A nursing quality control intelligent verification and scoring method, device, equipment and medium

Abstract

The application relates to a nursing quality control intelligent verification scoring method, device, equipment and medium. The method comprises the following steps: acquiring patient electronic medical record data based on bed identification information and performing entity relationship joint extraction to obtain a patient multi-dimensional pathological feature vector set; a static quality control rule tree is adaptively reconstructed into a graph to generate a customized quality control verification graph carrying adaptive weight coefficients; multi-modal verification data flow is acquired based on the customized quality control verification graph and cross-modal feature alignment is performed to obtain a multi-modal feature matrix; semantic conflict detection is performed on the multi-modal feature matrix to obtain verification node data that passes the check; and the verification node data is weighted and aggregated based on the adaptive weight coefficients to generate a global quality control score result. The method can dynamically adjust the verification weight based on the pathological features and exclude multi-modal semantic conflicts, thereby improving the objectivity of the quality control score.

Description

Technical Field

[0001] This invention belongs to the field of intelligent quality control technology, and in particular relates to a nursing quality control intelligent verification and scoring method, device, equipment and medium. Background Technology

[0002] With the development of medical informatization and smart ward technology, digital technologies such as electronic medical record systems, mobile nursing terminals, and medical image acquisition equipment have emerged. These technologies are characterized by the realization of electronic storage of nursing process data, multi-source heterogeneous acquisition, and convenient terminal access, which leads to the gradual shift of current nursing quality control and verification scoring methods from purely manual paper records to electronic-assisted verification.

[0003] In traditional techniques, nursing quality control scoring methods typically work as follows: Inspectors use mobile devices to scan patient bed identification codes to retrieve pre-configured electronic quality control forms. They then manually check each item on the form and check its status in the system. For any anomalies, inspectors manually input text descriptions or attach photos taken on-site as supporting evidence. After receiving all the checked results and supporting materials, the system simply adds up the results according to a pre-set fixed score ratio, ultimately generating a quality control scoring report.

[0004] However, current nursing quality control scoring methods have the following problems: On the one hand, existing scoring methods use completely fixed inspection rules and uniform score allocation standards, which cannot flexibly adjust the focus and importance weight of each quality control item according to the dynamic and personalized pathological characteristics of different patients, resulting in the quality control standards being out of touch with the actual condition of patients; on the other hand, during the verification process, the system only performs simple archiving and storage display of multimodal data such as manually entered text descriptions and on-site photographed images, lacking an automatic verification mechanism for the inherent logical consistency between multi-source data, and cannot identify semantic contradictions or omissions between text records and image representations, resulting in the final quality control scoring results having great subjective bias and verification blind spots, making it difficult to truly and objectively reflect the quality of clinical nursing. Summary of the Invention

[0005] Therefore, it is necessary to provide a nursing quality control intelligent verification and scoring method, device, equipment, and medium that can dynamically adjust the verification weight based on the patient's pathological characteristics and automatically detect semantic conflicts in multimodal data to address the above-mentioned technical problems.

[0006] Firstly, this application provides a nursing quality control intelligent verification and scoring method, including:

[0007] S1. Obtain patient electronic medical record data based on bed identification information, and perform entity relationship joint extraction on the patient electronic medical record data to obtain a set of multidimensional pathological feature vectors of the patient.

[0008] S2. Based on the ontology knowledge graph of the nursing quality control domain and the set of multidimensional pathological feature vectors of patients, the static quality control rule tree is reconstructed in an adaptive manner to generate a customized quality control check graph with adaptive weight coefficients.

[0009] S3. Obtain multimodal verification data stream based on customized quality control verification map, and perform cross-modal feature alignment on the multimodal verification data stream to obtain the aligned multimodal feature matrix;

[0010] S4. Perform semantic conflict detection on the aligned multimodal feature matrix to obtain the verification node data that has passed the verification.

[0011] S5. Based on the adaptive weight coefficients in the customized quality control verification map, the verification node data that has passed the verification is weighted and aggregated to generate a global quality control score result.

[0012] In one embodiment, S2 includes:

[0013] S21. Based on the ontology knowledge graph of the nursing quality control domain and the set of multidimensional pathological feature vectors of patients, a graph neural network is used to map the static quality control rule tree to obtain the initial quality control node feature matrix.

[0014] S22. Based on the topological connection relationship in the ontology knowledge graph of the nursing quality control domain, the feature matrix of the initial quality control node is aggregated with the features of neighboring nodes to obtain the updated quality control node feature matrix.

[0015] S23. Based on the updated quality control node feature matrix, the basic weights of each quality control item node in the static quality control rule tree are dynamically amplified and calculated to obtain the adaptive weight coefficients of each quality control item node. The formula for calculating the adaptive weight coefficients is as follows:

[0016]

[0017] In the formula, For the first Adaptive weighting coefficients for each quality control project node. For the first The basic weight of each quality control project node. The total number of feature entities in the patient's multidimensional pathological feature vector set. For the first Each feature entity vector For the first Vector representation of each quality control project node The cosine similarity function is used. For cross-attention calculation function, The penalty factor is amplified for weighting;

[0018] S24. Based on the adaptive weight coefficients and preset dynamic thresholds of each quality control project node, the static quality control rule tree is topologically sorted and reorganized to generate a customized quality control verification map carrying adaptive weight coefficients.

[0019] In one embodiment, S23 includes:

[0020] S231. Based on the updated quality control node feature matrix, calculate the similarity between each feature entity and each quality control item node in the patient's multidimensional pathological feature vector set to obtain the semantic similarity between each feature entity and each quality control item node.

[0021] S232. Calculate the attention weight allocation for each feature entity and each quality control item node to obtain the cross-attention score of each feature entity to each quality control item node.

[0022] S233. Based on semantic similarity and cross-attention scores, the basic weights of each quality control item node in the static quality control rule tree are dynamically amplified and calculated to obtain the adaptive weight coefficients of each quality control item node.

[0023] In one embodiment, S3 includes:

[0024] S31. Perform multimodal feature extraction on the multimodal verification data stream to obtain a multimodal basic feature set, which includes text feature vectors, visual feature vectors, and standard semantic anchor vectors;

[0025] S32. Based on the modal variance of text feature vectors and visual feature vectors, perform inverse variance weighted calculation on text feature vectors and visual feature vectors to obtain a balanced feature vector set;

[0026] S33. Perform multi-head cross-attention mapping on the balanced feature vector set to obtain the initial attention feature matrix;

[0027] S34. Perform residual connection calculation based on the initial attention feature matrix and standard semantic anchor vector to obtain the residual feature matrix;

[0028] S35. Perform layer normalization calculation on the residual feature matrix to obtain the aligned multimodal feature matrix.

[0029] In one embodiment, S4 includes:

[0030] S41. Decouple the aligned multimodal feature matrix to obtain the aligned text feature vector, the aligned visual feature vector, and the aligned standard semantic anchor vector.

[0031] S42. Based on the aligned standard semantic anchor vector, perform deviation feature mapping on the aligned text feature vector to obtain the text semantic deviation vector.

[0032] S43. Based on the aligned standard semantic anchor vector, perform deviation feature mapping on the aligned visual feature vector to obtain the visual semantic deviation vector.

