Medical asset management big data analysis method and system

Abstract

The present application relates to the technical field of big data analysis, and discloses a medical asset management big data analysis method and system, the method comprising: preprocessing to obtain a standardized text data set; using a pre-trained language model for feature learning to obtain a fused feature vector; constructing a trained multi-task model; extracting information to obtain structured medical asset management information; establishing a multi-factor variance analysis model and a generalized additive model based on the structured medical asset management information, and performing a multi-factor analysis of medical equipment utilization efficiency to obtain a comprehensive influencing factor analysis result; performing optimal scheduling of medical resources based on the comprehensive influencing factor analysis result to obtain an optimal resource allocation plan, and then automatically extracting key information from massive unstructured text data, deeply analyzing the multi-dimensional factors that affect the utilization efficiency of medical equipment, and realizing intelligent scheduling of medical resources based on data-driven decision support.

Description

Technical Field

[0001] The present application relates to the field of big data analysis technology, and in particular to a medical asset management big data analysis method and system. Background Art

[0002] With the rapid development of medical technology and the continuous growth of medical needs, medical institutions are facing increasingly complex asset management challenges. Large hospitals have a large number of medical equipment and assets of various types. The effective management of these assets is directly related to medical quality, operational efficiency and economic benefits. However, traditional medical asset management methods often rely on manual operations and experience judgment, which is difficult to cope with massive data and complex decision-making needs.

[0003] In the era of big data, medical institutions have generated a large amount of asset management-related data, including equipment usage records, maintenance logs, procurement information, etc. These data contain rich valuable information, but due to the lack of unified data formats, scattered information, and lack of effective analysis tools, it is difficult to fully explore and utilize them. In addition, medical asset management involves multiple links. How to achieve the optimal allocation of resources while ensuring the quality of medical services is a complex multi-objective optimization problem. Summary of the invention

[0004] The present application provides a medical asset management big data analysis method and system.

[0005] The first aspect of the present application provides a medical asset management big data analysis method, the medical asset management big data analysis method comprising:

[0006] Preprocess the historical medical asset management text data to obtain a standardized text dataset;

[0007] Using a pre-trained language model to perform feature learning on the standardized text dataset, extracting context-related first-order features through a masked language model and a next sentence prediction task, and generating second-order fused features through nonlinear combination to obtain a fused feature vector;

[0008] Building a multi-task learning framework based on the fused feature vector, training named entity recognition and relation extraction tasks simultaneously, and obtaining a trained multi-task model;

[0009] Extracting information from target medical asset management text data based on the trained multi-task model to obtain structured medical asset management information;

[0010] A multi-factor variance analysis model and a generalized additive model are established based on the structured medical asset management information, and a multi-factor analysis of medical equipment utilization efficiency is performed to obtain comprehensive influencing factor analysis results;

[0011] Based on the analysis results of the comprehensive influencing factors, medical resources are optimized and scheduled to obtain the optimal resource allocation plan.

[0012] A second aspect of the present application provides a medical asset management big data analysis system, the medical asset management big data analysis system comprising:

[0013] A preprocessing module is used to preprocess the historical medical asset management text data to obtain a standardized text data set;

[0014] A feature learning module, used to perform feature learning on the standardized text dataset using a pre-trained language model, extract context-related first-order features through a masked language model and a next sentence prediction task, and generate second-order fused features through nonlinear combination to obtain a fused feature vector;

[0015] A training module, used to construct a multi-task learning framework according to the fused feature vector, and simultaneously train named entity recognition and relationship extraction tasks to obtain a trained multi-task model;

[0016] An extraction module, used to extract information from the target medical asset management text data based on the trained multi-task model to obtain structured medical asset management information;

[0017] An analysis module, used to establish a multi-factor variance analysis model and a generalized additive model based on the structured medical asset management information, and to perform a multi-factor analysis of the medical equipment utilization efficiency to obtain a comprehensive influencing factor analysis result;

[0018] The scheduling module is used to optimize the scheduling of medical resources based on the analysis results of the comprehensive influencing factors to obtain the optimal resource allocation plan.

[0019] The third aspect of the present application provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the electronic device executes the above-mentioned medical asset management big data analysis method.

[0020] The fourth aspect of the present application provides a computer-readable storage medium, which stores instructions. When the computer-readable storage medium is run on a computer, it enables the computer to execute the above-mentioned medical asset management big data analysis method.

[0021] Compared with the prior art, the present application has the following beneficial effects: standardized processing of historical medical asset management text data is achieved through preprocessing. Deep feature learning: feature learning is performed using a pre-trained language model, context-related features are extracted through a masked language model and the next sentence prediction task, and fusion features are generated through nonlinear combination. It can capture deep semantic information in the text and improve the ability to understand content related to medical asset management. A multi-task learning framework for named entity recognition and relationship extraction tasks is constructed. It can share underlying feature representations, improve the generalization ability of the model, and at the same time improve the comprehensiveness and accuracy of medical asset management information extraction. Through the multi-factor variance analysis model and the generalized additive model, an in-depth analysis of the factors affecting the efficiency of medical equipment use is achieved. Not only can the main influencing factors be identified, but also the interaction effects and nonlinear relationships between factors can be captured. Based on the results of the comprehensive influencing factor analysis, a resource scheduling optimization model considering multiple objectives is constructed. Through local search optimization and multi-objective evaluation, the optimal resource allocation plan that takes into account resource utilization, scheduling cost and service quality can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0023] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0024] Figure 1 It is a flowchart of a medical asset management big data analysis method provided by an embodiment of the present invention;

[0025] Figure 2 It is a schematic block diagram of the structure of the medical asset management big data analysis system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0028] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0029] It should be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 , an embodiment of the medical asset management big data analysis method in the embodiment of the present application includes:

[0030] Step 100: preprocess the historical medical asset management text data to obtain a standardized text data set;

[0031] It is understandable that the execution subject of the present application may be a medical asset management big data analysis system, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0032] Specifically, the historical medical asset management text data is segmented, and the continuous text is split into meaningful words or phrases to obtain the initial segmentation results. Based on the predefined medical asset management field dictionary, the initial segmentation results are identified with professional terms, and the segmentation results of professional terms are identified and optimized to ensure that the field-specific vocabulary can be correctly processed during the segmentation process to obtain the optimized segmentation results. Based on the optimized segmentation results, a word frequency-inverse document frequency matrix is ​​constructed to quantify the importance of the words, and then an appropriate threshold is set to filter out common but meaningless stop words in text processing, and a stop word list is generated accordingly. Based on the generated stop word list, the optimized segmentation results are processed to reduce noise by removing stop words to obtain more refined text data. The numbers and unit representations in the stop word-removed text are standardized, and different forms of numerical values ​​and units are uniformly converted into a standardized format to obtain a standardized text. Using the pre-built synonym library in the medical asset management field, synonym replacement is performed on the standardized text, and the same concept in different expressions is unified to ensure the semantic consistency of the text data and obtain a semantically unified text. The word vector model is used to vectorize the semantically unified text, convert the text into a numerical form that can be processed by the machine, and obtain the initial word vector. In order to reduce the data dimension while retaining important information, the initial word vector is mapped to a low-dimensional space to reduce the amount of calculation and improve the efficiency of the model, and finally a standardized text dataset is obtained.

[0033] Step 200: Use a pre-trained language model to perform feature learning on a standardized text dataset, extract context-related first-order features through a masked language model and a next sentence prediction task, and generate second-order fused features through nonlinear combination to obtain a fused feature vector;

[0034] Specifically, the standardized text dataset is segmented to obtain input sequences suitable for the input model. The length of each input sequence does not exceed 512 tokens to ensure that the model does not exceed the computing resource limit when processing. The input sequence is processed by superposition of tag embedding, position embedding and segment embedding, where tag embedding is used to represent the lexical meaning of each word or token, position embedding is used to capture the position relationship of words in the sequence, and segment embedding is used to distinguish different sentences or paragraphs to obtain the initial input embedding. The initial input embedding is input into the encoder of the pre-trained BERT model for processing. The encoder consists of 12 layers of Transformer layers, each of which contains a multi-head self-attention sublayer and a feedforward neural network sublayer. The multi-head self-attention sublayer uses 8 attention heads, and the dimension of each attention head is 96 to ensure that the model can capture the correlation and contextual relationship between different tokens in the input sequence from multiple angles. The feedforward neural network sublayer includes two linear transformations, and the GELU activation function is used between the linear transformations for nonlinear transformation to enhance the model's ability to recognize complex patterns. In each Transformer layer of the encoder, the attention weight is calculated through the multi-head self-attention mechanism, and the dependency between each token and other tokens is captured by weight distribution to obtain the context-related intermediate representation. The context-related intermediate representation is nonlinearly transformed through a feedforward neural network to obtain the output representation of each layer, which will be passed as input to the next Transformer layer. The vector corresponding to the token is extracted from the output of the last Transformer layer of the encoder and used as the first-order feature representation of the entire sequence to effectively summarize the semantic information in the input sequence. The masked language model task and the next sentence prediction task are performed based on the first-order feature representation. The specific operation of the masked language model task is to randomly mask 15% of the input tokens, and then predict these masked tokens to learn the dependency within the sequence and calculate the reconstruction loss. At the same time, the next sentence prediction task performs a binary classification task by judging whether two sentences are continuous, and the binary classification loss is calculated. The reconstruction loss and the binary classification loss are weighted and summed to form the overall loss function, and the model parameters are updated through the back-propagation algorithm to ensure the effectiveness and accuracy of the model in feature learning. The output representations of the last three Transformer layers of the encoder are nonlinearly combined, and the output representations of the last three layers are fused through a fully connected layer and a ReLU activation function to obtain a fused feature vector.

