A patient and hospital collaborative matching intelligent referral decision method and storage medium

Abstract

This invention relates to an intelligent referral decision-making method and storage medium for collaborative matching of patients and hospitals. First, a triplet sample set integrating patient-hospital-fitness is constructed. Then, a heterogeneous network graph of the medical consortium, including patients, hospitals, and two types of related edges, is built. A graph attention network is used to mine deep correlation features between nodes. Compared to traditional manual or single-feature matching methods, the accuracy of patient-hospital fit assessment is significantly improved, reducing mismatches and incorrect matching. Next, relying on the graph attention network, feature aggregation and fit calculation are automatically completed, quickly outputting the fit score between the patient and each candidate hospital. This replaces the cumbersome process of manual screening and evaluation, significantly shortening the time required for referral decisions, guiding patients to hospitals with matching treatment capabilities and resource availability, avoiding overcrowding of high-quality medical resources, improving resource utilization in primary hospitals, and facilitating the implementation of hierarchical medical treatment within medical consortia.

Description

Technical Field

[0001] This invention relates to the fields of smart healthcare and medical information technology, and in particular to an intelligent referral decision-making method and storage medium for collaborative matching between patients and hospitals. Background Technology

[0002] Currently, referrals within medical consortia largely rely on doctors' experience, administrative orders, or simple geographical proximity. Although some information platforms have emerged, the following pain points still exist: One-sided referral decisions: a lack of systematic balance between medical quality, cost efficiency, and patient convenience; Fragmented pathway management: information is not shared between hospitals at different levels, and there is a lack of continuous pathway management and effect tracking after patients are transferred out; Difficulty in quantifying value objectives: The concept of "value-based healthcare" lacks operable technical tools for implementation, making it difficult to translate the goal of "achieving the best health outcomes at the lowest cost" into concrete treatment decisions.

[0003] With the development of artificial intelligence technology, intelligent referral decision-making models based on machine learning and deep learning have gradually become a research hotspot. In existing technologies, some solutions extract patient condition characteristics and hospital qualification characteristics, and use traditional machine learning algorithms such as logistic regression and random forests to construct a fit assessment model, achieving preliminary patient-hospital matching. However, such models only focus on "point-to-point" feature matching between patients and hospitals, ignoring topological information such as the collaborative relationships between hospitals within the medical consortium and the connectivity of referral paths. This can easily lead to unreasonable decisions such as "high fit but incompatible referral paths." Furthermore, the models have poor adaptability to individual patient differences and dynamic changes in hospital resources, making it difficult to support complex two-way referral scenarios.

[0004] To address the aforementioned issues, some improved solutions introduce graph neural network technology to construct heterogeneous network graphs containing patient and hospital nodes, leveraging the feature aggregation capabilities of graph networks to uncover the relationships between nodes. However, existing graph neural network referral models still suffer from the following key shortcomings:

[0005] Insufficient feature aggregation capability: It mostly uses single-layer graph convolution or simple attention mechanism, which cannot deeply capture multi-level association features of patient-hospital matching edge and hospital-hospital collaboration edge, resulting in low fitting accuracy.

[0006] The training strategy is unreasonable: the end-to-end one-time training mode is often used, without taking into account the characteristics of medical consortium data, such as "few hospital samples and sparse collaboration relationships". The model is prone to overfitting and has poor generalization ability in small sample scenarios.

[0007] The loss function design lacks specificity: existing models mostly use general loss functions such as mean squared error loss and cross-entropy loss, without considering the core needs of referral decision-making—the rationality of graph structure ranking and the balance of hospital resource load. On the one hand, the loss function does not constrain the model output to conform to the priority ranking of the medical consortium's referral path, causing the recommendation results to deviate from the actual diagnosis and treatment process; on the other hand, it does not consider the pressure on hospital resource load, which can easily lead to an imbalance of overcrowding in tertiary hospitals and idle resources in primary hospitals, violating the original intention of hierarchical medical treatment.

[0008] In addition, the existing intelligent referral models have poor interpretability, and the model output fit scores lack clear decision-making basis, making it difficult for medical staff to accept them. At the same time, the models are not adaptable to dynamic changes in medical alliance resources. When hospitals add new specialties or adjust collaborative relationships, the models need to be retrained, resulting in high maintenance costs.