[0033] S44. The text semantic deviation vector and visual semantic deviation features are fused and weighted to obtain the multimodal conflict probability value. The formula for calculating the multimodal conflict probability value is as follows:

[0034]

[0035] In the formula, This represents the multimodal conflict probability value. The aligned standard semantic anchor vector, This is the aligned text feature vector. The aligned visual feature vectors and It is a linear transformation matrix. and For bias terms, For activation function, This indicates a feature concatenation operation. To fuse the weight matrix, To incorporate the bias term, Use the Sigmoid activation function;

[0036] S45. Based on the multimodal conflict probability value and the preset adaptive conflict judgment threshold, the aligned multimodal feature matrix is ​​frozen by state marking to obtain the verification node data that has passed the verification.

[0037] In one embodiment, S5 includes:

[0038] S51. Based on the verification node data that has passed the verification, extract the multimodal conflict probability value corresponding to each verification node to obtain a set of conflict probability values.

[0039] S52. Based on the numerical range of each element in the conflict probability value set, perform confidence penalty factor matching on each verification node to obtain the confidence penalty coefficient of each verification node.

[0040] S53. Based on the confidence penalty coefficient of each verification node, perform a second decay product calculation on the adaptive weight coefficient in the customized quality control verification map to obtain the final effective weight of each verification node.

[0041] S54. Based on the final effective weight of each verification node, the data of the verification nodes that have passed the verification are weighted and aggregated to generate a global quality control score. The formula for calculating the global quality control score is as follows:

[0042]

[0043] In the formula, This is the overall quality control score result. This represents the total number of nodes actually verified. For the first The adaptive weight coefficients of each verification node For the first The confidence penalty coefficient for each verification node. For the first The structured checklist status codes for each verification node For the first The historical multimodal conflict probability value corresponding to each verification node. This is the node score mapping function.

[0044] In one embodiment, S53 includes:

[0045] S531. Construct a diagonal penalty matrix based on the confidence penalty coefficient of each verification node to obtain the node-level penalty tensor;

[0046] S532. The adaptive weight coefficients in the customized quality control verification map are arranged in a vectorized manner to obtain the basic weight vector;

[0047] S533. Perform Hadamard product calculation on the node-level penalty tensor and the basic weight vector to obtain the initial effective weight vector.

[0048] S534. Extract the extreme values ​​of the initial effective weight vector to obtain the maximum and minimum weight values ​​of the initial effective weight vector;

[0049] S535. Based on the maximum weight value of the vector, the minimum weight value of the vector, and the preset normalization interval, the range linear scaling calculation is performed on the initial effective weight vector to obtain the final effective weight of each verification node.

[0050] Secondly, this application also provides a nursing quality control intelligent verification and scoring device, comprising:

[0051] The feature extraction module is used to obtain patient electronic medical record data based on bed identification information, and to perform joint entity relation extraction on the patient electronic medical record data to obtain a set of multidimensional pathological feature vectors of the patient.

[0052] The graph reconstruction module is used to adaptively reconstruct the static quality control rule tree based on the ontology knowledge graph of the nursing quality control field and the set of multidimensional pathological feature vectors of patients, and generate a customized quality control verification graph with adaptive weight coefficients.

[0053] The cross-modal alignment module is used to obtain multimodal verification data streams based on customized quality control verification maps, and to perform cross-modal feature alignment on the multimodal verification data streams to obtain the aligned multimodal feature matrix;

[0054] The conflict detection module is used to perform semantic conflict detection on the aligned multimodal feature matrix to obtain the verification node data that has passed the verification.

[0055] The weighted scoring module is used to perform weighted aggregation scoring on the verified node data based on the adaptive weight coefficients in the customized quality control verification map, and generate a global quality control score result.

[0056] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the nursing quality control intelligent verification and scoring method as described in the first aspect.

[0057] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the nursing quality control intelligent verification and scoring method as described in the first aspect.

[0058] The aforementioned intelligent verification and scoring method, device, equipment, and medium for nursing quality control extracts a set of multidimensional pathological feature vectors from patient medical records based on bed identification information. It then combines this with a knowledge graph of the nursing quality control ontology to adaptively reconstruct a static rule tree, generating a customized quality control verification graph with adaptive weight coefficients. Based on this graph, it acquires multimodal verification data streams and performs cross-modal feature alignment to obtain a feature matrix. Semantic conflict detection is then performed on this feature matrix to filter out verified verification node data. Finally, based on the adaptive weight coefficients in the graph, the verified node data is weighted and aggregated to generate a global quality control score. This achieves dynamic adaptive adjustment of quality control scoring rules to the patient's personalized pathological characteristics and effectively eliminates interference from potential semantic contradictions between multimodal verification data. Ultimately, it significantly improves the objectivity and accuracy of nursing quality control scoring results and their consistency with the patient's actual clinical condition. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 A schematic diagram illustrating the implementation environment of the intelligent verification and scoring method for nursing quality control provided by this invention;

[0061] Figure 2 A flowchart illustrating an intelligent verification and scoring method for nursing quality control provided by this invention;

[0062] Figure 3 This is a schematic diagram of the structure of a nursing quality control intelligent verification and scoring device provided by the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] The nursing quality control intelligent verification and scoring method provided in this application embodiment can be applied to, for example... Figure 1 In the implementation environment shown, the quality control intelligent terminal 100 communicates with the server 200 via network 300. The data storage system 201 can store the data that the server needs to process. The data storage system 201 can be integrated into the server or placed in the cloud or other network servers. Specifically, the quality control intelligent terminal 100 is a mobile phone, tablet computer, or dedicated hardware terminal that supports mobile use within the ward. This terminal is equipped with a barcode scanning module 101 (such as a camera or barcode scanner), an image acquisition module 102, and a touch interface 103. It is used to scan the patient's bed identification code using the barcode scanning module 101 to obtain bed identification information, acquire images of the nursing environment using the image acquisition module 102, and receive text input via the touch interface 103. It then uploads the acquired bed identification information and multimodal verification data stream to the server 200. Server 200 communicates with the hardware nodes of the hospital's existing electronic medical record system 400 via network 300. Based on the received bed identification information, it retrieves the corresponding patient's electronic medical record data from the electronic medical record system 400, processes it, and finally sends the generated global quality control score result to the touch interface 103 of the quality control intelligent terminal 100 for visualization. Server 200 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0065] In one exemplary embodiment, such as Figure 2 As shown, a nursing quality control intelligent verification and scoring method is provided, which is applied to... Figure 1 Taking the quality control intelligent terminal as an example, the following steps are taken: S1 to S5. Among them:

[0066] S1. Obtain patient electronic medical record data based on bed identification information, and perform entity relationship joint extraction on the patient electronic medical record data to obtain a set of multidimensional pathological feature vectors of the patient.

[0067] Optionally, the quality control intelligent terminal reads the unique identity index corresponding to the bed identification information, and calls the medical data storage interface based on the identity index to retrieve the patient's electronic medical record data. The patient's electronic medical record data includes four types of text data: diagnostic text, test records, medication records, and nursing records.