[0035] The context-related intermediate representation is layer-normalized to standardize the distribution of the intermediate representation, eliminate the scale differences of the data in different dimensions, and obtain the normalized intermediate representation. Normalization helps stabilize the training process of the model and accelerates the convergence speed. The normalized intermediate representation is input into a two-layer fully connected neural network for processing. In this neural network, the number of neurons in the first layer is 3072. This layer expands the dimension of the feature through linear transformation to capture more complex and high-order feature relationships. The output of the first layer is passed to the second layer of the fully connected neural network, which has 768 neurons. The linear transformation of this layer maps the feature back to the original dimension to obtain the intermediate feature. The GELU activation function is applied to perform nonlinear transformation on the intermediate feature to obtain the nonlinear activation feature. The GELU activation function can enhance the model's ability to recognize complex patterns while retaining the linear information of the feature. In order to prevent the model from overfitting, the nonlinear activation feature is dropout processed, and the dropout rate is set to 0.1, that is, 10% of the neuron output is randomly discarded to increase the robustness of the model and obtain the regularized feature. The regularized features are input into the output layer, where the number of neurons is 768, matching the feature dimension of the input. After this layer of processing, the original output representation is obtained. In order to improve the stability and expressiveness of the feature representation, the original output representation is processed by residual connection, and the original output representation is added to the normalized intermediate representation of the input to obtain the residual output representation. Residual connection can effectively alleviate the gradient vanishing problem in deep neural networks by directly adding input and output, while retaining the original information of the input features. The residual output representation is subjected to a second layer normalization process to standardize the data distribution, ensure the consistency of the output representation of each layer, reduce the internal covariate shift, and obtain the output representation of each layer. From the output of the last layer of the encoder, the 768-dimensional vector corresponding to each tag is extracted. In this way, the semantic information of each tag is fully represented. The 768-dimensional vectors are concatenated into a matrix to form the first-order feature representation of the entire sequence.

[0036] Step 300: construct a multi-task learning framework based on the fused feature vector, and train named entity recognition and relationship extraction tasks at the same time to obtain a trained multi-task model;

[0037] It should be noted that the fusion feature vector is transformed in dimension and its dimension is adjusted to 512 through the fully connected layer to obtain the shared feature representation of the historical text, which effectively reduces the dimension of the feature vector. Based on the shared feature representation of the historical text, a bidirectional long short-term memory network (BiLSTM) is constructed. The network contains a 3-layer BiLSTM structure, and each BiLSTM structure contains 256 hidden units. The bidirectional processing method of BiLSTM can capture the contextual information in the text sequence, so that the model can better understand the contextual semantics of the text and obtain the sequence encoding representation of the historical text. The self-attention mechanism is applied to the sequence encoding representation of the historical text to enhance the model's perception of the internal relationship of the text. Eight attention heads are used in the self-attention mechanism, and the dimension of each attention head is 64, so that the model can analyze the information in the text from different angles and obtain the context-aware representation of the historical text. The context-aware representation can effectively summarize and represent important information in the text. The context-aware representation of the historical text is input into the specific layer of the named entity recognition task, which contains two layers of fully connected networks. The number of neurons in the first layer is 256, and the number of neurons in the second layer is equal to the number of entity categories. Through the processing of these layers, the entity recognition logits of the historical text are obtained. The conditional random field (CRF) layer is applied to the entity recognition logits. Through the optimization of the CRF layer, the optimal annotation sequence is calculated to obtain the named entity recognition results of the historical text. Based on the obtained named entity recognition results and context-aware representation, the entity pair representation is constructed. Through concatenation and pooling operations, this information is converted into the first entity pair vector of fixed dimension. The entity pair vector is input into the specific layer of the relationship extraction task. This layer contains three layers of convolutional neural network. The convolution kernel size of each layer is set to 3, 4, and 5 respectively, and the number of convolution kernels is 128. Through these convolution operations, text features of different scales can be extracted to form multi-scale feature maps. The multi-scale feature maps are subjected to maximum pooling operations to reduce the dimensionality of the data and retain the most important feature information. The pooled results are concatenated to obtain the relationship representation vector. The relation representation vector is input into the fully connected layer, and the number of neurons in the last layer is set to the number of relation categories. Through this process, the relation classification logits of the historical text are obtained. The logits are used to determine the relationship between entities in the text. In order to optimize the training process of the entire model, a joint loss function is constructed based on the entity recognition logits and relation classification logits of the historical text, and the Adam optimizer is used for parameter update. The Adam optimizer, with its efficient computing and adaptive learning rate, ensures that the model can converge quickly and stably, and obtains the final trained multi-task model.

[0038] Step 400: extract information from target medical asset management text data based on the trained multi-task model to obtain structured medical asset management information;

[0039] Specifically, the target medical asset management text data is preprocessed, including word segmentation, stop word removal, and standardization. Through word segmentation, the text is decomposed into words or phrases; through stop word removal, common words with less semantic contribution are deleted to reduce data noise; standardization unifies different forms of expression to ensure data consistency and processability, and obtains preprocessed target text data. Based on the preprocessed target text data, the pre-trained language model is used to extract features and extract high-quality fused feature vectors from the text. The fused feature vector of the target text is transformed in dimension, and its dimension is adjusted to 512 through the fully connected layer to obtain the shared feature representation of the target text. Based on the shared feature representation, the trained bidirectional long short-term memory network (BiLSTM) is used to sequence encode the target text. Through the BiLSTM structure, the contextual information in the text sequence is captured to obtain the sequence encoding representation of the target text. The self-attention mechanism is applied to the sequence encoding representation, and the context-aware representation of the target text is obtained by calculating the attention weights of different positions. The context-aware representation of the target text is input into the trained named entity recognition task-specific layer, where the model generates entity recognition logits for the target text. These logits are processed by the conditional random field (CRF) layer to calculate the optimal annotation sequence and obtain the named entity recognition results of the target text, which include the types and boundaries of different entities in the text. Based on the context-aware representation of the target text and the above named entity recognition results, an entity pair representation is constructed. Through concatenation and pooling operations, this information is processed into a second entity pair vector of fixed dimension. These vectors effectively summarize the semantic relationship between each pair of entities, and the second entity pair vector is input into the trained relation extraction task-specific layer. After processing, the relation classification logits of the target text are generated. These logits are used to represent the relationship types between different entities in the text. Based on the named entity recognition results and relation classification logits of the target text, structured medical asset management information is constructed. The structured information includes the type of entity, the boundary position of the entity in the text, and the relationship type between entities.

[0040] Step 500: Establish a multi-factor variance analysis model and a generalized additive model based on the structured medical asset management information, and perform a multi-factor analysis of medical equipment utilization efficiency to obtain a comprehensive influencing factor analysis result;

[0041] Specifically, feature extraction is performed on structured medical asset management information to extract key factors such as equipment location, implementation of operating specifications, equipment load, and use of intelligent auxiliary systems to obtain the initial feature set. Data cleaning and standardization are performed on the initial feature set to ensure data quality and consistency. This includes missing value filling to deal with possible gaps in the data; outlier processing to exclude extreme values ​​that may interfere with the analysis results; and numerical normalization so that data of different dimensions can be processed under the same analysis framework to obtain the preprocessed feature data set. Based on the preprocessed feature data set, a multi-factor variance analysis model is constructed. In this model, the medical equipment utilization efficiency is taken as the dependent variable. By calculating the main effect and interaction effect of each factor, the independent and interactive effects of these factors on the equipment utilization efficiency are analyzed to obtain the multi-factor variance analysis results. Variance analysis can reveal the degree of influence of each factor and identify those main effect factors and interaction effect factors that have a significant impact on the equipment utilization efficiency. Hypothesis testing is performed on the results of multi-factor variance analysis, and the F test method is used to calculate the significance level of the influence of each factor to obtain a set of significant factors. Based on the set of significant factors, a generalized additive model was constructed, and the spline function was used to fit the nonlinear relationship between each factor and the equipment utilization efficiency to obtain a nonparametric regression model. The generalized additive model can more accurately capture the impact of each factor on the equipment utilization efficiency by fitting the nonlinear relationship. Model diagnosis was performed on the nonparametric regression model, including residual analysis to check the fitting effect of the model; influence point detection to identify data points that may have a disproportionate impact on the model; and multicollinearity test to ensure that there is no excessive correlation between the variables in the model. Through these diagnoses, model optimization suggestions were obtained. According to the results of model diagnosis, the generalized additive model was adjusted, including variable selection to exclude irrelevant or redundant variables; smoothing parameter optimization to better fit the nonlinear trend of the data; and the addition of interaction terms to capture the complex relationship between multiple factors. Through adjustment, the optimized generalized additive model was obtained to reflect the relationship between the medical equipment utilization efficiency and each factor. Based on the optimized generalized additive model, the marginal effect and relative importance of each factor on the medical equipment utilization efficiency were calculated, and combined with the results of multifactor variance analysis, the comprehensive influencing factor analysis results were obtained. The comprehensive analysis results can fully reveal the role of various factors in the efficiency of medical equipment use and guide the optimal allocation of medical resources and the formulation of management strategies.