[0009] Therefore, how to improve it to achieve accurate, efficient and balanced referral decisions and promote the implementation of the hierarchical medical system in medical consortia is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0010] To address at least one of the aforementioned technical problems, this invention provides an intelligent referral decision-making method for collaborative matching of patients and hospitals, comprising:

[0011] Acquire patient and hospital data, label patient-hospital fit scores, and construct a triplet sample set integrating patient-hospital-fit.

[0012] Based on the triplet sample set, patient data is mapped to patient node feature vectors and bound to patient nodes, and hospital data is mapped to hospital node feature vectors and bound to hospital nodes; a heterogeneous network graph of the medical consortium containing patient nodes, hospital nodes, patient-hospital matching edges, and hospital-hospital collaboration edges is constructed.

[0013] The heterogeneous network graph of the medical consortium is input into a preset graph attention network. The attention weights of the nodes are calculated through the graph attention network, and the node features are weighted and aggregated to obtain the node aggregated features. Based on the node aggregated features, the patient-hospital fit score is calculated and output.

[0014] Referral decisions are made based on patient-hospital fit scores.

[0015] Further steps, including acquiring patient and hospital data and labeling patient-hospital fit scores, include:

[0016] Obtain information on hospital level, annual PCI procedure volume percentage, and response time to determine hospital qualification indicators;

[0017] The patient's disease types, severity of each disease, and the hospital's ability to treat each disease are obtained, and a weighted sum is used to determine the patient's condition indicators.

[0018] Obtain the estimated arrival time and straight-line distance between the patient and the hospital, as well as the maximum acceptable time and maximum acceptable distance for patient transfer, to determine geographical accessibility indicators;

[0019] To obtain information on the strength of the patient's past treatment relationship with the hospital, the convenience of follow-up, and the suitability of rehabilitation resources, and to determine indicators of treatment continuity;

[0020] Based on hospital qualification indicators, patient condition indicators, geographical accessibility indicators, and treatment continuity indicators, a patient-hospital fit assessment model is constructed, and a weighted score is determined for the patient-hospital fit.

[0021] Furthermore, the graph attention network is a two-layer cascaded graph attention network, including a cascaded first-layer graph attention module, a second-layer graph attention module, and an fitness calculation module;

[0022] The first layer graph attention module takes the node feature vectors of the heterogeneous network graph of the medical consortium as input, calculates the attention weights between adjacent nodes through a multi-attention head mechanism, and aggregates them to obtain the first layer aggregated features.

[0023] The second-layer graph attention module takes the aggregated features of the first layer as input and repeats the attention weight calculation and feature aggregation process to obtain the second-layer aggregated features.

[0024] The fit calculation module calculates the patient-hospital fit score based on the second-layer aggregated features.

[0025] Furthermore, the first-layer graph attention module includes: a feature input unit, an attention weight calculation unit, and a weighted aggregation unit;

[0026] The input end of the feature input unit is connected to the output end of the node feature vector of the heterogeneous network graph of the medical consortium, and is used to receive and verify the feature vectors of patient nodes and hospital nodes.

[0027] The input end of the attention weight calculation unit is connected to the output end of the feature input unit. It has several parallel attention head branches built in. Each attention head branch establishes a calculation connection with all adjacent node pairs in the heterogeneous graph. It performs a linear transformation on the feature vectors of the source node and the target node through a linear transformation matrix. Then, it outputs the normalized node attention weights through the sequentially connected attention coefficient calculation subunit, LeakyReLU activation subunit, and softmax normalization subunit.

[0028] The input of the weighted aggregation unit is connected to the output of the attention weight calculation unit. Based on the node attention weight, the features of neighboring nodes are weighted and summed to output the first layer of aggregated features.

[0029] The second-layer graph attention module includes: an aggregated input unit, an attention weight recalculation unit, and a weighted aggregated output unit;

[0030] The input end of the aggregation input unit is unidirectionally connected to the output end of the weighted aggregation unit in the first layer graph attention module, and is used to receive the first layer aggregation features and complete the dimension verification.

[0031] The input of the attention weight recalculation unit is connected to the output of the aggregation input unit. The topology, number of attention heads, and connection relationship of the internal sub-units of the attention weight recalculation unit are the same as those of the attention weight calculation unit of the first layer graph attention module. It is used to repeatedly perform attention weight calculation on the first layer aggregated features.

[0032] The input of the weighted aggregation output unit is connected to the output of the attention weight recalculation unit to output the second layer of aggregated features.