[0068] Preferably, feature extraction of text data is completed through an entity relation joint extraction model. The entity relation joint extraction model is trained under supervision using labeled nursing medical record corpora. The training process uses the medical record entity labeling results and entity relationship labeling results as training labels, and the cross-entropy loss function is used to iteratively optimize the model parameters.

[0069] Furthermore, the pathological entities and entity relationships in the medical record text are identified through the entity relationship joint extraction model. The pathological entities and entity relationships are mapped into fixed-dimensional numerical vectors, and all numerical vectors are combined to obtain a set of multidimensional pathological feature vectors of the patient.

[0070] S2. Based on the ontology knowledge graph of the nursing quality control domain and the set of multidimensional pathological feature vectors of patients, the static quality control rule tree is reconstructed by graph adaptation to generate a customized quality control check graph with adaptive weight coefficients.

[0071] Optionally, the nursing quality control domain ontology knowledge graph is a pre-constructed structured knowledge set, containing nursing quality control concept nodes, concept attributes, and concept relationships. The static quality control rule tree is a predefined hierarchical set of quality control verification logic. The patient's multi-dimensional pathological feature vector set is input into the nursing quality control domain ontology knowledge graph to complete semantic matching, obtaining a subset of quality control concepts that match the patient's pathological features. Based on this subset, the node hierarchy and node association links of the static quality control rule tree are adjusted. The importance value of each reconstructed node is calculated based on semantic matching similarity, and this importance value is mapped to an adaptive weight coefficient. The reconstructed rule nodes and the adaptive weight coefficients are then integrated to generate a customized quality control verification graph.

[0072] S3. Obtain multimodal verification data stream based on customized quality control verification map, and perform cross-modal feature alignment on the multimodal verification data stream to obtain the aligned multimodal feature matrix.

[0073] Optionally, based on the node types of the customized quality control verification map, text-based verification data, image-based verification data, and numerical verification data are retrieved separately, and multiple types of data are combined to obtain a multimodal verification data stream. Feature encoding is performed on each modality of the multimodal verification data stream to generate independent initial feature sequences for each modality. A cross-modal feature alignment algorithm is used to establish semantic mapping relationships between different modal feature sequences. Based on the semantic mapping relationships, the dimensions and semantic spaces of all modal features are unified, eliminating semantic heterogeneity between different modal data. Multiple sets of feature sequences with unified dimensions and semantic matching are combined into an aligned multimodal feature matrix.

[0074] S4. Perform semantic conflict detection on the aligned multimodal feature matrix to obtain the verification node data that has passed the verification.

[0075] Optionally, all feature vectors in the aligned multimodal feature matrix are traversed, and the semantic description information and quality control constraints corresponding to each feature vector are extracted. Semantic conflict detection is performed using semantic similarity matching to complete logical verification. The semantic similarity values ​​of different feature vectors under the same verification dimension are calculated, and the semantic similarity values ​​are compared with a predefined conflict judgment threshold. Feature vectors with semantic deviations are identified as conflicting data and removed. Feature vectors that meet the semantic consistency and quality control constraints are retained, and the data content corresponding to all retained feature vectors is integrated to obtain the verification node data that has passed the verification.

[0076] S5. Based on the adaptive weight coefficients in the customized quality control verification map, the verification node data that has passed the verification is weighted and aggregated to generate a global quality control score result.

[0077] Optionally, adaptive weight coefficients corresponding to each verification node in the customized quality control verification map are extracted, establishing a one-to-one mapping relationship between the adaptive weight coefficients and the verified verification node data. A weighted summation calculation is used to aggregate the scores, with the quantitative characteristic values ​​of the verified verification node data serving as the calculation base and the adaptive weight coefficients of the corresponding nodes as the weighting coefficients. The weighted calculation results of all nodes are summed to generate a global quality control score, which is used to characterize the overall compliance level of patient care quality control.

[0078] In the above-mentioned intelligent verification and scoring method for nursing quality control, precise pathological features are obtained by extracting entities from electronic medical records, and a customized quality control graph is reconstructed by combining ontology knowledge graph. After completing multimodal data alignment and conflict verification, weighted scoring is achieved, which effectively adapts to the pathological features of different patients, improves the accuracy and adaptability of nursing quality control verification, and reduces the probability of misjudgment in quality control verification.

[0079] In one embodiment, S2 may include:

[0080] S21. Based on the ontology knowledge graph of the nursing quality control domain and the set of multidimensional pathological feature vectors of patients, a graph neural network is used to map the static quality control rule tree to obtain the initial quality control node feature matrix.

[0081] Optionally, feature mapping processing of multi-source data is completed through graph neural networks. The graph neural network is trained under supervision using a nursing quality control node labeled dataset. The training process optimizes the node feature fitting accuracy and uses the mean squared error loss function to update the parameters iteratively.

[0082] Furthermore, the semantic features of the ontology knowledge graph in the nursing quality control domain, the pathological features of the patient's multidimensional pathological feature vector set, and the node attributes of the static quality control rule tree are integrated into a unified vector embedding. The discrete rule nodes and associated features are converted into continuous numerical representations. All node vectors are combined according to the node arrangement order to generate an initial quality control node feature matrix.

[0083] S22. Based on the topological connection relationship in the ontology knowledge graph of the nursing quality control domain, the feature matrix of the initial quality control node is aggregated with the features of neighboring nodes to obtain the updated feature matrix of the quality control node.

[0084] Optionally, edge connection data of all nodes in the ontology knowledge graph of the nursing quality control domain is extracted. The topological connection relationship is the semantic association link between concept nodes in the ontology knowledge graph, and the neighbor node is the associated node that has a direct topological connection with the target quality control project node. Node feature fusion is completed by summing and aggregation. For each target node in the initial quality control node feature matrix, the feature values ​​of all neighbor nodes of the target node are accumulated. The fused feature values ​​replace the original features of the target node, and the full node feature iterative update is completed. The updated quality control node feature matrix is ​​obtained by combining all the updated node vectors.

[0085] S23. Based on the updated quality control node feature matrix, the basic weights of each quality control item node in the static quality control rule tree are dynamically amplified and calculated to obtain the adaptive weight coefficients of each quality control item node. The formula for calculating the adaptive weight coefficients is as follows:

[0086]

[0087] In the formula, For the first Adaptive weighting coefficients for each quality control project node. For the first The basic weight of each quality control project node. The total number of feature entities in the patient's multidimensional pathological feature vector set. For the first Each feature entity vector For the first Vector representation of each quality control project node The cosine similarity function is used. For cross-attention calculation function, The penalty factor is used to amplify the weight.

[0088] Optionally, the vectors of each node in the updated quality control node feature matrix are extracted, and the pre-configured base weights in the static quality control rule tree are retrieved. These base weights are the general fixed weight values ​​for the quality control items. The cosine similarity function is used to calculate the directional matching degree between two vectors in the semantic space. The cross-attention calculation function is trained using the associated labeled corpus of pathological entities and quality control nodes, used to quantify the association contribution of feature entities to the quality control nodes. The weight amplification penalty factor is a predefined adjustment parameter used to constrain the weight amplification magnitude. The numerical calculations are completed by substituting the values ​​into the formula, and the adaptive weight coefficients corresponding to each quality control item node are output sequentially.

