the current signal. Next, the system presets a high-frequency anomaly energy percentage threshold, which is strictly set to 10% in this embodiment. The system multiplies the total spectral energy value by 10% to obtain the target high-frequency energy cutoff reference value. Subsequently, starting from the highest frequency end of the spectrum, the system accumulates energy point by point downwards along the frequency decrease direction. When the accumulated energy sum is first greater than or equal to the aforementioned target high-frequency energy cutoff reference value, the system immediately stops accumulating and extracts the value of the critical frequency point at this time. The system directly sets this critical frequency point as the dynamic high-pass filter threshold for the current time slice. The system uses this dynamically adaptively calculated threshold to perform filtering. The above mechanism can adapt to the high-frequency burst physical characteristics of different devices, thus improving the detection generalization ability.

[0043] Next, the system inputs the high-frequency component parameters and the pure state matrix in parallel into the nonlinear activation layer to perform a high-dimensional manifold transformation. Using exponential linear units with a preset bias constant (e.g., 0.1), the originally linearly distributed combination parameters are mapped and projected into a high-dimensional Riemannian manifold space. After nonlinear reconstruction, a feature manifold matrix containing complex nonlinear energy expression relationships is output. Finally, the feature manifold matrix is ​​input into the attention pool to perform local feature compression. By learning the weights of the primary and secondary importance distributions of each spatial location or feature element within the feature manifold matrix in representing the overall state of the node, the high-dimensional expanded manifold matrix is ​​weighted and summed along the spatial dimension, compressing and materializing it into a standard discrete node representation in the graph network topology, ultimately outputting a quality control node set.

[0044] By using a denoising autoencoder and Fourier transform to remove white noise and extract high-frequency components, the system captures minute state fluctuations of the equipment. It employs high-dimensional manifold transformation of nonlinear activation layers and self-attention pooling for local feature compression. The resulting quality control node set not only removes redundant interference but also highly condenses potential anomalies, providing a high-fidelity, materialized feature vector for the system to identify unknown state deviations.

[0045] Extract the diagnosis and treatment time sequence occurrence stamp parameters and physical spatial location parameters from the medical topology association parameters; perform absolute difference conversion on the diagnosis and treatment time sequence occurrence stamp parameters to output the collaborative delay matrix; calculate the displacement deviation square parameter corresponding to the physical spatial location parameter; perform square root derivation on the displacement deviation square parameter to output the spatial distance matrix; perform linear weighted summation on the collaborative delay matrix and the spatial distance matrix to output the joint distance matrix; combine the joint distance matrix with natural exponential negative decay to output the spatiotemporal decay matrix; calculate the feature covariance parameter corresponding to the quality control node set and the feature standard deviation parameters of the source node and the target node; divide the feature covariance parameter by the product of the feature standard deviation parameters of the source node and the target node to output the feature correlation matrix; perform Hadamard product operation on the feature correlation matrix and the spatiotemporal decay matrix to output the marginal weight matrix; perform topological mapping and marginal connection operations on the quality control node set and the marginal weight matrix to output the medical quality control directed graph.

[0046] Specifically, the construction process of the directed graph for medical quality control is as follows: The system first performs an absolute difference calculation on the timing stamp parameters of diagnosis and treatment events, obtaining the exact network timestamps of any two medical events (the event corresponding to the source node and the event corresponding to the target node), performs an arithmetic subtraction, and takes the absolute value of the result. This absolute difference reflects the relative lag effect of different clinical medical intervention actions or equipment alarm sequences. After calculating the timing difference for all quality control node pairs in pairs, a collaborative delay matrix is ​​output.

[0047] For device deployment associations in a physical environment, the system calculates the displacement deviation square parameter corresponding to the physical spatial location parameters. In this embodiment, the spatial distance conversion follows the physical definition of three-dimensional Euclidean geometric distance. The system extracts the position coordinates of any two monitoring devices in a specific hospital ward's three-dimensional Cartesian coordinate system, calculates the coordinate differences between them in the X, Y, and Z axes, squares each coordinate difference in each dimension, sums them to obtain the displacement deviation square parameter, and then takes the square root of this sum. The resulting value depicts the linear span of the physical deployment between different devices, outputting a spatial distance matrix representing the global spatial span.

[0048] To comprehensively evaluate spatiotemporal factors, the system first performs independent interval normalization on both the time and space distance matrices. The specific transformation logic is as follows: The global maximum and minimum elements are extracted from the collaborative delay matrix and the spatial distance matrix respectively. Using the calculation rule of subtracting the minimum value from the current element and then dividing by the difference between the maximum and minimum values, all elements within the collaborative delay matrix and the spatial distance matrix are uniformly mapped to a dimensionless numerical range of 0 to 1. The system then performs a linear weighted summation on the collaborative delay matrix and the spatial distance matrix, outputting a joint distance matrix. The specific weight coefficient selection method is as follows: The standard deviations of all elements within the normalized collaborative delay matrix and the normalized spatial distance matrix are calculated separately. Then, the standard deviation of the collaborative delay matrix is ​​divided by the sum of the standard deviations of both matrices, and the result is set as the time weight coefficient. Similarly, the standard deviation of the spatial distance matrix is ​​divided by the sum of the standard deviations of both matrices, and the result is set as the spatial weight coefficient.

[0049] Based on this, the joint distance matrix is ​​combined with the natural exponential negative decay operation. Specifically, the negative value of each element in the joint distance matrix is ​​used as the exponent to solve the problem, with the natural constant as the base, so as to output the spatiotemporal decay matrix, which effectively blocks meaningless graph connections.