[0033] Furthermore, the fit calculation module includes a feature splicing unit, a fully connected mapping unit, and a score output unit;

[0034] The input of the feature splicing unit is connected in two ways to the patient node aggregation feature end and the hospital node aggregation feature end output by the second layer graph attention module, which are used to fuse the aggregation features of the two types of nodes, namely patients and hospitals.

[0035] The input of the fully connected mapping unit is connected to the output of the feature concatenation unit, and the feature dimension transformation is completed through multiple levels of fully connected layers;

[0036] The input of the scoring output unit is connected to the output of the fully connected mapping unit, and is used to map the transformed features into patient-hospital fit scores and output them.

[0037] Furthermore, during training, the total loss function is determined by a weighted average of the main loss function and the balanced loss function; where:

[0038] The main loss function, used for ranking by fusing graph structure information, is expressed as:

[0039] ;

[0040] in, For sorting loss, The number of training samples, As a balance factor, For patients ; For the hospital , For the hospital , For the predicted fit between patient i and hospital j, The predicted fit between patient i and hospital k. The graph structure regularization coefficient is used. The number of nodes in the graph. For nodes The neighborhood group, For nodes and nodes Attention weights , for nodes ,node eigenvectors;

[0041] The balancing function, used to integrate the balancing losses of resource load, is expressed as:

[0042] ;

[0043] in, To balance the losses, For the total number of hospitals, The number of patients allocated to hospital h, For the capacity h of the hospital, Target resource utilization rate;

[0044] The total loss function, weighted, is as follows:

[0045] ;

[0046] in, For the total loss, For sorting loss, To balance the losses, This is the loss weighting coefficient.

[0047] Furthermore, the training process includes: pre-training of hospital node features:

[0048] The pre-trained data is input into the multilayer perceptron, and the model parameters are optimized using the main loss function. During training, the BatchNorm layer is used to accelerate convergence. If the hospital resource load difference coefficient is greater than a set threshold, the balanced loss function is fused to construct the total loss function. After training, the feature vectors output by the intermediate layers of the MLP are extracted as the initial features of the hospital nodes.

[0049] The initial features of hospital nodes are bound to the hospital nodes in the heterogeneous network graph of the medical consortium, serving as fixed basic features for subsequent training.

[0050] Furthermore, the training process also includes: graph structure training:

[0051] The heterogeneous network graph of the medical consortium is input into the graph attention network. The labeled patient-hospital fit scores and inter-hospital referral priority ranking are used as supervision. The error between the predicted ranking and the actual ranking is calculated using the main loss function. At the same time, real-time resource load data of hospitals is introduced to construct a balanced loss function that integrates resource load. Based on the gradient backpropagation of the total loss function, only the parameters of the attention weight calculation unit are updated. The association weights of patient-hospital matching edges and hospital-hospital collaboration edges in the heterogeneous network graph of the medical consortium are learned, as well as the fit adjustment rules under different hospital resource loads. The preliminarily trained graph attention network is output.

[0052] Furthermore, the training process also includes: end-to-end training:

[0053] Unlock the MLP parameters corresponding to the hospital node initialization features frozen during the pre-training phase, as well as the parameters of all modules in the graph attention network; optimize the parameters of the entire network with a small learning rate; retain the attention weight parameters learned during the graph structure training phase during training; the main loss function still adopts the main loss function that integrates graph structure information; add a resource load dynamic adjustment term to the balancing loss function; and integrate L2 regularization loss to suppress overfitting until the iteration termination condition is reached, at which point training stops and the current optimal model parameters are saved as the final training model parameters.

[0054] On the other hand, the present invention also provides a computer storage medium storing executable program code; the executable program code is used to execute any of the above-mentioned intelligent referral decision methods.