[0089] S24. Based on the adaptive weight coefficients and preset dynamic thresholds of each quality control project node, the static quality control rule tree is topologically sorted and reorganized to generate a customized quality control verification map carrying adaptive weight coefficients.

[0090] Optionally, a preset dynamic threshold is retrieved. This preset dynamic threshold is a predefined judgment value used to filter valid quality control nodes. The topology sorting and reorganization adjusts the node hierarchy and connection links based on the numerical distribution of the adaptive weight coefficients. The adaptive weight coefficients of each quality control item node are compared with the preset dynamic threshold. Invalid nodes with adaptive weight coefficients lower than the preset dynamic threshold are removed, while valid nodes with adaptive weight coefficients that meet the threshold requirements are retained. The hierarchical arrangement and topology connection relationship of the valid nodes are adjusted, and the adaptive weight coefficients corresponding to each valid node are bound to generate a customized quality control verification map carrying the adaptive weight coefficients.

[0091] In the above embodiments, node feature mapping is completed through graph neural networks, neighbor feature aggregation is completed by combining ontology topology, adaptive weights are calculated and topology reorganization is completed, so as to achieve deep adaptation between the quality control rule tree and the patient's pathological features, accurately quantify the importance of quality control nodes, and improve the personalization and verification accuracy of customized quality control verification maps.

[0092] In one embodiment, S23 may include:

[0093] S231. Based on the updated quality control node feature matrix, calculate the similarity between each feature entity and each quality control item node in the patient's multidimensional pathological feature vector set to obtain the semantic similarity between each feature entity and each quality control item node.

[0094] Optionally, vector representations of all quality control project nodes are extracted from the updated quality control node feature matrix. Simultaneously, vector data of all feature entities within the patient's multidimensional pathological feature vector set are retrieved, and bidirectional semantic matching is performed using cosine similarity calculation. Cosine similarity quantifies the semantic association between two sets of vectors based on the angle between them in the vector space, and outputs matching values ​​directly through standardized vector dot product operations. All combinations of feature entities and all quality control project nodes are traversed to complete full pairing calculations, generating a semantic similarity value set covering all pairing combinations. Higher semantic similarity values ​​indicate a higher semantic fit between the feature entity and the quality control project node.

[0095] S232. Calculate the attention weight allocation for each feature entity and each quality control item node to obtain the cross-attention score of each feature entity to each quality control item node.

[0096] Optionally, a cross-attention calculation logic is used to complete the weight allocation operation. The cross-attention calculation logic is trained with supervised training using annotated pathological entities and quality control nodes associated corpus. The training process aims to optimize the association matching accuracy and complete the convergence adjustment of parameters.

[0097] For example, feature entity vectors and quality control project node vectors are used as dual input data. Attention encoding is used to quantify the contribution of a single set of feature entities to a single set of quality control project nodes, and the corresponding cross-attention scores are output. By completing the operation of the full combination, a set of cross-attention scores corresponding one-to-one with the semantic similarity pairings is generated to characterize the strength of the association influence of feature entities on quality control project nodes.

[0098] S233. Based on semantic similarity and cross-attention scores, the basic weights of each quality control item node in the static quality control rule tree are dynamically amplified and calculated to obtain the adaptive weight coefficients of each quality control item node.

[0099] Optionally, the preset base weights of each quality control item node in the static quality control rule tree are retrieved. These base weights are fixed values ​​adapted to general nursing scenarios and are used to characterize the importance of basic checks on the quality control items. The semantic similarity and cross-attention scores under the same pair are multiplied and fused, and the fused values ​​corresponding to all feature entities are accumulated. Predefined adjustment parameters are then used to amplify the values, and the amplified values ​​are multiplied by the base weights. By performing the above calculation process on each quality control item node, the adaptive weight coefficients corresponding to each quality control item node are output, achieving personalized adaptation of the weight values ​​to the patient's pathological characteristics.

[0100] In the above embodiments, the semantic similarity calculation of feature entities and quality control nodes, the cross-attention score allocation, and the basic weight dynamic amplification operation are completed sequentially to accurately quantify the correlation strength between the two, generate adaptive weight coefficients that are adapted to the individual pathological characteristics of patients, and improve the accuracy and personalized adaptation capability of quality control weight allocation.

[0101] In one embodiment, S3 may include:

[0102] S31. Perform multimodal feature extraction on the multimodal verification data stream to obtain a multimodal basic feature set, which includes text feature vectors, visual feature vectors, and standard semantic anchor vectors.

[0103] Optionally, the multimodal verification data stream is processed by independent encoding for each modality. Text-type verification data is vectorized using pre-trained word embedding encoding, while visual-type verification data is mapped to dimensions using convolutional encoding. The standard semantic anchor vector is a predefined unified semantic benchmark vector for nursing quality control, used to unify the semantic reference standard of multimodal data. Vectors are classified and integrated according to modality type. Text feature vectors, visual feature vectors, and standard semantic anchor vectors are combined to generate a well-structured set of multimodal basic features.

[0104] S32. Based on the modal variance of the text feature vector and the visual feature vector, perform inverse variance weighted calculation on the text feature vector and the visual feature vector to obtain a balanced feature vector set.

[0105] Optionally, the global distribution dispersion of the text feature vector and the global distribution dispersion of the visual feature vector are calculated separately to generate the modality variance values ​​for the corresponding modalities. The modality variance is used to characterize the distribution fluctuation amplitude of the single-modality feature data. Modality feature balancing is performed using inverse variance weighting, which uses the reciprocal of the modality variance as the weighting coefficient. Element-wise weighting is applied to the text and visual feature vectors to weaken the feature proportion of high-fluctuation modes and strengthen the feature contribution of low-fluctuation modes. The results are then integrated to generate a balanced feature vector set, eliminating interference from distribution differences between modalities.

[0106] S33. Perform multi-head cross-attention mapping on the balanced feature vector set to obtain the initial attention feature matrix.

[0107] Optionally, a multi-head cross-attention coding layer is used to complete the feature mapping process. The multi-head cross-attention coding layer is trained under supervision using nursing quality control multimodal labeled corpus. The training process aims to achieve parameter convergence with the modality feature matching accuracy as the optimization objective.

[0108] Furthermore, the balanced feature vector set is used as dual-path input data, and multiple sets of parallel attention heads are split to complete independent semantic association operations, capturing fine-grained semantic association relationships between features of different modalities. The output features of all attention heads are concatenated to complete dimensional integration, generating an initial attention feature matrix with unified dimensions and enhanced semantic association.

[0109] S34. Perform residual connection calculation based on the initial attention feature matrix and standard semantic anchor vector to obtain the residual feature matrix.

[0110] Optionally, residual connection processing is performed using element-wise addition. This residual connection operation fuses semantic association features with baseline semantic features while preserving the core information of the original features, reducing the probability of information loss during feature mapping. The global expansion matrix of the standard semantic anchor vector is retrieved, and the initial attention feature matrix is ​​added dimension-wise to the global expansion matrix of the standard semantic anchor vector to complete the fusion embedding of the baseline semantics. This generates a residual feature matrix that fuses the baseline and association semantics, enhancing the semantic consistency of the features.

[0111] S35. Perform layer normalization calculation on the residual feature matrix to obtain the aligned multimodal feature matrix.