[0050] The system continues to calculate the feature covariance parameter corresponding to the quality control node set and the feature standard deviation parameters of the source and target nodes. The feature covariance parameter is then divided by the product of the source node's feature standard deviation parameter and the target node's feature standard deviation parameter. The standard deviation is calculated as the square root of the arithmetic mean of the sum of the squares of the deviations of the sample values ​​of each dimension of the node's features from their respective feature arithmetic means. Finally, the system outputs a dense feature correlation matrix composed of similarity scores.

[0051] Using graph structure mapping rules, the system performs a Hadamard product operation on the feature correlation matrix and the spatiotemporal decay matrix, outputting a marginal weight matrix that integrates spatiotemporal and attribute factors. Finally, the system performs topological mapping and marginal connection operations on the materialized quality control node set and the calculated marginal weight matrix. Nodes are defined as vertices in a directed graph, and marginal weights are defined as directed, weighted topological connections from the source vertex to the target vertex, outputting a directed medical quality control graph that integrates multi-source modalities and complex physical dependencies. The specific direction determination rule is as follows: the absolute direction of the marginal weight is always unidirectional, pointing from the source node with the earlier timestamp to the target node with the later timestamp; if there are concurrent medical event nodes with identical timestamps, the system calls a pre-set device network topology dictionary in the background, pointing unidirectionally from the data sending node to the data receiving node.

[0052] By extracting the diagnosis and treatment time sequence and physical spatial location, a spatiotemporal decay matrix is ​​calculated. Combined with the feature correlation matrix based on standard deviation and covariance, a marginal weight matrix is ​​generated using Hadamard product operation. This multi-dimensional deep integration of the physical spatial distance, time delay and equipment status correlation of medical equipment is achieved, and the generated medical quality control directed graph reflects the complex physical interaction topology in real-world scenarios.

[0053] Furthermore, the directed graph of medical quality control is input into a feature aggregation network for neighborhood aggregation, outputting a high-order node matrix; the high-order node matrix is ​​then input into a decoding network layer for inverse projection, outputting a reconstructed node matrix; deviation derivation is performed between the reconstructed node matrix and the high-order node matrix, outputting a reconstructed deviation set; this corresponds to the reconstructed deviation module; the specific implementation process includes: The feature aggregation network extracts the source node feature matrix and target node feature matrix from the directed graph of medical quality control; inputs the source node feature matrix and the target node feature matrix into a gating attention kernel to perform channel saliency discrimination operation, and outputs an attention weight matrix; obtains the cascade fault threshold of medical equipment; inputs the cascade fault threshold of medical equipment and the attention weight matrix into a sparse truncate unit to perform low-relevance edge filtering operation, and outputs a sparse weight matrix; performs neighborhood feature aggregation with the sparse weight matrix and the directed graph of medical quality control, and outputs an aggregated feature matrix; inputs the aggregated feature matrix into a nonlinear activation kernel with a leaky modified linear unit to perform gradient diffusion suppression, and outputs a high-order node matrix.

[0054] Specifically, the generation process of the higher-order node matrix is ​​as follows: First, the underlying feature aggregation network extracts the source node feature matrix and target node feature matrix from the directed graph of medical quality control through adjacency matrix indexing. These two feature matrices are then simultaneously input into the gating attention kernel to perform channel saliency identification. The gating attention kernel in this module draws on the gating logic principles of gated recurrent units and long short-term memory networks in complex temporal processing networks. Its core algorithm parameters include a standard convolutional mapping layer with 512 input / output channels and a 3×3 core size, and a 32-dimensional query key-value distribution space distributed across four independent parallel attention heads. This is achieved through the combined use of the Sigmoid activation function and the Tanh function. Simultaneously, relying on the similarity calculation of the attention heads, alignment weights between features are calculated to automatically focus the attention on channels with high clinical saliency responses. After completing the above identification, an attention weight matrix with attention distribution intensity is output.

[0055] To eliminate high-noise, sporadic interference edges lacking clear clinical causal relationships, the system pre-determines a cascading fault threshold for medical devices. In a real hospital information network with strong coupling, the propagation of cascading faults between individual device nodes and the resulting redundant alarm signals has a clear critical safety assessment threshold. If some false alarms or sporadic weak connections fall below this threshold, the risk of causing a global network collapse or substantial clinical decision-making bias approaches zero. This cascading fault threshold is determined by statistically analyzing the weight distribution of false alarm signals in the medical unit over the past year, using its 95th percentile as a fixed threshold; in this embodiment, it is specifically set to 0.15.

[0056] Therefore, the system synchronously inputs the cascade fault threshold of medical devices and the attention weight matrix into the sparse truncate to perform low-relevance edge filtering. If any edge connection weight generated by attention evaluation is lower than the cascade fault threshold in numerical comparison, its weight will be forcibly truncated to zero through hard constraints. This reduces the computational complexity of the dense graph, improves the robustness and noise resistance of the network backbone structure, and outputs the sparse weight matrix after network sparsification.

[0057] Subsequently, the sparse weight matrix is ​​subjected to neighborhood feature aggregation with the directed graph of medical quality control. Using standard graph convolution information propagation rules, the target node actively receives and integrates the feature states of its first-order and multi-order related neighbor nodes; the weighting coefficients constitute the sparse weight matrix. To effectively prevent excessive smoothing attenuation of node features after multiple information transmissions in deep graph networks, this embodiment strictly limits the depth of graph convolution information propagation to two hop orders (i.e., only two layers of graph convolutional mapping network). After information aggregation, each node, in addition to containing its own original attributes, also condenses the adjacency state pattern of its local network community, outputting an aggregated feature matrix.