[0055] This invention provides an intelligent referral decision-making method and storage medium for collaborative matching of patients and hospitals. First, it constructs a triplet sample set integrating patient-hospital-fitness. Then, it constructs a heterogeneous network graph of the medical consortium, including patients, hospitals, and two types of related edges. Combining graph attention networks to mine deep correlation features between nodes, the accuracy of patient-hospital fit assessment is significantly improved compared to traditional manual or single-feature matching methods, reducing mismatches and incorrect matching. Next, relying on the graph attention network, it automatically completes feature aggregation and fit calculation, quickly outputting the fit score between the patient and each candidate hospital. This replaces the cumbersome process of manual screening and evaluation, significantly shortening the time required for referral decisions, guiding patients to hospitals with matching treatment capabilities and resource availability, avoiding overcrowding of high-quality medical resources, improving resource utilization in primary hospitals, and facilitating the implementation of hierarchical medical treatment within medical consortia. Attached Figure Description

[0056] Figure 1 This is a flowchart of one embodiment of the intelligent referral decision-making method for patient-hospital collaborative matching according to the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0058] It should be noted that if the embodiments of the present invention involve directional indications, such as up, down, left, right, front, back, etc., these directional indications are only used to explain the relative positional relationships and movement of the components in a specific posture. If the specific posture changes, the directional indications will also change accordingly. Furthermore, if the embodiments of the present invention involve descriptions such as "first," "second," "S1," "S2," "step one," "step two," etc., these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance, or implicitly indicating the number of technical features indicated or the order of method execution. Those skilled in the art will understand that anything that does not violate the inventive concept and is within the scope of the present invention should be included in the protection scope of the present invention.

[0059] like Figure 1 As shown, this invention provides an intelligent referral decision-making method for collaborative matching between patients and hospitals, which can optionally adopt a three-layer platform architecture of "data perception layer - intelligent decision-making layer - collaborative application layer"; Data perception layer: through standardized interfaces, it aggregates HIS, EMR, PACS, DRG / DIP settlement systems, patient follow-up systems and external data from member hospitals within the medical consortium; Intelligent decision-making layer: deploys the core "fitness assessment model" and "medical consortium heterogeneous network diagram"; Collaborative application layer: provides multi-terminal applications for doctors, patients and administrators, carrying out business processes such as referral applications, path execution, and effect tracking.

[0060] Specifically, intelligent referral decision-making methods that collaboratively match patients and hospitals include:

[0061] S1: Obtain patient data and hospital data, label the patient-hospital fit score, and construct a triplet sample set that integrates patient-hospital-fit score;

[0062] Specifically, patient data may include, but is not limited to, disease characteristics, individual baseline characteristics, and treatment risk characteristics; hospital data may include, but is not limited to, medical qualification characteristics, resource allocation characteristics, service efficiency characteristics, and medical consortium collaboration characteristics; and based on the medical consortium referral and treatment guidelines, the compatibility relationship between patients and hospitals is labeled to construct a triplet sample set integrating patient P-hospital H-compatibility S.

[0063] In a preferred embodiment, step S1 includes:

[0064] S11: Obtain hospital level, annual PCI surgery volume percentage, and response time to determine hospital qualification indicators.

[0065] Specifically: Hospital qualification indicators, which can be represented as:

[0066] ;

[0067] in, For patients The matching score of Hospital H in terms of hospital qualification indicators The weighting coefficients for hospital qualification indicators are optional. , The hospital's grade score is 0-1. The number of PCI procedures per year This is a reference surgical volume (1.5 times the regional average). For green channel response time, The maximum acceptable response time is 90 minutes. , , These are the weighting factors for hospital level, percentage of surgeries, and corresponding time period, respectively; Example: , In this embodiment, the hospital's qualification level reflects its comprehensive diagnostic and treatment capabilities, the annual number of PCI surgeries reflects its technical proficiency, and the response time reflects its emergency treatment efficiency, thus fully reflecting the hospital's qualification indicators.

[0068] S12: Obtain information on the types of illnesses the patient has, the severity of each illness, and the hospital's ability to treat each illness, and determine the patient's condition indicators.

[0069] Specifically: Patient's condition indicators, which can be represented as:

[0070] ;

[0071] in, For patients The fit score between hospital H and the patient's condition indicators The weighting coefficients for patient condition indicators can be selected. , For the number of disease types, The hospital's ability to handle disease level d. The severity of the patient's condition (d). The weighting factor for disease d can be determined based on disease risk stratification.

[0072] S13: Obtain the estimated arrival time and straight-line distance between the patient and the hospital, as well as the maximum acceptable time and maximum acceptable distance for patient transfer, to determine geographic accessibility indicators.

[0073] ;

[0074] in, For patients The fit score of Hospital H on the geographical accessibility index The weighting coefficient for the geographical accessibility indicator is optional. , For the estimated arrival time, The maximum acceptable time is 120 minutes. The straight-line distance. The maximum acceptable distance is 100 kilometers.