[0112] Optionally, layer normalization is used to standardize the features. Layer normalization unifies the numerical distribution range of the feature matrix, eliminating semantic bias caused by differences in the magnitude of feature values. By calculating the mean and standard deviation of the residual feature matrix within a single dimension, a standardization transformation is performed on an eigenvalue-by-eigenvalue basis according to the normalization formula, mapping all feature values ​​to a unified distribution range. After completing the normalization operation of the entire matrix, an aligned multimodal feature matrix with a regular distribution and unified semantic space is generated, achieving accurate cross-modal semantic alignment of multimodal data.

[0113] In the above embodiments, multimodal feature extraction, inverse variance mode balancing, multi-head attention mapping, residual semantic fusion and layer normalization are completed sequentially, effectively eliminating modal heterogeneity and distribution differences, achieving accurate semantic alignment of multimodal data, and improving the feature reliability of subsequent quality control verification.

[0114] In one embodiment, S4 may include:

[0115] S41. Decouple the aligned multimodal feature matrix to obtain the aligned text feature vector, the aligned visual feature vector, and the aligned standard semantic anchor vector.

[0116] Optionally, based on preset modality dimension partitioning rules, dimensional splitting feature decoupling processing is performed on the aligned multimodal feature matrix. Feature decoupling is achieved through matrix dimension slicing operations, directly separating data according to the fixed storage dimensions of the modal features. This accurately separates independent-dimensional text feature data, visual feature data, and baseline semantic feature data, while simultaneously preserving the dimensional specifications and semantic representation integrity of all feature vectors, generating aligned text feature vectors, aligned visual feature vectors, and aligned standard semantic anchor vectors, respectively.

[0117] S42. Based on the aligned standard semantic anchor vector, perform deviation feature mapping on the aligned text feature vector to obtain the text semantic deviation vector.

[0118] Optionally, the aligned standard semantic anchor vector is used as a unified semantic benchmark. Deviation quantification is performed using vector difference operations combined with linear mapping. The linear mapping parameters are optimized based on the nursing quality control text semantic annotation dataset. The numerical difference between the aligned text feature vector and the aligned standard semantic anchor vector is calculated dimension by dimension. The difference vector is input into the linear mapping layer to complete the feature transformation of semantic deviation. The output is a text semantic deviation vector representing the degree of deviation between the text semantics and the benchmark semantics, thus achieving a preliminary quantitative representation of text modality semantic anomalies.

[0119] S43. Based on the aligned standard semantic anchor vector, perform deviation feature mapping calculation on the aligned visual feature vector to obtain the visual semantic deviation vector.

[0120] Optionally, using a unified deviation quantification logic, and taking the aligned standard semantic anchor vector as a benchmark, a dedicated deviation mapping process is performed on the visual modality features. The linear mapping parameters of the visual modality are adapted to the distribution characteristics of the visual features to achieve independent fitting. The dimensionality difference between the aligned visual feature vector and the aligned standard semantic anchor vector is calculated. By adapting the linear mapping of the visual features, the deviation feature transformation is completed, generating a visual semantic deviation vector. This accurately quantifies the degree of deviation between the visual modality data and the quality control benchmark semantics, forming a bimodal deviation representation with the text semantic deviation vector.

[0121] S44. The text semantic deviation vector and visual semantic deviation features are fused and weighted to obtain the multimodal conflict probability value. The formula for calculating the multimodal conflict probability value is as follows:

[0122]

[0123] In the formula, This represents the multimodal conflict probability value. The aligned standard semantic anchor vector, This is the aligned text feature vector. The aligned visual feature vectors and It is a linear transformation matrix. and For bias terms, For activation function, This indicates a feature concatenation operation. To fuse the weight matrix, To incorporate the bias term, This is the Sigmoid activation function.

[0124] Optionally, the bimodal deviation vector is retrieved for fusion operation. The linear transformation matrix, bias term, fusion weight matrix, and fusion bias term are all trained under supervision using a multimodal semantic conflict annotation dataset, with conflict recognition accuracy as the optimization objective to achieve parameter convergence. The Rectified Luminaire (ReLU) function is used to filter invalid negative bias features, feature concatenation is used to merge bimodal bias features, and the Sigmoid activation function is used to map the fused features to a standardized probability interval. The entire numerical computation process is performed according to the given formula, outputting multimodal conflict probability values ​​that characterize the degree of semantic conflict in the multimodal data.

[0125] S45. Based on the multimodal conflict probability value and the preset adaptive conflict judgment threshold, the aligned multimodal feature matrix is ​​frozen by state marking to obtain the verification node data that has passed the verification.

[0126] Optionally, a preset adaptive conflict judgment threshold is retrieved. This threshold is a predefined semantic conflict judgment boundary value adapted to the quality control scenario, used to distinguish between compliance features and conflict features. All feature units in the aligned multimodal feature matrix are traversed, and their corresponding multimodal conflict probability values ​​are compared with the adaptive conflict judgment threshold. Feature units whose conflict probabilities meet compliance requirements are marked with a status and frozen. Abnormal feature units with excessive conflict probabilities are removed. All frozen and retained compliance feature data are integrated to obtain the verification node data that has passed the verification.

[0127] In the above embodiments, bimodal features are decoupled and separated by multimodal features, the bimodal semantic deviation is quantified and the conflict probability is calculated by fusion, and feature filtering and freezing are completed by combining thresholds. Multimodal semantic conflict data is accurately identified and eliminated, effectively ensuring the semantic consistency of the verification node data and the reliability of quality control verification.

[0128] In one embodiment, S5 may include:

[0129] S51. Based on the verification node data that has passed the verification, extract the multimodal conflict probability value corresponding to each verification node to obtain a set of conflict probability values.

[0130] Optionally, a unique index association is established between the verified check node data and the multimodal conflict probability values. A targeted data retrieval operation is performed based on the unique node identifier, and the retrieval process uses dimensional matching to achieve accurate data extraction. All verified check nodes are traversed to extract all corresponding multimodal conflict probability values ​​in batches. The values ​​are then organized and combined according to the topological arrangement of the check nodes to generate a structurally ordered set of conflict probability values.

[0131] S52. Based on the numerical range of each element in the conflict probability value set, perform confidence penalty factor matching on each verification node to obtain the confidence penalty coefficient of each verification node.

[0132] Optionally, a pre-constructed mapping table between numerical intervals and penalty factors is retrieved. This mapping table is based on predefined grading matching rules adapted to the nursing quality control scenario. Different numerical intervals correspond to differentiated confidence penalty coefficients, used to quantify the credibility of node data. All numerical elements within the conflict probability value set are traversed, and the preset numerical interval to which each element belongs is determined. A one-to-one matching assignment is completed according to the mapping table, outputting a unique confidence penalty coefficient for each verification node, thus achieving a quantitative and grading representation of the credibility of node data.

[0133] S53. Based on the confidence penalty coefficient of each verification node, perform a second decay product calculation on the adaptive weight coefficient in the customized quality control verification map to obtain the final effective weight of each verification node.