[0058] The aggregated feature matrix is ​​input and gradient vanishing suppression is performed using a nonlinear activation kernel with leaked modified linear units. When processing matrix elements with negative input values, the activation function does not directly set them to 0, but retains them and assigns a preset, extremely small constant (such as 0.01) as its gradient descent slope. After the suppression operation is completed, a higher-order node matrix is ​​output.

[0059] By combining a cascade fault threshold for medical devices and filtering out low-relevance edges, invalid connections and computational noise in the network topology are reduced. Subsequently, neighborhood feature aggregation is performed, and modified linear units are used to suppress gradient vanishing in deep networks. This enables the system to focus on high-risk device cascade paths and efficiently extract high-order node matrices with deep topological associations, which is beneficial for stably mining key anomaly propagation features in complex data.

[0060] The decoding network layer inputs the higher-order node matrix into an inverse convolutional kernel to perform a high-dimensional manifold unrolling operation, outputting a manifold transition matrix; the manifold transition matrix is ​​input into a physical mapper to perform a clinical feature alignment operation using a dynamic time warping algorithm, outputting the reconstructed node matrix; the reconstructed node matrix and the higher-order node matrix are subjected to an absolute difference conversion operation, outputting a node deviation matrix; the reconstructed node matrix and the higher-order node matrix are input into an inner product derivation layer to perform a topology restoration operation, outputting a comparison topology matrix; the comparison topology matrix is ​​subjected to an association break detection operation, outputting a structural deviation matrix; the node deviation matrix and the structural deviation matrix are input into an adaptive penalizer to perform a joint weighted aggregation operation, outputting a reconstruction deviation set.

[0061] Specifically, the process of reconstructing the deviation set is as follows: The decoding network layer inputs the high-order node matrix into the inverse convolution kernel to perform a high-dimensional manifold unfolding operation. By spreading zero-value padding around the elements of the original feature matrix and superimposing learnable convolution kernel sliding operations, the high-order node matrix is ​​upsampled and expanded in the spatial resolution or temporal length dimension. It is then gradually nonlinearly mapped and unfolded back to the data scale of its input original observation space, and outputs the manifold transition matrix.

[0062] Due to the prevalent temporal misalignment problem among multimodal features, the system inputs the manifold transition matrix into the physical mapper and employs a dynamic time warping algorithm to perform clinical feature alignment. Specifically, for each feature time series, a two-dimensional cumulative alignment cost matrix is ​​constructed. The value of any grid coordinate position in this matrix is ​​determined by calculating the square of the feature data difference between the current source node sequence time point and the corresponding time point of the target node sequence (i.e., the basic metric of Euclidean distance), and adding the minimum value among the three accumulated path cost elements of that coordinate position in the cost matrix: its left immediate neighbor, its upper immediate neighbor, and its upper left diagonal immediate neighbor. Through iterative accumulation using the dynamic programming paradigm, the system optimizes from the starting grid to the diagonal ending grid to obtain the minimum nonlinear cumulative path cost between the two feature time series. Furthermore, the system applies Sakoe-Chiba band constraints in its computational constraints and strictly sets the search width threshold of the maximum warping window to 10% of the total length of the input feature sequences. Finally, the reconstructed node matrix is ​​output.

[0063] First, for individual attribute offsets, the system performs an absolute difference conversion operation on the reconstructed node matrix and the higher-order node matrix together. It iterates through and calculates the absolute mathematical difference between the vector value of each medical device node in the decoding mapping reconstruction space and the corresponding dimension vector element in the original higher-order observation space, and outputs a scalarized node bias matrix.

[0064] The reconstructed node matrix and the higher-order node matrix are then input into the inner product derivation layer to perform topology restoration. The inner product derivation layer calculates the dot product similarity between the feature vectors of any two nodes within these two matrix systems using matrix transpose and multiplication. A larger dot product value indicates a more similar feature distribution and a higher probability of hidden physical connections. Through this logical projection from features to structure, the system reconstructs the implicit adjacency graph between nodes based on the similarity score and outputs it as a comparison topology matrix.

[0065] Next, the system performs a link break detection operation on the generated comparison topology matrix, comparing row by row the original network physical adjacency graph established by edge weights with the structural reconstruction adjacency graph derived by decoding features. When it detects that some historically existing high-frequency normal edges have significant drops or absences in their edge weights in the reconstruction graph, or when it finds that erroneous topology clustering has been added between device nodes that were not originally physically related due to violent oscillations at the same frequency, these topology variations that cause structural damage are marked as high-risk, and a structural deviation matrix is ​​output.