[0075] S14: Obtain information on the strength of the patient's past treatment relationship with the hospital, the convenience of follow-up, and the suitability of rehabilitation resources to determine indicators of treatment continuity.

[0076] Specifically, the treatment continuity indicator can be represented as:

[0077] ;

[0078] in, For patients The fit score with hospital H on the continuity of treatment indicator. The weighting coefficients for the continuous treatment indicators can be selected. , The strength of the previous medical relationship (0-1 points). For ease of follow-up (0-1 point). The degree of fit for rehabilitation resources (0-1 points). These are weighting factors for the severity of the doctor-patient relationship, ease of follow-up, and suitability of rehabilitation resources; Example: In this embodiment, prior medical history is the core factor in ensuring the continuity of treatment, the suitability of rehabilitation resources is crucial for long-term prognosis, and the convenience of follow-up, as an auxiliary factor, can fully reflect the continuity of the patient's subsequent treatment.

[0079] S15: Based on hospital qualification indicators, patient condition indicators, geographical accessibility indicators, and treatment continuity indicators, construct a patient-hospital fit assessment model and determine the patient-hospital fit score by weighting.

[0080] Specifically, the patient-hospital fit assessment model is represented as follows:

[0081] ;

[0082] in, The fit score between patient P and hospital H is calculated. For the first The weighting coefficients of each indicator For patients Fit score with hospital H on the k-th indicator; example ,satisfy .

[0083] S2: Based on the triplet sample set, patient data is mapped to patient node feature vectors and bound to patient nodes, and hospital data is mapped to hospital node feature vectors and bound to hospital nodes; a heterogeneous network graph of the medical consortium containing patient nodes, hospital nodes, patient-hospital matching edges, and hospital-hospital collaboration edges is constructed.

[0084] Specifically, patient data can be sourced from patient medical records, consultation records, risk assessment results, etc., to extract several features, including disease risk (disease type, severity, complications), individual risk (age, mobility, underlying diseases), treatment risk (allergy history, surgical contraindications), etc., mapped to patient node feature vectors and bound to patient nodes. Specifically, patient node IDs can be configured: prefix P + sequence number, such as P001, P002, ..., Pn; the compiled patient node feature vectors are mapped one-to-one with patient nodes to form a patient node feature dictionary; the optional data format is set as: [Patient ID, feature vector dimension 1, feature vector dimension 2, ..., feature vector dimension N]. Example: P001, 0.8 (severe myocardial infarction), 0.7 (65 years old), 0.3 (mobility impairment), 0.0 (no allergies).

[0085] Hospital data can be sourced from 5 years of hospital operation data, qualification data, personnel data, configuration data, etc., to extract several features, including qualification features (specialty certification, technical access), resource features (equipment configuration, doctor titles), efficiency features (referral response time, rehabilitation follow-up capabilities), etc., which are mapped to hospital node feature vectors and bound to hospital nodes. Specifically, the hospital node ID can be configured as a prefix H + sequence number, such as H001, H002, ..., Hm; the compiled hospital node feature vectors are mapped one-to-one with hospital nodes to form a hospital node feature dictionary; the optional data format is set as: [Hospital ID, feature dimension 1, ..., feature dimension M], example: H003, 0.9 (chest pain center), 0.8 (PCI technology), 0.7 (percentage of associate senior physicians).

[0086] The patient-hospital matching edge can be constructed based on the association between the patient's condition and the hospital's treatment capabilities. It can optionally include a list of hospitals within the medical consortium that can treat the patient; for example, a myocardial infarction patient might only be matched with hospitals that have a chest pain center. Example: For each patient, hospitals within the medical consortium with treatment capabilities are selected based on their condition; example, P001 myocardial infarction patient → H001, H003, H006), constructing an edge index in COO format.

[0087] Hospital-to-hospital collaboration relationships can be built based on internal referrals or technical collaborations within the medical consortium, such as superior-subordinate referrals or peer-to-peer technical collaborations within the consortium. For example, collaboration relationships are extracted from the medical consortium's operational data: H001→H005: superior referral to primary care; H002→H003: peer-to-peer technical collaboration; converted to integer indices to construct edge indices.