[0042] The preprocessed feature data set is factor-encoded, and the categorical variables are converted into dummy variables. The categorical variables can be expressed in numerical form, and the coded feature matrix suitable for matrix operations is obtained. Based on the coded feature matrix, a full factorial design matrix is ​​constructed. The full factorial design matrix includes the main effect items of each factor, as well as the interaction effect items between the factors, which comprehensively reflects the interaction between the factors and obtains the design matrix. The design matrix is ​​QR decomposed to orthogonalize the design matrix and simplify the subsequent calculation process. The result of QR decomposition is an orthogonalized design matrix and an upper triangular matrix. The orthogonalized design matrix can reduce the multicollinearity problem in the model, improve the calculation efficiency and the stability of the model. Based on the orthogonalized design matrix, the sum of squares of each effect is calculated. First, the sum of squares of each main effect item and interaction effect item is calculated. These sums of squares represent the contribution of each effect to the efficiency of medical equipment use. By accumulating these sums of squares, the effect sum of square vector is obtained, which summarizes the contribution of each factor. According to the effect sum of square vector and the degrees of freedom of each effect, the mean square value is calculated to obtain the effect mean square vector. The mean square value is the ratio of the sum of squares of the effect to its corresponding degrees of freedom, which can reflect the average contribution of each effect. Based on the effect mean square and the error mean square, the F statistic is calculated to obtain the F value of each effect. The F value is used to determine whether each effect is statistically significant. A significance test is performed on the F value of each effect, and the significance of each effect is evaluated by calculating the p value. By comparing with the preset significance level, statistically significant effects are screened out to obtain a set of significant effects. Based on the set of significant effects, the variance explanation ratio and effect size of each significant effect are calculated to obtain the results of multifactor analysis of variance. The variance explanation ratio indicates the contribution of each significant effect in explaining the variance of the dependent variable, while the effect size provides a quantification of the impact of the effect.

[0043] Step 600: Optimize the scheduling of medical resources based on the comprehensive influencing factor analysis results to obtain the optimal resource allocation plan.

[0044] Specifically, the factor weights are calculated for the comprehensive influencing factor analysis results, and the factor weight vector is obtained by quantitatively analyzing the importance of each influencing factor. The weight vector reflects the relative importance of different factors in resource optimization. Based on the factor weight vector, the existing medical resource allocation status is evaluated. In this process, the comprehensive utility score of each resource unit is calculated, and these scores represent the effectiveness of each resource unit under the current configuration. By summarizing the scores, a resource utility evaluation matrix is ​​formed. Cluster analysis is performed on the resource utility evaluation matrix, and the resource units are divided into three categories: high efficiency, medium efficiency, and low efficiency, and the resource efficiency classification results are obtained. Based on the resource efficiency classification results, a resource scheduling optimization model is constructed. In the model, maximizing resource utilization is set as the objective function, and resource capacity and demand matching are used as constraints to ensure that resource scheduling can maximize utility and meet the constraints of actual needs. The initial optimization model is solved to generate a set of candidate resource allocation solutions. The candidate solution is an exploration of different solutions to the resource allocation problem. In order to improve the quality of the solution, local search optimization is performed based on the candidate solution set, and the optimized resource allocation solution set is obtained by refining the local solution. Find the optimal solution through continuous adjustment and optimization. Perform multi-objective evaluation on the optimized resource allocation scheme set. During the evaluation process, the three key dimensions of resource utilization, scheduling cost and service quality indicators are considered to calculate the comprehensive score of each scheme. These scores are weighted and scored to comprehensively reflect the performance of each scheme under multiple objectives and obtain the scheme ranking results. According to the scheme ranking results, the scheme with the highest comprehensive score is selected as the optimal resource allocation scheme. The optimal resource allocation scheme should include the optimal location arrangement, staffing plan and usage time arrangement of each medical equipment, so as to achieve the optimal scheduling and utilization of medical resources in actual operation.

[0045] In the embodiment of the present application, the standardized processing of historical medical asset management text data is achieved through preprocessing. Deep feature learning: feature learning is performed using a pre-trained language model, context-related features are extracted through a masked language model and the next sentence prediction task, and fusion features are generated through nonlinear combination. It can capture deep semantic information in the text and improve the ability to understand the content related to medical asset management. A multi-task learning framework for named entity recognition and relationship extraction tasks is constructed. It can share the underlying feature representation, improve the generalization ability of the model, and at the same time improve the comprehensiveness and accuracy of medical asset management information extraction. Through the multi-factor variance analysis model and the generalized additive model, an in-depth analysis of the factors affecting the efficiency of medical equipment use is achieved. Not only can the main influencing factors be identified, but also the interaction effects and nonlinear relationships between factors can be captured. Based on the comprehensive influencing factor analysis results, a resource scheduling optimization model considering multiple objectives is constructed. Through local search optimization and multi-objective evaluation, the optimal resource allocation plan that takes into account resource utilization, scheduling cost and service quality can be obtained.

[0046] In a specific embodiment, the process of executing step 100 may specifically include the following steps:

[0047] Perform word segmentation processing on historical medical asset management text data to obtain initial word segmentation results;

[0048] Based on the predefined medical asset management dictionary, professional terminology recognition is performed on the initial word segmentation results to obtain optimized word segmentation results;

[0049] Build a word frequency-inverse document frequency matrix based on the optimized word segmentation results, set a threshold to filter stop words, and obtain a stop word list;

[0050] The optimized word segmentation results are filtered based on the stop word list to obtain the text without stop words, and the numbers and unit expressions in the text without stop words are standardized to obtain the normalized text;

[0051] Based on the pre-built synonym database in the field of medical asset management, synonyms are replaced in the standardized text to obtain semantically unified text;

[0052] The word vector model is used to vectorize the semantically unified text to obtain the initial word vector, and the initial word vector is mapped to a low-dimensional space to obtain a standardized text dataset.

[0053] Specifically, the historical medical asset management text data is segmented to decompose the continuous text stream into smaller units, namely words or phrases, to better capture the information in the text and obtain the initial segmentation results. Based on the predefined medical asset management field dictionary, the segmentation results are identified with professional terms to obtain optimized segmentation results. Based on the optimized segmentation results, a term frequency-inverse document frequency (TF-IDF) matrix is ​​constructed. TF-IDF is a commonly used text representation method that is used to measure the importance of a word in a specific document. TF (Term Frequency) indicates the frequency of a word appearing in a document, while IDF (Inverse Document Frequency) is used to measure the prevalence of a word in all documents. The TF-IDF value can be calculated using the formula:

[0054] ;

[0055] Among them, TF Expressing words In the documentation The frequency of occurrence in , while IDF It's a word The inverse document frequency of is defined as:

[0056] ;

[0057] in, is the total number of documents, DF Indicates that it contains words The number of documents in the medical asset management text is 1. The TF-IDF matrix is ​​used to identify words with high importance in the medical asset management text, and then a threshold is set to filter out common stop words. The stop word screening process aims to remove words that contribute less to text analysis and generate a stop word list. Based on the generated stop word list, the optimized word segmentation results are filtered to remove these stop words to obtain a more concise text representation. The numbers and unit representations in the text without stop words are standardized to ensure the consistency and comparability of the data and obtain a normalized text. Based on the pre-built synonym library in the field of medical asset management, the normalized text is replaced with synonyms to unify different expressions and maintain consistency at the semantic level. Synonym replacement can ensure that the same concept in the text is expressed consistently and obtain a semantically unified text. The word vector model is used to vectorize the semantically unified text to obtain the initial word vector. Word vector models (such as Word2Vec, GloVe, etc.) map words to a low-dimensional vector space so that words with similar semantics are closer in the vector space. Assume that the Word2Vec model is used. Through training, a fixed-dimensional vector, such as 100 or 300 dimensions, is generated for each word, so that each word in the text can be represented by a vector to obtain the initial word vector. The initial word vector is mapped to a low-dimensional space. For example, through dimensionality reduction techniques such as principal component analysis or t-SNE, the high-dimensional word vector is reduced to a lower dimension, such as 50 or 20 dimensions. Through this dimensionality reduction operation, the redundancy of features can be reduced and the most important semantic information can be retained to obtain the final standardized text dataset.

[0058] In a specific embodiment, the process of executing step 200 may specifically include the following steps:

[0059] Segment the standardized text dataset to obtain input sequences, each of which has a length no greater than 512 tokens.

[0060] Perform a superposition process of token embedding, position embedding and segment embedding on the input sequence to obtain the initial input embedding;

[0061] The initial input is embedded into the encoder of the pre-trained BERT model. The encoder consists of 12 Transformer layers. Each Transformer layer contains a multi-head self-attention sublayer and a feed-forward neural network sublayer. The multi-head self-attention sublayer uses 8 attention heads, and the dimension of each attention head is 96. The feed-forward neural network sublayer contains two linear transformations, and the intermediate activation function is GELU.

[0062] In each Transformer layer of the encoder, the attention weights are calculated through the multi-head self-attention mechanism to obtain the context-dependent intermediate representation;

[0063] The context-related intermediate representation is nonlinearly transformed through a feedforward neural network to obtain the output representation of each layer, and the vector corresponding to the tag is extracted from the output of the last Transformer layer of the encoder as the first-order feature representation of the entire sequence;

[0064] Based on the first-order feature representation, the masked 15% input tokens are predicted through the masked language model task to obtain the reconstruction loss. Based on the first-order feature representation, the next sentence prediction task is used to determine whether two sentences are continuous, resulting in the binary classification loss.

[0065] The reconstruction loss and the binary classification loss are weighted summed to obtain the overall loss function, and the model parameters are updated through back propagation;

[0066] The output representations of the last three Transformer layers of the encoder are nonlinearly combined. Specifically, the output representations of the last three Transformer layers are fused through a fully connected layer and a ReLU activation function to obtain a fused feature vector.