[0134] Optionally, element-wise multiplication is used to perform weight decay. The calculation process uses the adaptive weight coefficients in the customized quality control verification map as the base weight values ​​and the confidence penalty coefficients of the verification nodes as the decay adjustment values. The two are directly multiplied to complete the secondary weight correction. A unified multiplication process is performed on all verification nodes, weakening the weight contribution of nodes with high conflict probability and strengthening the weight proportion of high-confidence nodes. After the calculation, the final effective weights for each verification node are generated, providing accurate weight basis for global score aggregation.

[0135] S54. Based on the final effective weight of each verification node, the data of the verification nodes that have passed the verification are weighted and aggregated to generate a global quality control score. The formula for calculating the global quality control score is as follows:

[0136]

[0137] In the formula, This is the overall quality control score result. This represents the total number of nodes actually verified. For the first The adaptive weight coefficients of each verification node For the first The confidence penalty coefficient for each verification node. For the first The structured checklist status codes for each verification node For the first The historical multimodal conflict probability value corresponding to each verification node. This is the node score mapping function.

[0138] Optionally, the associated parameters of all verification nodes are retrieved and weighted normalized aggregation operations are performed. The structured checkmark status code is a discrete coded value representing the quality control compliance status of the node. The node score mapping function uses linear mapping logic to complete the standardization transformation of the status values. The numerator weighted score accumulation and denominator weight accumulation operations are completed according to the given formula. The normalized global quality control score result is output through ratio calculation. The global quality control score result is used to quantitatively represent the overall compliance level of patient care quality control across all dimensions, realizing the standardized and quantitative output of quality control results.

[0139] In the above embodiments, by extracting the node conflict probability, matching the confidence penalty coefficient, correcting the effective weight of the node and completing the weighted normalization score, and combining the data credibility to dynamically adjust the weight contribution, the overall quality control compliance level is accurately quantified, thereby improving the objectivity and accuracy of the nursing quality control score results.

[0140] In one embodiment, S53 may include:

[0141] S531. Construct a diagonal penalty matrix based on the confidence penalty coefficient of each verification node to obtain the node-level penalty tensor.

[0142] Optionally, using the total number of verification nodes as the matrix dimension specification, a diagonal matrix construction logic is employed to construct the numerical matrix. The values ​​on the main diagonal of the diagonal matrix are sequentially filled with the confidence penalty coefficients corresponding to each verification node, while the values ​​on the non-main diagonal positions are uniformly set to zero. The constructed two-dimensional diagonal matrix is ​​then converted into a node-level penalty tensor, with the tensor dimension perfectly matching the number of verification nodes, ensuring that each confidence penalty coefficient forms a unique dimensional mapping with its corresponding verification node. This process is implemented using basic matrix operations.

[0143] S532. The adaptive weight coefficients in the customized quality control verification map are arranged in a vectorized manner to obtain the basic weight vector.

[0144] Optionally, based on the node topology sorting rules of the customized quality control verification map, a sequential reorganization process is performed on the adaptive weight coefficients corresponding to all verification nodes. The reorganization process strictly follows the hierarchical arrangement order of the nodes and does not change the original value of any adaptive weight coefficient. The discretely distributed adaptive weight coefficients are integrated into a one-dimensional ordered vector structure to generate a basic weight vector. The length of the basic weight vector is consistent with the total number of verification nodes, achieving dimensional uniformity of the weight values ​​and ensuring that it has a fully matched computational dimension with the node-level penalty tensor, providing compliant input data for element-wise multiplication operations.

[0145] S533. Perform Hadamard product calculation on the node-level penalty tensor and the basic weight vector to obtain the initial effective weight vector.

[0146] Optionally, the Hadamard product is used to perform the fusion operation between the tensor and the vector. The Hadamard product is an element-wise multiplication operation of data of the same dimension. During the operation, the dimensional structure of the data is not changed, and only the multiplication of corresponding values ​​is performed. The node-level penalty tensor and the basic weight vector are multiplied correspondingly in each dimension. The confidence penalty is corrected for the adaptive weight coefficient of each check node, weakening the weight contribution intensity of nodes with high conflict probability. After the operation is completed, an initial effective weight vector with unified dimensions and corrected values ​​is generated, realizing the initial precise adjustment of node weights.

[0147] S534. Extract the extreme values ​​of the initial effective weight vector to obtain the maximum and minimum weight values ​​of the initial effective weight vector.

[0148] Optionally, the extreme value extraction operation is performed by comparing the values ​​of all elements in the initial effective weight vector through a full vector traversal. During the traversal, the values ​​of all elements in the initial effective weight vector are compared sequentially, and the global maximum and global minimum values ​​in the vector are selected through iterative comparisons. The global maximum value obtained by selection is marked as the maximum weight value of the initial effective weight vector, and the global minimum value obtained by selection is marked as the minimum weight value of the initial effective weight vector.

[0149] S535. Based on the maximum weight value of the vector, the minimum weight value of the vector, and the preset normalization interval, the range linear scaling calculation is performed on the initial effective weight vector to obtain the final effective weight of each verification node.

[0150] Optionally, the upper and lower limits of a predefined normalization interval are retrieved. The preset normalization interval is a uniformly set weight value distribution interval in the nursing quality control scenario, used to eliminate differences in the magnitude of the initial effective weight vector and ensure the numerical stability of the global scoring calculation. Element-by-element numerical transformation is completed using the range normalization linear mapping formula, which is:

[0151]

[0152] In the formula, For the first The final effective weight of each verification node, This is the lower limit value of the preset normalization interval. To set the upper limit value of the preset normalization interval, The maximum weight value of the initial effective weight vector. The minimum weight value of the initial effective weight vector. For the first The initial effective weight values ​​corresponding to each verification node are traversed. All elements in the initial effective weight vector are substituted into the formula one by one to complete the linear scaling operation, mapping all original weight values ​​to a unified normalization interval without difference. Finally, the final effective weights of each verification node with consistent dimensions and uniform units are output.

[0153] In the above embodiments, by constructing a diagonal penalty tensor, reorganizing the weight vector, performing element-wise product correction, extracting extreme values, and normalizing the range, the accurate penalty correction and standardization of the weights are completed, eliminating the differences in weight dimensions and improving the rationality of weight allocation and the stability of global scoring calculation.

[0154] In the aforementioned intelligent verification and scoring method, device, equipment, and medium for nursing quality control, targeting nursing quality control verification application scenarios, patient electronic medical record data is retrieved using bed identification information as an index. A multidimensional pathological feature vector set of patients is constructed through entity relation joint extraction. Based on the ontology knowledge graph of the nursing quality control domain, graph neural network feature mapping and neighbor aggregation operations are employed. Semantic similarity and cross-attention are combined to dynamically amplify the weights of quality control nodes. A topological reorganization of the static quality control rule tree generates a customized quality control verification graph. Simultaneously, multimodal verification data streams are collected, and cross-modal feature alignment is achieved through modal balancing, cross-attention mapping, and normalization. Conflict detection and compliant node data are completed through semantic deviation fusion operations. Then, a tensor is constructed based on the confidence penalty coefficient to complete weight decay correction and normalization processing. Finally, a global quality control score result is generated through weighted normalization aggregation operations. This technical solution effectively addresses the technical problems of existing nursing quality control, such as fixed rules, semantic heterogeneity of multimodal data, and lack of data credibility correction in quality control scoring, leading to poor verification adaptability, high conflict and misjudgment rates, and insufficient objectivity and accuracy of scoring results. It achieves personalized adaptive verification of nursing quality control, accurate multimodal data validation, and objective quantitative scoring, comprehensively improving the accuracy, adaptability, and reliability of nursing quality control verification.