[0088] S3: Input the heterogeneous network graph of the medical consortium into a preset graph attention network (GAT), calculate the attention weights of the nodes through the graph attention network, and perform weighted aggregation of the node features to obtain the node aggregated features; calculate and output the patient-hospital fit score based on the node aggregated features;

[0089] (1) Construction of graph attention network:

[0090] Specifically, the graph attention network is a two-layer cascaded graph attention network (GAT), which includes a cascaded first-layer graph attention module, a second-layer graph attention module, and an fitness calculation module;

[0091] The first layer graph attention module takes the node feature vectors of the heterogeneous network graph of the medical consortium as input, calculates the attention weights between adjacent nodes through a multi-attention head mechanism, and aggregates them to obtain the first layer aggregated features.

[0092] The second-layer graph attention module takes the aggregated features of the first layer as input and repeats the attention weight calculation and feature aggregation process to obtain the second-layer aggregated features.

[0093] The fit calculation module calculates the patient-hospital fit score based on the second-layer aggregated features.

[0094] More specifically:

[0095] A: The first layer graph attention module includes: feature input unit, attention weight calculation unit, and weighted aggregation unit;

[0096] The input end of the feature input unit is connected to the output end of the node feature vector of the heterogeneous network graph of the medical consortium, and is used to receive and verify the feature vectors of patient nodes and hospital nodes.

[0097] The input end of the attention weight calculation unit is connected to the output end of the feature input unit. It has several parallel attention head branches built in. Each attention head branch establishes a calculation connection with all adjacent node pairs (patient-hospital node pairs, hospital-hospital node pairs) in the heterogeneous graph. It performs a linear transformation on the feature vectors of the source node and the target node through a linear transformation matrix. Then, it outputs the normalized node attention weights through the sequentially connected attention coefficient calculation subunit, LeakyReLU activation subunit, and softmax normalization subunit.

[0098] The input of the weighted aggregation unit is connected to the output of the attention weight calculation unit. Based on the node attention weight, the features of neighboring nodes are weighted and summed to output the first layer of aggregated features.

[0099] B: The second-layer graph attention module includes: an aggregated input unit, an attention weight recalculation unit, and a weighted aggregated output unit;

[0100] The input end of the aggregation input unit is unidirectionally connected to the output end of the weighted aggregation unit in the first layer graph attention module, and is used to receive the first layer aggregation features and complete the dimension verification.

[0101] The input of the attention weight recalculation unit is connected to the output of the aggregation input unit. The topology, number of attention heads, and connection relationships of the internal sub-units (linear transformation sub-unit, attention coefficient calculation sub-unit, LeakyReLU activation sub-unit, softmax normalization sub-unit) of the attention weight recalculation unit are consistent with those of the attention weight calculation unit of the first layer graph attention module. It is used to repeatedly perform attention weight calculation on the first layer aggregated features.

[0102] The input of the weighted aggregation output unit is connected to the output of the attention weight recalculation unit to output the second-layer aggregation features (node ​​aggregation features).

[0103] Preferably, in a graph attention network (GAT), the node attention weights are calculated as follows:

[0104] ;

[0105] in, For nodes and nodes Attention weights For attention coefficient vectors, , for nodes ,node ,node eigenvectors, For nodes The set of neighbors.

[0106] More specifically, the node output features are represented as:

[0107] ;

[0108] in, For nodes Updated feature vectors, For activation function, For nodes and nodes Attention weights It is a linear transformation matrix. Let j be the feature vector of node j. For nodes The set of neighbors.

[0109] More specifically, each graph attention layer is set to have 32 attention heads, the hidden layer dimension is 128, the activation function is LeakyReLU and the negative slope is set to 0.2.

[0110] C: Fit calculation module: including feature splicing unit, fully connected mapping unit, and score output unit;

[0111] The input of the feature splicing unit is connected in two ways to the patient node aggregation feature end and the hospital node aggregation feature end output by the second layer graph attention module, which are used to fuse the aggregation features of the two types of nodes, namely patients and hospitals.

[0112] The input of the fully connected mapping unit is connected to the output of the feature concatenation unit, and the feature dimension transformation is completed through multiple levels of fully connected layers;

[0113] The input of the scoring output unit is connected to the output of the fully connected mapping unit, and is used to map the transformed features into patient-hospital fit scores and output them.