[0066] Finally, the system synchronously inputs the node deviation matrix and the structural deviation matrix into the adaptive penalizer to perform a joint weighted aggregation operation. This adaptive penalizer introduces an adaptive penalty function strategy with dynamic constraint capabilities into its underlying architecture. It uses Lagrange multiplier parameters as dynamic weight adjustment anchors, adjusting the combined weight ratio of individual node attribute deviations and global network topology deviations in the final cost calculation formula based on the size and severity of irregular node feature groups violating constraints in each iteration cycle. The two types of penalty scores are aggregated to generate a reconstruction deviation set for qualitative judgment. The specific mapping relationship for adjusting the weight ratio is as follows: the system calculates the ratio of the number of broken edges in the comparison topology matrix to the total number of edges in the initial medical quality control directed graph, defining it as the topology destruction rate; it extracts a preset structural deviation base weight (e.g., 0.3), multiplies the topology destruction rate by a penalty amplification factor (e.g., 1.5), adds it to the base weight, and outputs the dynamic structural weight, with the maximum threshold of this dynamic structural weight rigidly limited to 0.8; subsequently, a fixed value of 1 is subtracted from the dynamic structural weight to output the corresponding node attribute deviation weight. This explicit numerical mapping ensures that when a local anomaly causes a large-scale network topology collapse, the system can autonomously shift the focus of punishment towards structural deviations.

[0067] By using high-dimensional manifold unrolling with inverse convolutional kernels and dynamic time warping algorithms, not only is the node deviation matrix of a single device calculated, but the structural deviation matrix is ​​also output synchronously through inner product derivation layer and associated broken link detection. This process can comprehensively quantify the variation of node attributes and the collapse of system topology connections, thereby identifying various unknown deep-seated abnormal states.

[0068] Furthermore, the reconstructed deviation set is subjected to distribution fitting and identification, and medical anomaly identification parameters are output; a network blocking order is generated based on the medical anomaly identification parameters and transmitted to the corresponding target terminal for port blocking; corresponding to the clinical anomaly identification blocking module; parameter reading Figure 3 The specific implementation process includes: Acquire a clinical historical baseline sample set and medical equipment service life parameters; input the clinical historical baseline sample set into an isolated forest distribution evaluator to perform a hyperplane random cutting operation, and output a baseline contour matrix; perform local outlier factor conversion on the reconstructed deviation set and the baseline contour matrix, and output density deviation parameters; perform aging offset conversion on the medical equipment service life parameters, and output time aging compensation parameters; input the time aging compensation parameters and the density deviation parameters into an adaptive comparator to perform a dynamic baseline correction operation, and output dynamic out-of-bounds parameters; perform hard threshold binarization truncation on the dynamic out-of-bounds parameters, and output medical anomaly identification parameters.

[0069] Specifically, the generation process of medical anomaly marker parameters is as follows: The system first retrieves the clinical historical baseline sample set and the service life parameters of medical equipment directly linked to all controlled devices from the central archive database. The clinical historical baseline sample set details the historical graph reconstruction deviation distribution data reflecting normal system fluctuations generated during the target medical unit's normal observation period when the equipment was confirmed to be fault-free.

[0070] The system inputs the extracted clinical history baseline sample set into an Isolation Forest distribution estimator based on an unsupervised learning paradigm to perform hyperplane random cutting operations. The core ensemble tree estimator of the Isolation Forest distribution estimator has a core parameter of 256, and the threshold for the maximum random feature sampling number for each tree is also set to 256 sampling points. By performing multiple random hyperplane splits at feature extrema in the high-dimensional feature space, the average depth level of the decision tree path required to segment the sample data of a specific node is automatically calculated. Based on this average depth distance, an outer boundary barrier model for judging the massive amount of normal data is constructed, thereby outputting a baseline contour matrix representing the extreme values ​​of the healthy historical operating range.

[0071] Subsequently, the reconstructed deviation set and the baseline profile matrix are jointly input, and a local outlier factor conversion step is performed. During the conversion, the detection and evaluation range is first defined, and the calculated local average data spatial density around a designated evaluation point is compared with the overall average local density of several neighboring nodes within its designated neighborhood. In the implementation parameter optimization configuration of this embodiment, the neighborhood point investigation scale parameter is set to 20, while the upper limit of the expected empirical outlier contamination rate for medical network faults is set to 0.1 (i.e., 10%).

[0072] If the calculated density of a device reconstruction deviation vector point in the set hyperspace dimension is drastically lower than the density of the nearest healthy point cloud defined in the historical baseline contour matrix, then this test sample will be assigned a spatial outlier score with a relatively high value in this local outlier factor conversion operation. After traversing the reconstruction deviation set, the output is a density deviation parameter that directly quantifies the relative fault severity index of each monitoring device.

[0073] Given that physical medical devices undergo wear and tear on their components and natural aging of precision, to prevent reasonable numerical deviations caused by natural physical aging from being incorrectly reconstructed and misjudged by the intelligent system, the system extracts the time marked on the label of the tested device, calls up the service life parameters of the medical device, and independently performs aging offset conversion. The specific calculation logic is as follows: a fixed constant base is extracted as the basic reaction enhancement coefficient factor representing the rate of chemical reaction during physical aging (this coefficient is usually forced to a value of 2 under the general model parameter configuration). The calculation logic of its exponential part is defined as the difference between the actual measured chip ambient temperature of the medical device currently operating under high load or high heat environment and the normal reference room temperature of the standard cleanroom, and then divided by 10 to obtain the specific power-order parameter exponent. Finally, the system outputs the time aging compensation parameter through derivation and calculation. The normal reference room temperature of the standard cleanroom is a constant environmental parameter preset in the system's background configuration file, and its value can be fixedly extracted according to the national medical building environment specifications (e.g., fixed value of 25 degrees Celsius).