[0155] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0156] Based on the same inventive concept, this application also provides an apparatus for implementing the above-mentioned intelligent verification and scoring method for nursing quality control. The solution provided by this apparatus is similar to the solution described in the above method; therefore, the specific limitations of one or more embodiments of the intelligent verification and scoring apparatus for nursing quality control provided below can be found in the above-described limitations of the intelligent verification and scoring method for nursing quality control, and will not be repeated here.

[0157] In one exemplary embodiment, such as Figure 3 As shown, a nursing quality control intelligent verification and scoring device 10 is provided, comprising:

[0158] The feature extraction module 11 can be used to obtain patient electronic medical record data based on bed identification information, and to perform joint entity relation extraction on the patient electronic medical record data to obtain a set of multidimensional pathological feature vectors of the patient.

[0159] The graph reconstruction module 12 can be used to adaptively reconstruct the static quality control rule tree based on the ontology knowledge graph of the nursing quality control field and the patient's multidimensional pathological feature vector set, and generate a customized quality control check graph with adaptive weight coefficients.

[0160] The cross-modal alignment module 13 can be used to obtain multimodal verification data streams based on customized quality control verification maps, and to perform cross-modal feature alignment on the multimodal verification data streams to obtain the aligned multimodal feature matrix;

[0161] The conflict detection module 14 can be used to perform semantic conflict detection on the aligned multimodal feature matrix to obtain the verification node data that has passed the verification.

[0162] The weighted scoring module 15 can be used to perform weighted aggregation scoring on the verification node data that has passed the verification based on the adaptive weight coefficients in the customized quality control verification map, and generate a global quality control scoring result.

[0163] In one embodiment, the map reconstruction module can be used to:

[0164] S21. Based on the ontology knowledge graph of the nursing quality control domain and the set of multidimensional pathological feature vectors of patients, a graph neural network is used to map the static quality control rule tree to obtain the initial quality control node feature matrix.

[0165] S22. Based on the topological connection relationship in the ontology knowledge graph of the nursing quality control domain, the feature matrix of the initial quality control node is aggregated with the features of neighboring nodes to obtain the updated quality control node feature matrix.

[0166] S23. Based on the updated quality control node feature matrix, the basic weights of each quality control item node in the static quality control rule tree are dynamically amplified and calculated to obtain the adaptive weight coefficients of each quality control item node.

[0167] S24. Based on the adaptive weight coefficients and preset dynamic thresholds of each quality control project node, the static quality control rule tree is topologically sorted and reorganized to generate a customized quality control verification map carrying adaptive weight coefficients.

[0168] In one embodiment, the map reconstruction module can also be used for:

[0169] S231. Based on the updated quality control node feature matrix, calculate the similarity between each feature entity and each quality control item node in the patient's multidimensional pathological feature vector set to obtain the semantic similarity between each feature entity and each quality control item node.

[0170] S232. Calculate the attention weight allocation for each feature entity and each quality control item node to obtain the cross-attention score of each feature entity to each quality control item node.

[0171] S233. Based on semantic similarity and cross-attention scores, the basic weights of each quality control item node in the static quality control rule tree are dynamically amplified and calculated to obtain the adaptive weight coefficients of each quality control item node.

[0172] In one embodiment, the cross-modal alignment module can be used to:

[0173] S31. Perform multimodal feature extraction on the multimodal verification data stream to obtain a multimodal basic feature set, which includes text feature vectors, visual feature vectors, and standard semantic anchor vectors;

[0174] S32. Based on the modal variance of text feature vectors and visual feature vectors, perform inverse variance weighted calculation on text feature vectors and visual feature vectors to obtain a balanced feature vector set;

[0175] S33. Perform multi-head cross-attention mapping on the balanced feature vector set to obtain the initial attention feature matrix;

[0176] S34. Perform residual connection calculation based on the initial attention feature matrix and standard semantic anchor vector to obtain the residual feature matrix;

[0177] S35. Perform layer normalization calculation on the residual feature matrix to obtain the aligned multimodal feature matrix.

[0178] In one embodiment, the collision detection module can be used to:

[0179] S41. Decouple the aligned multimodal feature matrix to obtain the aligned text feature vector, the aligned visual feature vector, and the aligned standard semantic anchor vector.

[0180] S42. Based on the aligned standard semantic anchor vector, perform deviation feature mapping on the aligned text feature vector to obtain the text semantic deviation vector.

[0181] S43. Based on the aligned standard semantic anchor vector, perform deviation feature mapping on the aligned visual feature vector to obtain the visual semantic deviation vector.

[0182] S44. Perform fusion weight calculation on the text semantic deviation vector and the visual semantic deviation feature to obtain the multimodal conflict probability value;

[0183] S45. Based on the multimodal conflict probability value and the preset adaptive conflict judgment threshold, the aligned multimodal feature matrix is ​​frozen by state marking to obtain the verification node data that has passed the verification.

[0184] In one embodiment, the weighted scoring module can be used to:

[0185] S51. Based on the verification node data that has passed the verification, extract the multimodal conflict probability value corresponding to each verification node to obtain a set of conflict probability values.

[0186] S52. Based on the numerical range of each element in the conflict probability value set, perform confidence penalty factor matching on each verification node to obtain the confidence penalty coefficient of each verification node.

[0187] S53. Based on the confidence penalty coefficient of each verification node, perform a second decay product calculation on the adaptive weight coefficient in the customized quality control verification map to obtain the final effective weight of each verification node.

[0188] S54. Based on the final effective weight of each verification node, perform weighted aggregation scoring on the verification node data that has passed the verification, and generate a global quality control score result.

[0189] In one embodiment, the weighted scoring module can also be used for:

[0190] S531. Construct a diagonal penalty matrix based on the confidence penalty coefficient of each verification node to obtain the node-level penalty tensor;

[0191] S532. The adaptive weight coefficients in the customized quality control verification map are arranged in a vectorized manner to obtain the basic weight vector;

[0192] S533. Perform Hadamard product calculation on the node-level penalty tensor and the basic weight vector to obtain the initial effective weight vector.

[0193] S534. Extract the extreme values ​​of the initial effective weight vector to obtain the maximum and minimum weight values ​​of the initial effective weight vector;

[0194] S535. Based on the maximum weight value of the vector, the minimum weight value of the vector, and the preset normalization interval, the range linear scaling calculation is performed on the initial effective weight vector to obtain the final effective weight of each verification node.

[0195] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the nursing quality control intelligent verification and scoring method as described above.

[0196] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the nursing quality control intelligent verification and scoring method as described above.