[0114] Specifically, the fitness calculation module uses a fitness score prediction head, including a 3-layer fully connected network: 128→64→32→1; BatchNorm layer: accelerates training and improves stability; Dropout layer: dropoutrate=0.2; Linear activation: directly outputs a fitness score of 0-10, which can be optionally represented as:

[0115] ;

[0116] in, The predicted fit between patient P and hospital j. For the patient, using their feature vector as input, Let j be the hospital, and its feature vector be the input. This is a diagram of a heterogeneous network within a medical consortium.

[0117] (2) Training of graph attention network

[0118] Specifically, the total loss function can be determined by a weighted average of the main loss function and the balanced loss function; where: the main loss function, used for ranking loss that integrates graph structure information, is expressed as:

[0119] ;

[0120] in, For sorting loss, The number of training samples, As a balance factor, For patients ; For the hospital , For the hospital , For the predicted fit between patient i and hospital j, The predicted fit between patient i and hospital k. The graph structure regularization coefficient is used. The number of nodes in the graph. For nodes The neighborhood group, For nodes and nodes Attention weights , for nodes ,node eigenvectors.

[0121] Optional, when Should be ranked Before ,otherwise ; Through nodes and nodes attention weights and feature difference items Incorporate graph structure information.

[0122] The balancing function, used to integrate the balancing losses of resource load, is expressed as:

[0123] ;

[0124] in, To balance the losses, For the total number of hospitals, The number of patients allocated to hospital h, For the capacity h of the hospital, To achieve the target resource utilization rate, the following options are available. ;

[0125] The total loss function, weighted, is as follows:

[0126] ;

[0127] in, For the total loss, For sorting loss, To balance the losses, The loss weighting coefficient can be selected. .

[0128] More specifically, training strategies may include layered training, including:

[0129] S1: Pre-training of hospital node features, which can optionally use 5 years of hospital operation data to train basic features; including: inputting the pre-training data into a multilayer perceptron (MLP), optimizing model parameters using the main loss function, and using a BatchNorm layer to accelerate convergence during training; if the hospital resource load difference coefficient is >0.3 (uneven load), then the balanced loss function is fused to construct the total loss function; after training, the feature vectors output by the intermediate layers of the MLP are extracted as the initial features of the hospital nodes; the initial features of the hospital nodes are bound to the hospital nodes in the heterogeneous network graph of the medical consortium as fixed basic features for subsequent training;

[0130] S2: Graph structure training, with optional fixed node features, learning the weights of relationships between hospitals; including: inputting the heterogeneous network graph of the medical consortium into the graph attention network, using labeled patient-hospital fit scores and inter-hospital referral priority ranking as supervision, and using the main loss function to calculate the error between the predicted ranking and the actual ranking; simultaneously introducing real-time hospital resource load data to construct a balanced loss function that integrates resource load; based on gradient backpropagation of the total loss function, only updating the parameters of the attention weight calculation unit, learning the association weights of patient-hospital matching edges and hospital-hospital collaboration edges in the heterogeneous network graph of the medical consortium, as well as the fit adjustment rules under different hospital resource loads, and outputting the pre-trained graph attention network;

[0131] S3: End-to-end training, with optional joint optimization of node features and graph structure; including: unlocking the MLP parameters corresponding to the hospital node initialization features frozen in the first stage, as well as the parameters of all modules in the graph attention network (including the first / second layer graph attention module and the fitness calculation module); using a small learning rate (1e-5) to optimize the parameters of the entire network, retaining the attention weight parameters learned in the second stage during training, and still using the main loss function that integrates graph structure information (to enhance the rationality of referral path ranking). The balance loss function adds a dynamic adjustment term for resource load (real-time access to hospital bed and equipment load data), and integrates L2 regularization loss to suppress overfitting. If the total loss of the validation set does not decrease after 10 consecutive rounds, training is stopped, and the current optimal model parameters are saved as the final training model parameters.

[0132] More specifically, the optimizer can be the Adam optimizer, with the following parameters: initial learning rate: 5e-5; learning rate scheduling: decay by 0.5 times every 5 epochs; batch size: 16 (each batch contains 16 patients and their corresponding hospital sets).

[0133] More specifically: the negative sampling strategy can be optionally set to pair each positive hospital with 3 negative hospitals, with a hard negative case mining strategy: priority is given to hospitals that are highly similar to positive hospitals but have low fit scores.

[0134] S4. Determine referral decisions based on patient-hospital fit scores.