[0067] Specifically, the standardized text dataset is segmented to ensure that the length of each input sequence does not exceed 512 tokens. The input sequence is processed by superposition of token embedding, position embedding, and segment embedding to obtain the initial input embedding. Token embedding maps each token to a fixed-dimensional vector that captures the semantic information of the vocabulary; position embedding preserves the order information of words in the sequence by adding vectors to each token position, so that the model can understand the relative position of words; segment embedding is used to distinguish tokens in different sentences or paragraphs. For example, in BERT, if the input sequence contains two sentences, these sentences will be labeled as sentence A and sentence B respectively, and represented by segment embedding vectors. These embeddings are superimposed to form the initial input embedding, which will be used as the input data of the model. The initial input embedding is input to the encoder of the pre-trained BERT model for processing. The encoder of the BERT model contains 12 Transformer layers, each of which consists of a multi-head self-attention sublayer and a feedforward neural network sublayer. In each Transformer layer, the multi-head self-attention mechanism generates attention weights by calculating the correlation between different tokens in the input sequence. Attention weights are used to capture the importance of each token in the context. The multi-head self-attention sublayer in the BERT model uses 8 attention heads, each with a dimension of 96, to process input data in parallel from different perspectives to improve the model's perception. After the multi-head self-attention mechanism calculates the attention weights, it generates context-dependent intermediate representations that contain the semantic information of each token in the input sequence. The context-dependent intermediate representations are nonlinearly transformed through a feedforward neural network. The feedforward neural network contains two linear transformation layers, with the GELU activation function used in the middle. GELU is more suitable for processing complex nonlinear relationships than the traditional ReLU (Rectified Linear Unit) and can better preserve semantic information. After nonlinear transformation, each Transformer layer outputs a new representation, which is further processed in subsequent layers. From the output of the last Transformer layer of the encoder, the vector corresponding to the token is extracted as the first-order feature representation of the entire sequence. The first-order feature representation summarizes the semantic information of each token in the input sequence and can be used as the basic input for downstream tasks. Based on the first-order feature representation, the masked language model task is used to predict 15% of the randomly masked input tokens. The core of the masked language model is to randomly mask some input tokens, and the model needs to predict the masked words based on the context. The loss function of this task is usually defined as the cross-entropy loss, which is used to calculate the difference between the predicted value and the actual value. Through this process, the reconstruction loss of MLM is obtained. The next sentence prediction task is performed based on the first-order feature representation to determine whether the two sentences are continuous, and a binary classification loss is generated, which is usually also calculated using the cross-entropy loss.The reconstruction loss of the MLM task and the binary classification loss of the NSP task are weighted and summed to obtain the overall loss function. The form of the loss function can usually be expressed as:

[0068] ;

[0069] in, and It is the weight coefficient in the loss function, which is used to balance the loss of the two tasks. Through the back propagation algorithm, the gradient information of the loss function is passed to the parameters of each layer of the model, and the model parameters are updated to gradually approach the optimal solution. The output representations of the last three layers of the encoder Transformer layer are nonlinearly combined. The output representations of the three layers are fused through the fully connected layer and the ReLU activation function to obtain the fused feature vector.

[0070] In a specific embodiment, the execution step performs a nonlinear transformation on the context-related intermediate representation through a feedforward neural network to obtain the output representation of each layer, and extracts the vector corresponding to the tag from the output of the last Transformer layer of the encoder as the first-order feature representation of the entire sequence, which may specifically include the following steps:

[0071] Perform layer normalization on the context-related intermediate representation to obtain a normalized intermediate representation;

[0072] The normalized intermediate representation is input into a two-layer fully connected neural network, with 3072 neurons in the first layer and 768 neurons in the second layer, to obtain the intermediate features;

[0073] Apply the GELU activation function to the intermediate features to obtain nonlinear activation features, and perform dropout processing on the nonlinear activation features with a dropout rate of 0.1 to obtain regularized features;

[0074] The regularized features are input into the output layer with 768 neurons to obtain the original output representation, and the original output representation is residually connected and added to the normalized intermediate representation of the input to obtain the residual output representation;

[0075] Perform a second layer normalization on the residual output representation to obtain the output representation of each layer;

[0076] From the output representation of each layer of the last Transformer layer of the encoder, the 768-dimensional vector corresponding to each token is extracted, and the 768-dimensional vectors are concatenated into a matrix to obtain the first-order feature representation of the entire sequence.

[0077] Specifically, the context-related intermediate representation is layer-normalized to obtain a normalized intermediate representation. Layer normalization is a standardization technique that normalizes the input data to keep the input distribution of each layer stable, which helps to improve the training efficiency and convergence speed of deep neural networks. The normalized intermediate representation is input into a two-layer fully connected neural network for processing. In the neural network, the number of neurons in the first layer is 3072, which aims to expand the dimension of the features, increase the complexity of the data through linear transformation, and capture higher-order feature information. Assume that the normalized intermediate representation of the input is , then the output of the first layer of fully connected neural network can be expressed as:

[0078] ;

[0079] in, is the weight matrix of the first layer of neural network, is the bias term, is the intermediate feature output by the first layer. The intermediate feature is passed to the second layer of fully connected neural network, which has 768 neurons, and the expanded features of the previous layer are remapped back to a lower dimension to obtain a more compact feature representation. This process can be expressed as:

[0080] ;

[0081] in, is the weight matrix of the second layer neural network, is the bias term, is the output of the second layer, which is represented as the final intermediate feature. Applying the GELU activation function to the intermediate feature, we get the nonlinear activation feature. The GELU activation function can perform nonlinear transformation on the input data while retaining the linear information. The formula is:

[0082] ;

[0083] Unlike the traditional ReLU activation function, the GELU activation function is more flexible in dealing with complex nonlinear relationships and can capture more delicate feature information. Through the processing of this activation function, nonlinear activation features are obtained. In order to prevent the model from overfitting, the nonlinear activation features are subjected to dropout processing, and the dropout rate is set to 0.1. During each training process, 10% of the neuron outputs will be randomly set to zero to increase the robustness of the model and obtain regularized features. The regularized features are input into the output layer for processing. The number of neurons in the output layer is 768, which is consistent with the feature dimension of the input. The regularized features are mapped to the required output dimension to obtain the original output representation. The residual connection is used to add the original output representation to the normalized intermediate representation of the input to obtain the residual output representation. The formula for the residual connection is as follows:

[0084] ;

[0085] in, is the residual output representation, It is a normalized intermediate representation. Through this connection, the model can retain the original input information while effectively utilizing the features learned by the deep network. This operation helps alleviate the gradient vanishing problem and improves the training effect of the deep network. The residual output representation is subjected to a second layer normalization process to standardize the distribution of the residual output so that the features of each layer output have the same scale, ensuring the stability and consistency of subsequent processing. The 768-dimensional vector corresponding to each token is extracted from each layer of the output representation of the last layer of the Transformer layer of the encoder. These vectors are high-dimensional representations of the semantic and contextual information of each token in the input sequence. The 768-dimensional vectors of all tokens are concatenated into a matrix to obtain the first-order feature representation of the entire sequence.

[0086] In a specific embodiment, the process of executing step 300 may specifically include the following steps:

[0087] The fusion feature vector is transformed in dimension and the dimension is adjusted to 512 through a fully connected layer to obtain the shared feature representation of the historical text.

[0088] Based on the shared feature representation of historical texts, a bidirectional long short-term memory network is constructed, which includes a 3-layer BiLSTM structure. Each BiLSTM structure contains 256 hidden units to obtain the sequence encoding representation of historical texts.

[0089] Apply the self-attention mechanism to the sequence encoding representation of historical texts, using 8 attention heads, each with a dimension of 64, to obtain a context-aware representation of historical texts;

[0090] The context-aware representation of the historical text is input into the named entity recognition task-specific layer, which contains a 2-layer fully connected network. The number of neurons in the first layer is 256, and the number of neurons in the second layer is equal to the number of entity categories, to obtain the entity recognition logits of the historical text;

[0091] Apply the conditional random field layer to the entity recognition logits of the historical text, calculate the optimal annotation sequence, and obtain the named entity recognition results of the historical text;

[0092] Based on the context-aware representation of historical texts and the named entity recognition results of historical texts, an entity pair representation is constructed, and a first entity pair vector of fixed dimension is obtained through concatenation and pooling operations;

[0093] The first entity pair vector is input into the relation extraction task-specific layer, which includes three layers of convolutional neural network. The convolution kernel sizes of each layer are 3, 4, and 5 respectively, and the number of convolution kernels is 128, to obtain a multi-scale feature map;

[0094] Perform the maximum pooling operation on the multi-scale feature maps and concatenate the pooling results to obtain the relationship representation vector, which is then input into the fully connected layer with the number of neurons equal to the number of relationship categories to obtain the relationship classification logits of the historical text.

[0095] A joint loss function is constructed based on the entity recognition logits of historical texts and the relationship classification logits of historical texts, and the Adam optimizer is used to update the parameters to obtain the trained multi-task model.