[0074] In this embodiment, a truncated mean filtering mechanism based on a sliding time window is introduced. A buffer pool with a length of 10 sampling periods is constructed. The collected temperature values ​​are fully permuted, and after removing the extreme values ​​at both ends, the arithmetic mean of the middle values ​​is calculated to output a smoothed center temperature. Specifically, when processing the raw temperature readings directly collected from the physical device, in order to eliminate transient thermal shock interference, the system adds a truncated mean filtering step before substituting the exponential difference calculation. The specific calculation process is as follows: the system no longer directly uses the instantaneous temperature reading at a single moment, but establishes a buffer data pool with a length of 10 sampling periods, continuously collecting and storing the chip ambient temperature values ​​of the most recent 10 measurements. When the buffer data pool is full, the system first fully permutes these 10 temperature values ​​in descending order. Then, the system performs an extreme value removal operation, fixing and removing the two highest temperature values ​​at the front and the two lowest temperature values ​​at the back, thereby forcibly filtering out non-physical temperature spurious peaks caused by instantaneous short circuits of the sensor or accidental external thermal shocks. Subsequently, the system calculates the arithmetic mean of the remaining six temperature values ​​in the middle range, sums all six values, and divides by 6 to obtain a smooth and stable center temperature estimate. The system strictly uses this optimal smoothed temperature value obtained after truncated mean calculation as the final, truly compliant measured ambient temperature value of the chip. The above mechanism filters out transient extreme spurious peaks with extremely simple sorting and truncation logic, ensuring the purity of the aging algorithm and temperature difference calculation from the underlying data source.

[0075] The system inputs the time aging compensation parameters and density deviation sub-parameters in parallel to the adaptive comparator and performs dynamic benchmark correction operations. Utilizing the reasonable tolerance margin variable values ​​at the data points corresponding to the time aging compensation parameter matrix introduced by the system, it performs reverse subtraction on extreme value fluctuations in the density deviation sub-parameter results, outputting dynamic out-of-bounds parameters.

[0076] The system performs hard threshold binarization truncation on dynamic out-of-bounds parameters. The truncation logic is as follows: the system sets an absolute system security isolation value red line in the background configuration file. When the value of the detected and calculated dynamic out-of-bounds parameter is indeed higher than the threshold red line after comparison, the operator directly forces a "1" output downstream (that is, the system confirms that the corresponding device port has a serious anomaly and has the nature of threat propagation); conversely, when the dynamic out-of-bounds parameter is calculated and its result value is lower than or exactly equal to the set threshold red line, a system Boolean logic instruction "0" representing safety will be output (that is, the system recognizes that the fluctuation at this time is in a safe and controllable state and marks it as normal release).

[0077] The objective acquisition process of the system safety isolation numerical red line is as follows: Extract all dynamic out-of-bounds parameters generated by the target medical unit during a continuous historical normal operation cycle confirmed to be fault-free, forming a historical out-of-bounds parameter sample set; calculate the arithmetic mean and standard deviation of this historical out-of-bounds parameter sample set; based on statistical principles, add the arithmetic mean to three times the standard deviation to output the calculation boundary baseline; further, call the objective database of the National Medical Device Classification Catalog to extract the risk level quantification factor corresponding to the target medical terminal (for example, for Class III medical devices, the objective value of the risk level quantification factor is 0.8; for Class II medical devices for routine physiological monitoring, the value is 1.0); multiply the calculation boundary baseline by the risk level quantification factor, and finally dynamically derive and output the system safety isolation numerical red line.

[0078] Then, all abnormal terminals with the trigger action attribute "Logic 1" are aggregated into medical abnormality identifier parameters that have the characteristics of a set of instructions that determine the underlying cut-off logic semantics.

[0079] By generating a baseline profile through isolated forests, calculating density deviation parameters using local outlier factors, and introducing medical equipment service life parameters to calculate time aging compensation parameters, the deviation of normal operation indicators caused by natural aging of equipment is fully considered, effectively avoiding false alarms in business operations, and helping the anomaly identification parameters output by the final hard threshold truncation to reflect unexpected abnormal states.

[0080] Obtain the medical IoT intranet topology library; perform routing resolution and media control addressing between the medical anomaly identification parameters and the medical IoT intranet topology library, and output the physical address parameters of the target medical terminal; perform a secure message encapsulation operation on the medical anomaly identification parameters, and output the network blocking order; input the network blocking order and the physical address parameters into the bus scheduler, and output a directed isolation control frame; transmit the directed isolation control frame to the target medical terminal corresponding to the medical anomaly identification parameters; input the directed isolation control frame into a preset isolation control interface to perform a local area network link stripping operation.

[0081] Specifically, when the medical anomaly flag parameter outputs a Boolean value of 1, the system triggers a physical port blocking mechanism. The specific execution flow is as follows: First, the system calls the medical IoT intranet topology library and uses the association mapping kernel to perform reverse route tracing and media access control addressing. The system compares the virtual tags in the anomaly identification parameters with the topology library to extract the physical address parameters (such as firmware MAC address or static IP) of the target medical terminal of the faulty device.

[0082] The system uses the AES-256 symmetric encryption algorithm to encapsulate the abnormal identification parameters into secure messages and outputs a network blocking order with dual security protection.

[0083] Subsequently, the blocking order, along with the physical address parameters, is input to the bus scheduler. The scheduler, based on the underlying communication protocol, translates and reconstructs it into a targeted addressing isolation control frame. This control frame is configured with 8 core data bits and 2 stop check bits, and, with the help of the underlying congestion prevention and data acknowledgment retransmission mechanisms, delivers the instruction to the target terminal.