[0197] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0198] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A nursing quality control intelligent verification and scoring method, characterized in that, The method includes: S1. Obtain patient electronic medical record data based on bed identification information, and perform entity relationship joint extraction on the patient electronic medical record data to obtain a set of patient multidimensional pathological feature vectors. S2. Based on the ontology knowledge graph of the nursing quality control domain and the patient's multidimensional pathological feature vector set, the static quality control rule tree is reconstructed in an adaptive manner to generate a customized quality control check graph with adaptive weight coefficients. S3. Based on the customized quality control verification map, obtain the multimodal verification data stream, and perform cross-modal feature alignment on the multimodal verification data stream to obtain the aligned multimodal feature matrix; S4. Perform semantic conflict detection on the aligned multimodal feature matrix to obtain the verification node data that has passed the verification. S5. Based on the adaptive weight coefficients in the customized quality control verification map, the verification node data that has passed the verification are weighted and aggregated for scoring to generate a global quality control score result.

2. The method according to claim 1, characterized in that, S2 includes: S21. Based on the ontology knowledge graph of the nursing quality control domain and the patient's multidimensional pathological feature vector set, perform graph neural network mapping on the static quality control rule tree to obtain the initial quality control node feature matrix; S22. Based on the topological connection relationship in the ontology knowledge graph of the nursing quality control domain, the initial quality control node feature matrix is ​​aggregated with neighbor node features to obtain the updated quality control node feature matrix. S23. Based on the updated quality control node feature matrix, the basic weights of each quality control item node in the static quality control rule tree are dynamically amplified and calculated to obtain the adaptive weight coefficients of each quality control item node. The formula for calculating the adaptive weight coefficients is as follows: In the formula, For the first Adaptive weighting coefficients for each quality control project node. For the first The basic weight of each quality control project node. The total number of feature entities in the patient's multidimensional pathological feature vector set. For the first Each feature entity vector For the first Vector representation of each quality control project node The cosine similarity function is used. For cross-attention calculation function, The penalty factor is amplified for weighting; S24. Based on the adaptive weight coefficients and preset dynamic thresholds of each quality control project node, the static quality control rule tree is topologically sorted and reorganized to generate the customized quality control verification map carrying adaptive weight coefficients.

3. The method according to claim 2, characterized in that, S23 includes: S231. Based on the updated quality control node feature matrix, perform similarity calculation on each feature entity and each quality control item node in the patient multidimensional pathological feature vector set to obtain the semantic similarity between each feature entity and each quality control item node. S232. Perform attention weight allocation calculation on each of the feature entities and each of the quality control project nodes to obtain the cross-attention score of each feature entity to each of the quality control project nodes; S233. Based on the semantic similarity and the cross-attention score, the basic weights of each quality control item node in the static quality control rule tree are dynamically amplified and calculated to obtain the adaptive weight coefficients of each quality control item node.

4. The method according to claim 1, characterized in that, The S3 includes: S31. Perform multimodal feature extraction on the multimodal verification data stream to obtain a multimodal basic feature set, wherein the multimodal basic feature set includes text feature vectors, visual feature vectors, and standard semantic anchor vectors; S32. Based on the modal variance of the text feature vector and the visual feature vector, perform inverse variance weighted calculation on the text feature vector and the visual feature vector to obtain a balanced feature vector set; S33. Perform multi-head cross-attention mapping on the balanced feature vector set to obtain the initial attention feature matrix; S34. Perform residual connection calculation based on the initial attention feature matrix and the standard semantic anchor vector to obtain the residual feature matrix; S35. Perform layer normalization calculation on the residual feature matrix to obtain the aligned multimodal feature matrix.

5. The method according to claim 4, characterized in that, The S4 includes: S41. Decouple the aligned multimodal feature matrix to obtain the aligned text feature vector, the aligned visual feature vector, and the aligned standard semantic anchor vector. S42. Based on the aligned standard semantic anchor vector, perform deviation feature mapping calculation on the aligned text feature vector to obtain the text semantic deviation vector. S43. Based on the aligned standard semantic anchor vector, perform deviation feature mapping calculation on the aligned visual feature vector to obtain the visual semantic deviation vector. S44. Perform fusion weight calculation on the text semantic deviation vector and the visual semantic deviation feature to obtain the multimodal conflict probability value. The formula for calculating the multimodal conflict probability value is as follows: In the formula, The multimodal conflict probability value is... The aligned standard semantic anchor vector. The aligned text feature vector. The aligned visual feature vector, and It is a linear transformation matrix. and For bias terms, For activation function, This indicates a feature concatenation operation. To fuse the weight matrix, To incorporate the bias term, Use the Sigmoid activation function; S45. Based on the multimodal conflict probability value and the preset adaptive conflict determination threshold, the aligned multimodal feature matrix is ​​frozen with state marking to obtain the verification node data that has passed the verification.

6. The method according to claim 5, characterized in that, The S5 includes: S51. Based on the verified node data, extract the multimodal conflict probability value corresponding to each verification node to obtain a set of conflict probability values. S52. Based on the numerical range of each element in the set of conflict probability values, perform confidence penalty factor matching on each of the verification nodes to obtain the confidence penalty coefficient of each of the verification nodes. S53. Based on the confidence penalty coefficient of each of the verification nodes, perform a second decay product calculation on the adaptive weight coefficient in the customized quality control verification map to obtain the final effective weight of each of the verification nodes. S54. Based on the final effective weight of each verification node, the data of the verification nodes that have passed the verification are weighted and aggregated to generate a global quality control score result. The calculation formula for the global quality control score result is as follows: In the formula, The global quality control score result is as follows. This represents the total number of nodes actually verified. For the first The adaptive weight coefficients of each verification node For the first The confidence penalty coefficient for each verification node, For the first The structured checklist status codes for each verification node For the first The historical multimodal conflict probability value corresponding to each verification node. This is the node score mapping function.

7. The method according to claim 6, characterized in that, S53 includes: S531. Construct a diagonal penalty matrix based on the confidence penalty coefficient of each of the verification nodes to obtain a node-level penalty tensor; S532. The adaptive weight coefficients in the customized quality control verification map are arranged in a vectorized manner to obtain the basic weight vector; S533. Perform the Hadamard product calculation on the node-level penalty tensor and the basic weight vector to obtain the initial effective weight vector. S534. Extract the extreme values ​​of the initial effective weight vector to obtain the maximum and minimum weight values ​​of the initial effective weight vector; S535. Based on the maximum weight value of the vector, the minimum weight value of the vector, and the preset normalization interval, the initial effective weight vector is subjected to range linear scaling calculation to obtain the final effective weight of each of the verification nodes.

8. A nursing quality control intelligent verification and scoring device, characterized in that, The device includes: The feature extraction module is used to obtain patient electronic medical record data based on bed identification information, and to perform entity relation joint extraction on the patient electronic medical record data to obtain a set of multidimensional pathological feature vectors of the patient. The graph reconstruction module is used to adaptively reconstruct the static quality control rule tree based on the ontology knowledge graph of the nursing quality control field and the patient's multidimensional pathological feature vector set, and generate a customized quality control check graph with adaptive weight coefficients. The cross-modal alignment module is used to obtain a multimodal verification data stream based on the customized quality control verification map, and to perform cross-modal feature alignment on the multimodal verification data stream to obtain an aligned multimodal feature matrix; The conflict detection module is used to perform semantic conflict detection on the aligned multimodal feature matrix to obtain the verification node data that has passed the verification. The weighted scoring module is used to perform weighted aggregation scoring on the verified node data based on the adaptive weight coefficients in the customized quality control verification map, and generate a global quality control score result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images