[0096] Specifically, the fusion feature vector is transformed in dimension and its dimension is adjusted to 512 through the fully connected layer to obtain the shared feature representation of the historical text. Assume that the input fusion feature vector is ,in is the original dimension, which is converted into a 512-dimensional vector through the linear transformation of the fully connected layer , the formula is as follows:

[0097] ;

[0098] in, is the weight matrix, is the bias vector. Based on the shared feature representation of historical text, a bidirectional long short-term memory network (BiLSTM) is constructed. The network contains a 3-layer BiLSTM structure, and each BiLSTM structure contains 256 hidden units. BiLSTM can capture the context information in the sequence and is suitable for processing data with time or order relationships. Assume that the shared feature representation of the input is , then after being processed by the BiLSTM network, the output can be expressed as a sequence encoding representation ,in is the time step The hidden state vector of is the length of the sequence. This sequence encoding representation captures the temporal dependencies in the text and can provide rich contextual information. The self-attention mechanism is applied to the sequence encoding representation of historical text to improve the model's ability to capture important information. The self-attention mechanism can calculate the attention weights of different positions in the sequence and highlight the most critical parts for the current task. In this process, 8 attention heads are used, and the dimension of each attention head is 64. Assume that the input is , the self-attention mechanism will generate context-aware representations ,in is the time step The context representation of is as follows:

[0099] ;

[0100] in, , K and are the query matrix, key matrix and value matrix obtained by linear transformation of H, is the dimension of the attention head (i.e. 64). The self-attention mechanism enables the model to focus on the context related to a specific task and generate a more accurate representation. The context-aware representation of the historical text is input into the named entity recognition task-specific layer for processing. The task-specific layer consists of a 2-layer fully connected network with 256 neurons in the first layer and the number of neurons in the second layer equal to the number of entity categories. Through these fully connected layers, the model is able to extract and classify entity information in the sequence. Assume that the number of entity categories is , then the output of this layer can be expressed as entity recognition logits ,in Represents the time step The conditional random field (CRF) layer is applied to the entity recognition logits of historical texts. The CRF layer can use the dependency of sequence labels to calculate the optimal label sequence and improve the accuracy of named entity recognition. The goal of CRF is to maximize the conditional probability of sequence labels, as follows:

[0101] ;

[0102] in, represents a tag sequence, It is an energy function composed of logits and label conversion scores. Through the Viterbi algorithm, find the label sequence that maximizes the above conditional probability and obtain the named entity recognition result of the historical text. After the named entity recognition is completed, the entity pair representation is constructed based on the context-aware representation of the historical text and the named entity recognition results. Through splicing and pooling operations, these representations are integrated into the first entity pair vector of fixed dimension. The entity pair vector is input into the specific layer of the relationship extraction task. The task-specific layer consists of 3 layers of convolutional neural networks. The convolution kernel sizes of each layer are 3, 4, and 5 respectively, and the number of convolution kernels is 128. Through convolution kernels of different sizes, the model can extract feature maps of different scales from entity pairs. These feature maps capture local patterns and semantic information in the text. Assume that the input entity pair vector is , the output multi-scale feature map can be expressed as ,in is the feature map after processing with different convolution kernels. Perform the maximum pooling operation on the multi-scale feature map, select the maximum value of each feature map, reduce the dimension of the data, and retain the most significant feature information. By concatenating the pooled results, we get the relationship representation vector r and input it into the fully connected layer. The number of neurons is equal to the number of relationship categories. The output logits can be expressed as , where each element represents the category score of the relation. A joint loss function is constructed based on the entity recognition logits and relation classification logits of historical texts for multi-task learning. Assume that the entity recognition loss is (such as cross entropy loss), the relation classification loss is , then the joint loss can be expressed as:

[0103] ;

[0104] in, and is a hyperparameter for adjusting weights. The Adam optimizer is used to update the model parameters to minimize the joint loss and obtain the trained multi-task model.

[0105] In a specific embodiment, the process of executing step 400 may specifically include the following steps:

[0106] Preprocess the target medical asset management text data, including word segmentation, stop word removal and standardization, to obtain the preprocessed target text data;

[0107] Based on the preprocessed target text data, use the pre-trained language model to extract features and obtain the fused feature vector of the target text;

[0108] The fusion feature vector of the target text is transformed in dimension and the dimension is adjusted to 512 through the fully connected layer to obtain the shared feature representation of the target text;

[0109] Based on the shared feature representation of the target text, the trained bidirectional long short-term memory network is used for sequence encoding to obtain the sequence encoding representation of the target text;

[0110] Applying self-attention mechanism to the sequence encoding representation of target text to obtain context-aware representation of target text;

[0111] The context-aware representation of the target text is input into the trained named entity recognition task-specific layer to obtain the entity recognition logits of the target text, and the optimal annotation sequence is calculated through the conditional random field layer to obtain the named entity recognition result of the target text;

[0112] Based on the context-aware representation of the target text and the named entity recognition results of the target text, an entity pair representation is constructed, and a second entity pair vector of a fixed dimension is obtained through concatenation and pooling operations. The second entity pair vector is input into the trained relation extraction task-specific layer to obtain the relation classification logits of the target text.

[0113] According to the named entity recognition results of the target text and the relationship classification logits of the target text, structured medical asset management information is constructed. The structured medical asset management information includes entity types, entity boundaries, and relationship types between entities.

[0114] Specifically, the target medical asset management text data is preprocessed. Preprocessing includes word segmentation, stop word removal and standardization. Word segmentation is to divide continuous text data into separate words or phrases. Stop word removal is to delete common words that contribute little to the semantics of the text, which helps to reduce noise in the data. Then standardization is performed to unify the different forms of numbers, units, etc. in the text to obtain the preprocessed target text data. Based on the preprocessed target text data, the pre-trained language model is used to extract features to obtain the fused feature vector of the target text. By inputting the preprocessed text into the pre-trained model, feature representations rich in contextual information are extracted. For example, using the BERT model, deep semantic representations of each word are obtained, which capture the meaning of the word in different contexts. Assume that the input text is , through the processing of the pre-trained model, the fusion feature vector is obtained ,in is the set of vector representations of each word. The fusion feature vector of the target text is transformed in dimension. The dimension is adjusted to 512 through the fully connected layer to obtain the shared feature representation of the target text. The formula can be expressed as:

[0115] ;

[0116] in, is the weight matrix for dimensionality transformation, is the bias term, is the transformed shared feature representation with a dimension of 512. After obtaining the shared feature representation, the trained bidirectional long short-term memory network (BiLSTM) is used for sequence encoding. BiLSTM is a neural network model that can capture the context information of sequence data and is suitable for processing time series data or text with sequential relationships. Assume that the shared feature representation is , then BiLSTM will generate a sequence encoding representation ,in is the time step The hidden state vector of the target text. These hidden states represent the semantic and contextual information of the text at each position. The self-attention mechanism is applied to the sequence encoding representation of the target text to enhance the model's ability to focus on key information. The self-attention mechanism generates a context-aware representation by calculating the attention weights of different positions in the sequence, highlighting the most relevant parts of the current task. Assume that the input is , the output of the self-attention mechanism is ,in is the time step The context representation of . The formula is expressed as:

[0117] ;

[0118] in, , and They are respectively The query, key and value matrices obtained by linear transformation, is the dimension of the attention head. Through the self-attention mechanism, the model is able to capture richer contextual relationships. The generated context-aware representation is input into the trained named entity recognition (NER) task-specific layer, which usually includes several layers of fully connected networks to process and classify each labeled entity category. Assume that the named entity recognition layer consists of two layers of fully connected networks, with 256 neurons in the first layer and the number of neurons in the second layer equal to the number of entity categories. The output logits can be expressed as ,in Represents the time step Entity category score. A conditional random field (CRF) layer is applied to the entity recognition logits. The CRF layer can exploit the dependency of sequence labels to calculate the optimal label sequence and improve the accuracy of entity recognition. CRF generates named entity recognition results for the target text by maximizing the conditional probability of sequence labels and finding the optimal label sequence. An entity pair representation is constructed based on the context-aware representation of the target text and the named entity recognition results. By concatenating and pooling these representations, a second entity pair vector of fixed dimension is obtained. The constructed entity pair vector is input into the trained relationship extraction task-specific layer for processing. This layer is usually composed of several layers of convolutional neural networks to extract complex relationships between entity pairs. Assume that the convolutional layer uses convolution kernels of different sizes to extract multi-scale feature maps, and finally obtains a compact relationship representation vector through the maximum pooling operation. The output logits can be expressed as , where each element represents the relationship category score of an entity pair. Based on the named entity recognition results and relationship classification loqits of the target text, structured medical asset management information is constructed. This information includes the recognized entity types, entity boundaries, and relationship types between entities. This structured information can be used to automatically manage medical assets, such as identifying the usage scenarios of equipment, recording the relationship between equipment and venues, etc.

[0119] In a specific embodiment, the process of executing step 500 may specifically include the following steps:

[0120] Perform feature extraction on structured medical asset management information, extract factors such as equipment location, execution of operating specifications, equipment load, and use of intelligent auxiliary systems, and obtain an initial feature set;

[0121] Perform data cleaning and standardization on the initial feature set, including missing value filling, outlier processing and numerical normalization, to obtain the preprocessed feature data set;

[0122] A multi-factor variance analysis model was constructed based on the pre-processed feature data set. The medical equipment utilization efficiency was taken as the dependent variable, and the main effect and interaction effect of each factor were calculated to obtain the multi-factor variance analysis results.

[0123] Conduct hypothesis testing on the results of multi-factor variance analysis, use the F test method to calculate the significance level of each factor, and obtain a set of significant factors;

[0124] A generalized additive model was constructed based on the significant factor set, and the nonlinear relationship between each factor and the utilization efficiency was fitted using the spline function to obtain a nonparametric regression model.

[0125] Perform model diagnosis on nonparametric regression models, including residual analysis, influence point detection and multicollinearity test, and obtain model optimization suggestions;

[0126] The generalized additive model was adjusted according to the model optimization suggestions, including variable selection, smoothing parameter optimization and interaction term addition, to obtain the optimized generalized additive model;

[0127] Based on the optimized generalized additive model, the marginal effect and relative importance of each factor on the utilization efficiency of medical equipment were calculated, and combined with the results of multi-factor variance analysis, the results of comprehensive influencing factor analysis were obtained.