[0084] Finally, the targeted isolation control frame is sent to the target device's pre-configured isolation control interface via the local area network. Based on the isolation command, the device invokes its local network interface management logic to disable communication, disconnect the session, or isolate the port. At this point, the abnormal terminal is isolated from the network.

[0085] By matching the physical address of the target terminal to the medical IoT intranet topology library, abnormal parameters are encapsulated into security messages and targeted isolation control frames. The instructions are then directly sent to the pre-configured isolation control interface to perform local area network link stripping. This media control and physical network isolation operation, which bypasses upper-layer application software and acts directly on the hardware layer, achieves millisecond-level network blocking for abnormal medical devices, improving the efficiency of quality control execution.

[0086] The joint training process of the parameters of the multi-source analytical model and the feature aggregation network in this system is as follows: First, multi-source heterogeneous data generated by the target medical unit during a continuous historical normal operation cycle with no confirmed faults are extracted as benchmark training samples, and end-to-end training is carried out using an unsupervised self-reconstruction learning paradigm.

[0087] The core training objective of the system is to minimize the feature reconstruction loss. The specific calculation method of the feature reconstruction loss is as follows: calculate the sum of the squared differences of the tensor elements at all corresponding positions between the high-order node matrix output by the feature aggregation network and the reconstructed node matrix output by the decoding network layer, and then calculate the total arithmetic mean of the sum of squares (i.e., a purely textual derivation of the mean squared error algorithm), which is used as the sole scalar penalty basis for backpropagation.

[0088] An adaptive moment estimation optimizer is used to perform gradient descent and dynamic updates of the network weights. The initial learning rate parameter is set to 0.001, and a cosine annealing decay strategy is introduced during training to prevent the model from getting trapped in local optima.

[0089] The system backend is set with a dynamic early stop mechanism. When the value of the feature reconstruction loss decreases by less than 0.0001 within 10 consecutive complete training iterations, or when the global maximum preset number of iterations (such as 500 times) is reached, the system automatically determines that the model network parameters have converged to the optimal solution, forcibly terminates the training process and fixes the network weights.

[0090] This invention provides a full-process quality control management system based on multi-source data fusion in medical institutions. It extracts multimodal features from electronic medical record diagnostic texts, medical equipment operation logs, and medical image slices using a multi-source analysis model. Combining Gaussian kernel functions and topological projection, it constructs a directed graph for medical quality control, achieving deep, structured mapping of massive amounts of heterogeneous medical data. Secondly, employing a graph feature aggregation network and a decoding network inverse projection mechanism, by calculating the feature deviation set between the higher-order node matrix and the reconstructed node matrix, it can utilize the structural reconstruction error of the underlying topology to identify hidden unknown abnormal state deviations in massive data. Finally, based on the generated medical anomaly identification parameters, it directly generates network blocking orders and transmits them to the target terminal to execute underlying hardware port blocking. This achieves automated interception of commands directly reaching the physical port, reducing the execution lag problem of the quality control system and facilitating safe response in medical scenarios.

[0091] Example 2 This embodiment describes the implementation steps of applying a full-process quality control management system based on multi-source data fusion in medical units to medical equipment (such as standard multi-parameter monitors and automated infusion pumps) in general medical wards for collaborative network monitoring.

[0092] First, the system acquires a set of electronic medical record diagnostic texts and inputs them into the medical record semantic parsing branch of the multi-source parsing model. A pre-trained language model is used to perform semantic mapping, outputting a 768-dimensional medical record semantic feature matrix. Simultaneously, it acquires a set of medical device logs from bedside routine monitors and infusion pumps, inputting them into the device temporal parsing branch for feature aggregation, outputting a device vital signs temporal matrix. Then, it acquires a set of standard-format medical image slices, extracts pixel spacing and slice thickness parameters, and normalizes them into a 512×512×100 pixel matrix, inputting it into the image space parsing branch to output an image space feature matrix. These three matrices are then input into an orthogonal multi-head bilinear attention network for dot product transformation, outputting modal attention scores. After multiplicative weighting and concatenation with channel dimensions, the final output is a dimensionality-reduced and aligned set of clinical feature matrices.

[0093] Secondly, principal component analysis (PCA) is performed on the clinical feature matrix set for orthogonal dimensionality reduction, with the cumulative variance contribution rate threshold set to 95%. This outputs the device node state parameters and the medical topology correlation parameters. After historical mean interpolation, the device node state parameters are input to a node mapper and then fed into a denoising autoencoder for background white noise filtering, outputting a clean state matrix. High-frequency component parameters are extracted via discrete Fourier transform and subjected to high-dimensional manifold transformation, followed by pooling to output the quality control node set. Subsequently, the occurrence stamps and physical coordinates in the medical topology correlation parameters are derived using absolute difference and Euclidean geometric distance, combined with natural exponential decay to output a spatiotemporal decay matrix. This matrix is ​​then multiplied element-wise with the feature correlation matrix calculated based on the Pearson correlation coefficient to output a marginal weight matrix. This is then concatenated with the quality control node set to output a directed medical quality control graph.

[0094] Next, the directed graph for medical quality control is input into a gated attention kernel. Nonlinear activation functions such as hyperbolic tangent control information retention and forgetting to perform channel saliency identification, outputting an attention weight matrix. After sparse truncation and neighborhood aggregation using a cascaded fault threshold for medical devices, a higher-order node matrix is ​​output. This matrix is ​​then inversely convolved and expanded into a manifold transition matrix. Subsequently, the system calls a dynamic time warping algorithm to perform clinical feature alignment, outputting a reconstructed node matrix. This reconstructed node matrix is ​​then converted to an absolute difference and topologically restored using an inner product with the higher-order node matrix, generating a node bias matrix and a structural bias matrix, respectively. These two matrices are dynamically weighted by an adaptive penalizer to output a reconstructed bias set.

[0095] Finally, the system inputs the acquired clinical historical baseline sample set into the isolated forest distribution evaluator, performs a hyperplane random cutting operation, and outputs a baseline contour matrix. The reconstruction deviation set and this matrix are then jointly input into the local outlier factor algorithm to perform density outlier conversion. The algorithm's neighborhood point parameter is set to 20, and the upper limit of the abnormal contamination rate is set to 0.1, outputting density deviation parameters. To eliminate normal physical wear and tear from long-term high-load operation of ordinary medical equipment, the system introduces the accelerated aging test conversion standard for sterile medical device packaging and substitutes it into the medical equipment's service life parameters. Under parameters where the baseline response enhancement factor is 2, the normal baseline room temperature is set to 25 degrees Celsius, and the measured high-load temperature is 55 degrees Celsius, the system converts and outputs time aging compensation parameters. After the two sets of parameters are corrected by an adaptive comparator and subjected to hard threshold truncation, a medical anomaly identifier parameter with a Boolean value of 1 is output due to the detection of a reconstruction collapse at a certain infusion pump node. The system then extracts the physical address parameters of the abnormal target terminal from the internal network topology library, encapsulates them in a 256-bit symmetric encrypted secure message, and outputs a network blocking order. The bus scheduler translates this instruction into a directed isolation control frame that notifies the underlying control to disconnect. This data frame requires the receiving network card to transmit the data to the target medical terminal under underlying fault-tolerant timing constraints, thus performing a hard disconnection and blocking of the local area network communication link of the abnormal device.

[0096] 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 full-process quality control management system based on multi-source data fusion in medical institutions, characterized in that: include: Clinical multimodal data module: acquires electronic medical record diagnosis and treatment text sets, medical device log sets, and image lesion slice sets, inputs them into a multi-source parsing model to perform cross-modal attention allocation and weighted concatenation, and outputs a set of clinical feature matrices; The medical quality control network construction module extracts the device node status from the clinical feature matrix set to participate in the medical topology association parameters; the device node status parameters are vectorized through a node mapper to output the quality control node set; the medical topology association parameters are input into a Gaussian kernel function to perform similarity conversion, outputting a marginal weight matrix, which is then topologically projected and connected with the quality control node set to output a directed medical quality control graph; The reconstruction deviation module: It inputs the directed graph of medical quality control into a feature aggregation network for neighborhood aggregation, outputting a high-order node matrix; it then inputs the high-order node matrix into a decoding network layer for inverse projection, outputting a reconstructed node matrix; finally, it performs deviation derivation between the reconstructed node matrix and the high-order node matrix, outputting a reconstruction deviation set. The specific generation process of the reconstruction deviation set includes: the decoding network layer inputs the high-order node matrix into an inverse convolution kernel to perform a high-dimensional manifold unfolding operation, outputting a manifold transition matrix; the manifold transition matrix is ​​input into a physical mapper and a dynamic time warping algorithm is used to perform clinical feature alignment, outputting the reconstructed node matrix; the... The reconstructed node matrix and the higher-order node matrix are subjected to an absolute difference conversion operation to output a node deviation matrix; the reconstructed node matrix and the higher-order node matrix are input into an inner product derivation layer to perform a topology restoration operation to output a comparison topology matrix; the comparison topology matrix is ​​subjected to an association break detection operation to output a structure deviation matrix; the node deviation matrix and the structure deviation matrix are input into an adaptive penalizer to perform a joint weighted aggregation operation to output a reconstruction deviation set; wherein the feature aggregation network and the decoding network layer adopt an unsupervised self-reconstruction learning paradigm for end-to-end training, and the training objective is to minimize the feature reconstruction loss; The clinical anomaly identification and blocking module performs distribution fitting identification on the reconstructed deviation set and outputs medical anomaly identification parameters. Specifically, it performs aging conversion on the service life parameters of medical equipment, using the basic improvement coefficient of physical aging reaction as the base and the difference between the measured ambient temperature and the normal baseline room temperature as the exponent to derive time aging compensation parameters; it uses the tolerance margin of the time aging compensation parameters to perform reverse subtraction on the fluctuation extreme values ​​of the density deviation sub-parameters, outputting dynamic out-of-bounds parameters; it performs hard threshold binarization truncation on the dynamic out-of-bounds parameters, outputting the medical anomaly identification parameters; it generates a network blocking order based on the medical anomaly identification parameters and transmits it to the corresponding target terminal to execute port blocking; the port blocking includes: calling the internal network topology library to perform reverse routing tracing and media access control addressing to extract the physical address parameters of the target terminal; it uses a symmetric encryption algorithm to encapsulate the medical anomaly identification parameters and outputs the network blocking order; it inputs the network blocking order and physical address parameters into the bus scheduler to translate them into a directed isolation control frame, and performs a local area network link stripping operation through a preset isolation control interface.