[0128] Specifically, feature extraction is performed on structured medical asset management information to extract factors such as equipment location, implementation of operating specifications, equipment load, and use of intelligent assistance systems. Equipment location refers to the physical location or department where the equipment is located, which may affect the frequency of use of the equipment; the implementation of operating specifications measures whether the operating specifications are strictly followed during the use of the equipment; the equipment load reflects the work intensity or usage rate of the equipment; and the use of intelligent assistance systems indicates whether there is support from intelligent systems when the equipment is used. Through the extraction of these features, an initial feature set is formed. Data cleaning and standardization are performed on the initial feature set. For missing values, mean filling, interpolation, or filling based on similar records are used to complete them. For outlier processing, box plots or standard deviation methods are used to identify and process outliers to avoid the impact of these data on the model. Numerical normalization is performed to convert data of different dimensions into the same range, such as scaling all eigenvalues ​​to the interval of [0, 1] to obtain the preprocessed feature data set. Based on the preprocessed feature data set, a multifactor analysis of variance model is constructed. Multifactor analysis of variance is used to study the impact of multiple factors on the dependent variable. The efficiency of medical equipment use is taken as the dependent variable, and the main effect and interaction effect of each factor are analyzed. The main effect refers to the direct impact of a single factor on the dependent variable, while the interaction effect examines the joint impact of multiple factors. For example, there may be an interaction effect between the equipment location and the equipment load, and the equipment in different locations may show different efficiency under different load conditions. By calculating the sum of the squares of each factor, the contribution of each factor is obtained, and the formula is as follows:

[0129] ;

[0130] in, is the sum of the squares of a factor, is the number of levels of the factor, is the number of observations at each level, is the mean of the level, is the overall mean. By calculating the sum of squares and mean square (MS=SS / df, where df is the degree of freedom), the main effect and interaction effect of each factor are obtained. Hypothesis test is performed on the results of multi-factor variance analysis, and the F test method is used to calculate the significance level of each factor's influence on the efficiency of medical equipment use. The F value calculation formula is:

[0131] ;

[0132] in, is the mean square of the factors, is the mean square of the error. By comparing the calculated F value with the critical F value, the significance of each factor is judged and a set of significant factors is formed. The set of significant factors includes factors that have a significant impact on the efficiency of medical equipment use. Based on the set of significant factors, a generalized additive model (GAM) is constructed. GAM allows the influence of each factor on the dependent variable to be nonlinear, and fits the nonlinear relationship between each factor and the efficiency of use through a spline function. The form of the generalized additive model is:

[0133] ;

[0134] in, is the dependent variable, i.e., the efficiency of medical equipment use; is the intercept term; is the first The effect of a factor; is the error term. Through GAM, the impact of various factors on utilization efficiency can be captured more flexibly. Model diagnosis is performed on the preliminary nonparametric regression model. Including residual analysis, influence point detection and multicollinearity test. Residual analysis is used to check the goodness of model fit. Ideally, the residuals should follow a normal distribution and have no obvious pattern; influence point detection is used to identify individual data points that have an abnormal impact on the model; multicollinearity test uses indicators such as VIF (variance inflation factor) to determine whether there is a high correlation between independent variables. Through these diagnoses, model optimization suggestions are obtained. According to the model optimization suggestions, the generalized additive model is adjusted. Including variable selection, smoothing parameter optimization and interaction term addition. For example, if some variables contribute less to the model, they can be removed from the model; smoothing parameter optimization involves adjusting the smoothness of the spline function to avoid overfitting or underfitting; the addition of interaction terms can further capture the complex relationship between factors. Through these adjustments, the optimized generalized additive model is obtained. Based on the optimized generalized additive model, the marginal effect and relative importance of each factor on the utilization efficiency of medical equipment are calculated. The marginal effect reflects the influence of each factor on the utilization efficiency at different levels, while the relative importance measures the contribution of each factor in the overall model. Combined with the results of multi-factor variance analysis, the results of comprehensive influencing factor analysis are obtained.

[0135] In a specific embodiment, the execution step constructs a multi-factor variance analysis model based on the preprocessed feature data set, takes the medical equipment utilization efficiency as the dependent variable, calculates the main effect and interaction effect of each factor, and obtains the multi-factor variance analysis result. The process may specifically include the following steps:

[0136] Factor encode the preprocessed feature data set, convert the categorical variables into dummy variables, and obtain the encoded feature matrix;

[0137] Based on the coded feature matrix, a full factorial design matrix is ​​constructed, including main effect terms and interaction effect terms, to obtain the design matrix;

[0138] Perform QR decomposition on the design matrix to obtain the orthogonalized design matrix and upper triangular matrix;

[0139] Based on the orthogonalized design matrix, the sum of squares of each effect is calculated, including the sum of squares of the main effect and the sum of squares of the interaction effect, and the effect sum of squares vector is obtained;

[0140] According to the effect square sum vector and the degree of freedom of each effect, the mean square value is calculated to obtain the effect mean direction quantity;

[0141] Based on the effect mean square and error mean square, the F statistic is calculated to obtain the F value of each effect;

[0142] Perform a significance test on the F value of each effect, calculate the p value, and screen the significant effects according to the preset significance level to obtain a significant effect set;

[0143] Based on the set of significant effects, the variance explanation ratio and effect size of each significant effect were calculated to obtain the results of multifactor analysis of variance.

[0144] Specifically, the categorical variables in the data set are factor-encoded, and the categorical variables are converted into dummy variables to form an encoded feature matrix. Through dummy variable encoding, the different values ​​(or categories) of each categorical variable are converted into a set of binary variables. Based on the encoded feature matrix, a full factorial design matrix is ​​constructed. This matrix contains all possible main effect terms and interaction effect terms. The main effect term refers to the direct effect of a single factor on the dependent variable, while the interaction effect term represents the effect of the joint effect of multiple factors on the dependent variable. For example, consider two factors A and B, A has two levels (such as high and low), and B also has two levels (such as presence and absence). The full factorial design matrix will contain the main effect terms of A and B (the effects of A and B separately), and the interaction effect term between A and B (the combined effect of A and B). Perform QR decomposition on the design matrix. QR decomposition is a common method in linear algebra that decomposes a matrix into an orthogonal matrix (Q matrix) and an upper triangular matrix (R matrix). The column vectors of an orthogonal matrix are orthogonal to each other, that is, they are independent of each other, while an upper triangular matrix is ​​a matrix with zero elements above (or below) the diagonal. Through QR decomposition, the design matrix is ​​orthogonalized to obtain the orthogonalized design matrix and upper triangular matrix, eliminating the multicollinearity problem between variables. Based on the orthogonalized design matrix, the sum of squares of each effect is calculated, including the sum of squares of the main effect and the sum of squares of the interaction effect. The sum of squares is an important indicator to measure the contribution of each effect to the dependent variable. Assume that the design matrix is , the dependent variable is , then the sum of squares of a certain effect can be calculated by the following formula:

[0145] ;

[0146] in, is the regression coefficient associated with the effect, is the orthogonalized design matrix, is the dependent variable vector. Through this process, we get the effect sum of squares vector , where each element represents the sum of squares of different effects. The mean square value is calculated based on the effect sum of squares vector and the degrees of freedom of each effect. The mean square value is the ratio of the sum of squares to the degrees of freedom, which represents the size of the sum of squares per unit degree of freedom. The calculation formula for the mean square value is:

[0147] ;

[0148] in, is the mean square of the effect, is the sum of the squares of the effects, is the degree of freedom of the effect. Through these mean square values, we can understand the impact of each effect on the dependent variable. Based on the effect mean square and the error mean square, the F statistic is calculated to obtain the F value of each effect. The F statistic is used to compare the mean square value of each effect with the error mean square value to determine whether the effect is significant. The calculation formula of the F value is:

[0149] ;

[0150] in, is the mean square value of the error. By calculating the F value, the significance of each effect is measured and its importance in the model is determined. The F value of each effect is tested for significance and the p value is calculated. The p value indicates whether the observed effect is significant at a given significance level. If the p value is less than the preset significance level (usually 0.05), the effect is considered significant and can enter the final analysis model. According to the results of the significance test, the significant effects are screened out to obtain a set of significant effects. Based on the set of significant effects, the variance explanation ratio and effect size of each significant effect are calculated. The variance explanation ratio indicates the contribution of each significant effect to the total variance, while the effect size measures the influence of the effect. Through these calculations, the final results of the multi-factor analysis of variance are obtained, and it is clear which factors or combinations of factors have a significant impact on the dependent variable (such as the efficiency of the use of medical equipment).

[0151] In a specific embodiment, the process of executing step 600 may specifically include the following steps:

[0152] Calculate the factor weights of the comprehensive influencing factor analysis results to obtain the factor weight vector;

[0153] Based on the factor weight vector, the existing medical resource allocation status is evaluated, the comprehensive utility score of each resource unit is calculated, and the resource utility evaluation matrix is ​​obtained;

[0154] Cluster analysis is performed on the resource utility evaluation matrix, and resource units are divided into three categories: high efficiency, medium efficiency, and low efficiency, and the resource efficiency classification results are obtained;

[0155] According to the resource efficiency classification results, a resource scheduling optimization model is constructed, and the maximization of resource utilization is used as the objective function, and resource capacity and demand matching are used as constraints to obtain the initial optimization model;

[0156] Solving the initial optimization model to obtain a set of candidate resource allocation solutions, and performing local search optimization based on the set of candidate resource allocation solutions to obtain an optimized set of resource allocation solutions;

[0157] Perform multi-objective evaluation on the optimized resource allocation scheme set, consider resource utilization, scheduling cost and service quality indicators, calculate the comprehensive score of each scheme, and obtain the scheme ranking result;

[0158] According to the scheme ranking results, the scheme with the highest comprehensive score is selected as the optimal resource allocation scheme. The optimal resource allocation scheme includes the optimal location, staffing and usage schedule of each medical equipment.

[0159] Specifically, the factor weights are calculated for the comprehensive influencing factor analysis results to obtain the factor weight vector. Assume that three key factors are identified in the comprehensive influencing factor analysis, namely, equipment location, operation specification execution, and equipment load. The contribution of each factor to resource utilization can be calculated through multivariate regression analysis or other statistical methods to determine their respective weights. Suppose these weights are , and , then the factor weight vector can be expressed as ,in Based on the calculated factor weight vector , evaluate the existing medical resource allocation, calculate the comprehensive utility score of each resource unit, and form a resource utility evaluation matrix. Each resource unit (such as a medical device or a medical team) can calculate its utility score based on the weight of the above factors and the actual performance of each factor. For example, for a medical device, its utility score It can be calculated by the following formula:

[0160] ;

[0161] in, , and Respectively represent the equipment location, operation specification execution and equipment load on the The resource utility evaluation matrix is ​​formed by performing similar calculations on all resource units. , where each element represents the comprehensive utility score of a resource unit. Cluster analysis is performed on the resource utility evaluation matrix to divide the resource units into three categories: high efficiency, medium efficiency, and low efficiency, and the resource efficiency classification results are obtained. Cluster analysis is an unsupervised learning method used to identify natural groupings in data. Using the K-means clustering algorithm, three categories (k=3) are determined and then classified according to the comprehensive utility scores of resource units. For example, the efficient class contains those resource units with the highest utility scores, while the inefficient class contains the units with the lowest scores. The medium efficiency class is between the two. This classification helps to identify bottlenecks and advantages in the current resource configuration. Based on the resource efficiency classification results, a resource scheduling optimization model is constructed, with maximizing resource utilization as the objective function, and at the same time, matching resource capacity and demand as constraints. The objective function can be defined as:

[0162] ;

[0163] in, is the sum of resource utilization, It is The comprehensive utility score of each resource unit, is a decision variable, indicating whether to choose Resource Units Indicates choice, indicates no selection). Constraints can include resource capacity and demand For example:

[0164] ;

[0165] in, It is The resource demand of each demand. After constructing the initial optimization model, the set of candidate resource allocation solutions is obtained by solving the model. Local search optimization is performed based on the set of candidate solutions. The purpose of local search optimization is to find a better resource allocation solution through small-scale adjustments and searches based on the candidate solutions. For example, simulated annealing or genetic algorithm is used for local search to obtain a set of optimized resource allocation solutions. Multi-objective evaluation is performed on the optimized set of resource allocation solutions, and the comprehensive score of each solution is calculated by considering multiple indicators such as resource utilization, scheduling cost and service quality. The calculation of the comprehensive score can be performed by the weighted summation method:

[0166] ;

[0167] in, It is is the resource utilization rate, is the scheduling cost, It is the quality of service. , and is the weight coefficient of the corresponding indicator, . According to the results of the scheme ranking, the scheme with the highest comprehensive score is selected as the optimal resource allocation scheme. The optimal scheme should find the best balance between resource utilization, cost and quality. For example, if after analysis, a certain scheme has the highest utilization rate, moderate cost and high quality score, then this scheme will become the best choice. The optimal resource allocation scheme includes the optimal location, staffing and usage schedule of each medical equipment. The location of the equipment can determine its frequency of use, staffing is related to the operating efficiency of the equipment, and the usage schedule affects the overall utilization and maintenance plan of the equipment.

[0168] The above describes the medical asset management big data analysis method in the embodiment of the present application. The following describes the medical asset management big data analysis system 10 in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the medical asset management big data analysis system 10 includes:

[0169] A preprocessing module 11 is used to preprocess the historical medical asset management text data to obtain a standardized text data set;

[0170] A feature learning module 12 is used to perform feature learning on a standardized text data set using a pre-trained language model, extract context-related first-order features through a masked language model and a next sentence prediction task, and generate second-order fused features through nonlinear combination to obtain a fused feature vector;

[0171] A training module 13 is used to construct a multi-task learning framework according to the fused feature vector, and simultaneously train named entity recognition and relationship extraction tasks to obtain a trained multi-task model;

[0172] An extraction module 14 is used to extract information from the target medical asset management text data based on the trained multi-task model to obtain structured medical asset management information;

[0173] The analysis module 15 is used to establish a multi-factor variance analysis model and a generalized additive model based on the structured medical asset management information, and to perform a multi-factor analysis of the medical equipment utilization efficiency to obtain a comprehensive influencing factor analysis result;

[0174] The scheduling module 16 is used to optimize the scheduling of medical resources based on the comprehensive influencing factor analysis results to obtain the optimal resource allocation plan.

[0175] Through the collaboration of the above components, the standardized processing of historical medical asset management text data is achieved through preprocessing. Deep feature learning: The pre-trained language model is used for feature learning, and the context-related features are extracted through the masked language model and the next sentence prediction task, and the fusion features are generated through nonlinear combination. It can capture the deep semantic information in the text and improve the understanding of the content related to medical asset management. A multi-task learning framework for named entity recognition and relationship extraction tasks is constructed. It can share the underlying feature representation, improve the generalization ability of the model, and at the same time improve the comprehensiveness and accuracy of the extraction of medical asset management information. Through the multi-factor variance analysis model and the generalized additive model, an in-depth analysis of the factors affecting the efficiency of medical equipment use is achieved. It can not only identify the main influencing factors, but also capture the interaction effects and nonlinear relationships between factors. Based on the comprehensive influencing factor analysis results, a resource scheduling optimization model considering multiple objectives is constructed. Through local search optimization and multi-objective evaluation, the optimal resource allocation plan that takes into account resource utilization, scheduling cost and service quality can be obtained.

[0176] The present application also provides an electronic device, which includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor executes the steps of the medical asset management big data analysis method in the above-mentioned embodiments.

[0177] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the medical asset management big data analysis method.

[0178] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0179] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0180] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A medical asset management big data analysis method, characterized in that: The medical asset management big data analysis method comprises: Preprocess the historical medical asset management text data to obtain a standardized text dataset; Using a pre-trained language model to perform feature learning on the standardized text dataset, extracting context-related first-order features through a masked language model and a next sentence prediction task, and generating second-order fused features through nonlinear combination to obtain a fused feature vector; Building a multi-task model based on the fused feature vector, and training named entity recognition and relationship extraction tasks at the same time to obtain a trained multi-task model; Extracting information from target medical asset management text data based on the trained multi-task model to obtain structured medical asset management information; A multi-factor variance analysis model and a generalized additive model are established based on the structured medical asset management information, and a multi-factor analysis of the medical equipment utilization efficiency is performed to obtain the comprehensive influencing factor analysis results; specifically, the following steps are performed: feature extraction is performed on the structured medical asset management information to extract factors such as equipment location, execution of operating specifications, equipment load, and usage of intelligent auxiliary systems to obtain an initial feature set; data cleaning and standardization are performed on the initial feature set, including missing value filling, outlier processing, and numerical normalization to obtain a preprocessed feature data set; factor encoding is performed on the preprocessed feature data set to convert categorical variables into virtual variables to obtain an encoded feature matrix; based on the encoded feature matrix, a full factorial design matrix is ​​constructed, including main effect terms and interaction effect terms to obtain a design matrix; QR decomposition is performed on the design matrix to obtain an orthogonalized design matrix and an upper triangular matrix; based on the orthogonalized design matrix, the sum of squares of each effect is calculated, including the sum of squares of the main effect and the sum of squares of the interaction effect to obtain an effect sum of squares vector; the mean square value is calculated based on the effect sum of squares vector and the degrees of freedom of each effect to obtain the effect mean square vector; based on the effect The mean square and error mean square are used to calculate the F statistic and obtain the F value of each effect; the F value of each effect is tested for significance, the p value is calculated, and the significant effects are screened according to the preset significance level to obtain a significant effect set; based on the significant effect set, the variance explanation ratio and effect size of each significant effect are calculated to obtain the multi-factor analysis of variance result; the multi-factor analysis of variance result is tested for hypothesis, the significance level of each factor is calculated by the F test method, and a significant factor set is obtained; a generalized additive model is constructed based on the significant factor set, and the spline function is used to compare each factor with the The nonlinear relationship between the utilization efficiency and the utilization efficiency is fitted to obtain a nonparametric regression model; the nonparametric regression model is diagnosed, including residual analysis, influence point detection and multicollinearity test, to obtain model optimization suggestions; the generalized additive model is adjusted according to the model optimization suggestions, including variable selection, smoothing parameter optimization and interaction term addition, to obtain an optimized generalized additive model; based on the optimized generalized additive model, the marginal effect and relative importance of each factor on the utilization efficiency of medical equipment are calculated, and combined with the multifactor variance analysis results, the comprehensive influencing factor analysis results are obtained; Based on the analysis results of the comprehensive influencing factors, medical resources are optimized and scheduled to obtain the optimal resource allocation plan.