2. The end-to-end quality control management system based on multi-source data fusion in medical units according to claim 1, characterized in that, The specific generation process of the clinical feature matrix set includes: the multi-source parsing model includes a medical record semantic parsing branch, a device temporal parsing branch, and an image spatial parsing branch; the electronic medical record diagnosis and treatment text set is input into the medical record semantic parsing branch to perform semantic mapping operation, and outputs the medical record semantic feature matrix; the medical device log set is input into the device temporal parsing branch to perform feature aggregation operation, and outputs the device vital signs temporal matrix; the medical image lesion slice set is input into the image spatial parsing branch to perform visual extraction operation, and outputs the image spatial feature matrix; the medical record semantic feature matrix, the device vital signs temporal matrix, and the image spatial feature matrix are input into the cross-attention layer to perform dot product conversion, and output modal attention score; the modal attention score and the image spatial feature matrix are multiplied and weighted to output the image weighted feature set; the medical record semantic feature matrix, the device vital signs temporal matrix, and the image weighted feature set are concatenated along the channel dimension to output the clinical feature matrix set.

3. The end-to-end quality control management system based on multi-source data fusion in medical units according to claim 1, characterized in that, The specific generation process of the quality control node set includes: performing principal component analysis orthogonal dimensionality reduction on the clinical feature matrix set to output device node state parameters for medical topology correlation; inputting the device node state parameters into a node mapper containing a linear projection layer and a nonlinear activation layer, performing historical mean interpolation on the device node state parameters, and outputting a complete state matrix; inputting the complete state matrix into a denoising autoencoder to perform background white noise filtering, and outputting a clean state matrix; extracting and calculating the high-frequency component parameters corresponding to the clean state matrix through Fourier transform; inputting the high-frequency component parameters into the clean state matrix to perform high-dimensional manifold transformation on the nonlinear activation layer, and outputting a feature manifold matrix; inputting the feature manifold matrix into an attention pool to perform local feature compression, and outputting the quality control node set.

4. The end-to-end quality control management system based on multi-source data fusion in medical units according to claim 1, characterized in that, The specific construction process of the directed graph for medical quality control includes: extracting the diagnosis and treatment time sequence occurrence stamps and physical spatial location parameters contained in the medical topology correlation parameters; performing absolute difference conversion on the diagnosis and treatment time sequence occurrence stamp parameters to output a collaborative delay matrix; calculating the displacement deviation square parameter corresponding to the physical spatial location parameter; performing square root derivation on the displacement deviation square parameter to output a spatial distance matrix; performing linear weighted summation on the collaborative delay matrix and the spatial distance matrix to output a joint distance matrix; combining the joint distance matrix with the natural exponential negative decay to output a spatiotemporal decay matrix; calculating the feature covariance parameters corresponding to the quality control node set and the respective feature standard deviation parameters of the source node and the target node; dividing the feature covariance parameter by the product of the feature standard deviation of the source node and the feature standard deviation of the target node to output a feature correlation matrix; performing a Hadamard product operation on the feature correlation matrix and the spatiotemporal decay matrix to output a marginal weight matrix; performing topological mapping and marginal connection operations on the quality control node set and the marginal weight matrix to output a directed graph for medical quality control.

5. The end-to-end quality control management system based on multi-source data fusion in medical units according to claim 1, characterized in that, The specific generation process of the higher-order node matrix includes: a feature aggregation network extracts the source node feature matrix and the target node feature matrix contained in the directed graph of medical quality control; the source node feature matrix and the target node feature matrix are input into a gated attention kernel to perform channel saliency discrimination operation, and the attention weight matrix is ​​output; the cascade fault threshold of medical equipment is obtained; the cascade fault threshold of medical equipment and the attention weight matrix are input into a sparse truncate unit to perform low-relevance edge filtering operation, and the sparse weight matrix is ​​output; the sparse weight matrix and the directed graph of medical quality control are subjected to neighborhood feature aggregation, and the aggregated feature matrix is ​​output; the aggregated feature matrix is ​​input into a nonlinear activation kernel with a leaky modified linear unit to perform gradient diffusion suppression, and the higher-order node matrix is ​​output.

6. The end-to-end quality control management system based on multi-source data fusion in medical units according to claim 1, characterized in that, The specific generation process of the medical anomaly identifier parameters includes: acquiring a clinical historical baseline sample set and medical equipment service life parameters; inputting the clinical historical baseline sample set into an isolated forest distribution evaluator to perform a hyperplane random cutting operation and outputting a baseline contour matrix; performing local outlier factor conversion on the reconstructed deviation set and the baseline contour matrix to output density deviation parameters; performing aging offset conversion on the medical equipment service life parameters to output time aging compensation parameters; inputting the time aging compensation parameters and the density deviation parameters into an adaptive comparator to perform a dynamic baseline correction operation and outputting dynamic out-of-bounds parameters; performing hard thresholding binarization truncation on the dynamic out-of-bounds parameters and outputting medical anomaly identifier parameters.

7. The end-to-end quality control management system based on multi-source data fusion in medical units according to claim 1, characterized in that, The specific process of executing port blocking includes: obtaining the medical IoT intranet topology library; performing routing resolution and media control addressing on the medical anomaly identifier in the medical IoT intranet topology library, and outputting the physical address parameter of the target medical terminal; performing a security message encapsulation operation on the medical anomaly identifier parameter, and outputting the network blocking order; inputting the network blocking order and the physical address parameter into the bus scheduler, and outputting a directed isolation control frame; transmitting the directed isolation control frame to the target medical terminal corresponding to the medical anomaly identifier parameter; and inputting the directed isolation control frame into a preset isolation control interface to perform a local area network link stripping operation.

Images