2. The medical asset management big data analysis method according to claim 1, characterized in that: The historical medical asset management text data is preprocessed to obtain a standardized text data set, including: Perform word segmentation processing on historical medical asset management text data to obtain initial word segmentation results; Performing professional terminology recognition on the initial word segmentation result based on a predefined medical asset management field dictionary to obtain an optimized word segmentation result; Constructing a word frequency-inverse document frequency matrix according to the optimized word segmentation results, and setting a threshold to filter stop words to obtain a stop word list; Filtering the optimized word segmentation results based on the stop word list to obtain a text without stop words, and standardizing the numbers and unit representations in the text without stop words to obtain a normalized text; Perform synonym replacement on the standardized text based on a pre-built synonym library in the field of medical asset management to obtain a semantically unified text; A word vector model is used to vectorize the semantically unified text to obtain an initial word vector, and the initial word vector is mapped to a low-dimensional space to obtain a standardized text data set.

3. The medical asset management big data analysis method according to claim 1, characterized in that: The method uses a pre-trained language model to perform feature learning on the standardized text dataset, extracts context-related first-order features through a masked language model and a next sentence prediction task, and generates second-order fused features through nonlinear combination to obtain a fused feature vector, including: Segmenting the standardized text data set to obtain input sequences, each of which has a length not exceeding 512 tokens; Performing a superposition process of tag embedding, position embedding and segment embedding on the input sequence to obtain an initial input embedding; Embedding the initial input into the encoder of the pre-trained BERT model, the encoder includes 12 Transformer layers, each Transformer layer includes a multi-head self-attention sublayer and a feed-forward neural network sublayer, wherein the multi-head self-attention sublayer uses 8 attention heads, each attention head has a dimension of 96, and the feed-forward neural network sublayer includes two linear transformations, and the intermediate activation function is GELU; In each Transformer layer of the encoder, attention weights are calculated through a multi-head self-attention mechanism to obtain a context-dependent intermediate representation; The context-related intermediate representation is nonlinearly transformed through a feedforward neural network to obtain an output representation of each layer, and a vector corresponding to the tag is extracted from the output of the last Transformer layer of the encoder as a first-order feature representation of the entire sequence; Based on the first-order feature representation, 15% of the randomly masked input tokens are predicted through the masked language model task to obtain the reconstruction loss, and based on the first-order feature representation, whether two sentences are continuous is judged through the next sentence prediction task to obtain the binary classification loss; Performing a weighted summation of the reconstruction loss and the binary classification loss to obtain an overall loss function, and updating the model parameters through back propagation; The output representations of the last three Transformer layers of the encoder are nonlinearly combined, specifically, the output representations of the last three Transformer layers are fused through a fully connected layer and a ReLU activation function to obtain a fused feature vector.

4. The medical asset management big data analysis method according to claim 3 is characterized in that: The context-related intermediate representation is subjected to nonlinear transformation through a feedforward neural network to obtain an output representation of each layer, and a vector corresponding to a tag is extracted from the output of the last Transformer layer of the encoder as a first-order feature representation of the entire sequence, including: Performing layer normalization processing on the context-related intermediate representation to obtain a normalized intermediate representation; Input the normalized intermediate representation into a two-layer fully connected neural network, where the number of neurons in the first layer is 3072 and the number of neurons in the second layer is 768, to obtain an intermediate feature; Applying a GELU activation function to the intermediate feature to obtain a nonlinear activation feature, and performing a dropout process on the nonlinear activation feature with a dropout rate of 0.1 to obtain a regularized feature; The regularized feature is input into the output layer, the number of neurons is 768, and the original output representation is obtained. The original output representation is residually connected and added to the normalized intermediate representation of the input to obtain the residual output representation; Performing a second layer normalization process on the residual output representation to obtain an output representation of each layer; From the output representation of each layer of the last Transformer layer of the encoder, a 768-dimensional vector corresponding to each tag is extracted, and the 768-dimensional vectors are concatenated into a matrix to obtain a first-order feature representation of the entire sequence.

5. The medical asset management big data analysis method according to claim 1, characterized in that: The multi-task model is constructed according to the fused feature vector, and named entity recognition and relationship extraction tasks are trained simultaneously to obtain a trained multi-task model, including: Transforming the dimension of the fused feature vector, adjusting the dimension to 512 through a fully connected layer, and obtaining a shared feature representation of the historical text; Based on the shared feature representation of the historical text, a bidirectional long short-term memory network is constructed, including a 3-layer BiLSTM structure, each layer of the BiLSTM structure includes 256 hidden units, and a sequence encoding representation of the historical text is obtained; Applying a self-attention mechanism to the sequence encoding representation of the historical text, using 8 attention heads, each with a dimension of 64, to obtain a context-aware representation of the historical text; Input the context-aware representation of the historical text into a named entity recognition task-specific layer, including a 2-layer fully connected network, where the number of neurons in the first layer is 256 and the number of neurons in the second layer is equal to the number of entity categories, to obtain entity recognition logits of the historical text; Applying a conditional random field layer to the entity recognition logits of the historical text, calculating the optimal annotation sequence, and obtaining the named entity recognition result of the historical text; Based on the context-aware representation of the historical text and the named entity recognition result of the historical text, construct an entity pair representation, and obtain a first entity pair vector of fixed dimension through concatenation and pooling operations; Inputting the first entity pair vector into a relation extraction task-specific layer, comprising three layers of convolutional neural networks, where the convolution kernel sizes of each layer are 3, 4, and 5 respectively, and the number of convolution kernels is 128, to obtain a multi-scale feature map; Performing a maximum pooling operation on the multi-scale feature map, and concatenating the pooling results to obtain a relationship representation vector, and inputting the relationship representation vector into a fully connected layer, where the number of neurons is equal to the number of relationship categories, to obtain relationship classification logits of the historical text; A joint loss function is constructed based on the entity recognition logits of the historical text and the relationship classification logits of the historical text, and the Adam optimizer is used to update the parameters to obtain a trained multi-task model.

6. The medical asset management big data analysis method according to claim 5, characterized in that: The step of extracting information from the target medical asset management text data based on the trained multi-task model to obtain structured medical asset management information includes: Preprocessing the target medical asset management text data, including word segmentation, stop word removal and standardization, to obtain preprocessed target text data; Based on the preprocessed target text data, extract features using a pre-trained language model to obtain a fused feature vector of the target text; Transforming the dimension of the fused feature vector of the target text, adjusting the dimension to 512 through a fully connected layer, and obtaining a shared feature representation of the target text; Based on the shared feature representation of the target text, a trained bidirectional long short-term memory network is used to perform sequence encoding to obtain a sequence encoding representation of the target text; Applying a self-attention mechanism to the sequence encoding representation of the target text to obtain a context-aware representation of the target text; Inputting the context-aware representation of the target text into the trained named entity recognition task-specific layer to obtain entity recognition logits of the target text, and calculating the optimal annotation sequence through the conditional random field layer to obtain the named entity recognition result of the target text; Based on the context-aware representation of the target text and the named entity recognition result of the target text, construct an entity pair representation, obtain a second entity pair vector of fixed dimension through concatenation and pooling operations, and input the second entity pair vector into the trained relation extraction task-specific layer to obtain relation classification logits of the target text; According to the named entity recognition results of the target text and the relationship classification logits of the target text, structured medical asset management information is constructed, and the structured medical asset management information includes entity types, entity boundaries, and relationship types between entities.

7. The medical asset management big data analysis method according to claim 1, characterized in that: The optimizing and scheduling of medical resources based on the comprehensive influencing factor analysis results to obtain the optimal resource allocation plan includes: Calculating the factor weights of the comprehensive influencing factor analysis results to obtain a factor weight vector; Based on the factor weight vector, the existing medical resource allocation status is evaluated, the comprehensive utility score of each resource unit is calculated, and a resource utility evaluation matrix is ​​obtained; Performing cluster analysis on the resource utility evaluation matrix, dividing resource units into three categories: high efficiency, medium efficiency and low efficiency, and obtaining resource efficiency classification results; A resource scheduling optimization model is constructed according to the resource efficiency classification result, and the maximization of resource utilization is used as the objective function, and resource capacity and demand matching are used as constraints to obtain an initial optimization model; Solving the initial optimization model to obtain a set of candidate resource allocation solutions, and performing local search optimization based on the set of candidate resource allocation solutions to obtain an optimized set of resource allocation solutions; Performing a multi-objective evaluation on the optimized resource allocation scheme set, taking into account resource utilization, scheduling cost and service quality indicators, calculating the comprehensive score of each scheme, and obtaining a scheme ranking result; According to the scheme ranking results, the scheme with the highest comprehensive score is selected as the optimal resource allocation scheme, and the optimal resource allocation scheme includes the optimal location, staffing and usage time arrangement of each medical device.

8. A medical asset management big data analysis system, characterized in that: Used to execute the medical asset management big data analysis method according to any one of claims 1 to 7, the medical asset management big data analysis system comprises: A preprocessing module is used to preprocess the historical medical asset management text data to obtain a standardized text data set; A feature learning module, used to perform feature learning on the standardized text dataset using a pre-trained language model, extract context-related first-order features through a masked language model and a next sentence prediction task, and generate second-order fused features through nonlinear combination to obtain a fused feature vector; A training module, used to construct a multi-task model according to the fused feature vector, and simultaneously train named entity recognition and relationship extraction tasks to obtain a trained multi-task model; An extraction module, used to extract information from the target medical asset management text data based on the trained multi-task model to obtain structured medical asset management information; An analysis module, used to establish a multi-factor variance analysis model and a generalized additive model based on the structured medical asset management information, and to perform a multi-factor analysis of the medical equipment utilization efficiency to obtain a comprehensive influencing factor analysis result; The scheduling module is used to optimize the scheduling of medical resources based on the analysis results of the comprehensive influencing factors to obtain the optimal resource allocation plan.

Images