[0135] Specifically, hospitals whose fit scores meet a preset threshold can be selected as candidate referral hospitals. The optimal referral decision is then output based on the medical consortium's referral process. For example, the top-3 recommended hospitals, their ranking probabilities, fit scores, and reasons for recommendation are output. A referral decision support report is then generated, including a list of recommended hospitals, an analysis of each hospital's strengths and weaknesses, and real-time resource status.

[0136] In summary, this invention provides an intelligent referral decision-making method for collaborative matching between patients and hospitals. First, it constructs a triplet sample set integrating patient-hospital-fitness. Then, it builds a heterogeneous network graph of the medical consortium, including patients, hospitals, and two types of related edges. Combining graph attention networks to mine deep correlation features between nodes, the accuracy of patient-hospital fit assessment is significantly improved compared to traditional manual or single-feature matching methods, reducing mismatches and incorrect matching. Next, relying on the graph attention network, it automatically completes feature aggregation and fit calculation, quickly outputting the fit score between the patient and each candidate hospital. This replaces the cumbersome process of manual screening and evaluation, significantly shortening the time required for referral decisions, guiding patients to hospitals with matching treatment capabilities and resource availability, avoiding overcrowding of high-quality medical resources, improving resource utilization in primary hospitals, and facilitating the implementation of hierarchical medical treatment within medical consortia.

[0137] In a preferred embodiment, this invention addresses the pain points of traditional medical consortium referral decision-making, namely "reliance on human experience, single dimension of fit assessment, and disconnect between referral pathways and medical resources." It employs a technical solution of constructing a 2-layer GAT feature aggregation layer using a heterogeneous network graph, followed by hierarchical training and weighted loss optimization. This achieves multi-dimensional, high-precision, and interpretable intelligent referral decision-making, with at least the following technical effects:

[0138] 1. By constructing a compatibility assessment system for "patient characteristics and medical resource capabilities" from four dimensions—"hospital qualification indicators, patient condition indicators, geographical accessibility indicators, and treatment continuity indicators"—this system can fundamentally solve the subjective and one-sided problems of traditional referrals within medical consortia, achieving precise matching between patients and medical resources.

[0139] Referral decisions are shifting from "experience-driven" to "data-driven," improving matching accuracy: hospital qualification indicators anchor the hospital's hard power of "what diseases it can treat"; patient condition indicators (such as disease type, severity, and complications) clarify the patient's core needs of "what treatment they need." Direct matching of the two can avoid resource mismatches such as "primary hospitals being unable to handle complex cases" and "higher-level hospitals accepting ordinary chronic diseases," ensuring that every patient is directed to the optimal institution capable of diagnosing and treating their disease.

[0140] Balancing "medical needs" with "convenience of access to medical care" to improve patient experience and compliance.

[0141] Geographic accessibility indicators (such as commuting distance and transportation convenience) address the shortcomings of "only considering medical capabilities without considering actual accessibility"—especially for patients with chronic diseases, elderly patients, and patients with mobility difficulties. Prioritizing the matching of patients with closer facilities among hospitals of equivalent medical capabilities can significantly reduce patients' medical time and economic costs. The inclusion of this dimension can effectively improve patients' acceptance of referral programs and reduce treatment interruptions caused by "too far to seek medical care."

[0142] Strengthen the closed-loop management of two-way referrals within medical consortia to ensure continuity of treatment.

[0143] Treatment continuity indicators (such as the hospital's rehabilitation follow-up capabilities, collaboration mechanisms with higher and lower-level institutions, and the degree of medical record sharing) focus on the key need of patients to "receive continuous management after treatment." For example, after a higher-level hospital completes surgery or emergency treatment, the assessment system will prioritize matching patients with primary care hospitals that have rehabilitation capabilities and collaboration with the higher-level hospital, achieving a seamless connection between "acute treatment - recovery rehabilitation - long-term chronic disease management," and avoiding situations where patients are "stuck in higher-level hospitals" or "no one follows up after discharge."

[0144] Optimize the efficiency of medical resource allocation within the medical consortium and reduce overall operating costs.

[0145] The assessment system can guide patients with mild or chronic illnesses to remain at the primary care level, while accurately referring patients with complex or severe illnesses to higher-level hospitals, thus achieving the goal of "minor illnesses treated in the community and major illnesses treated within the medical consortium." For the medical consortium, this reduces the burden on higher-level hospitals while fully leveraging the advantages of primary care hospitals in chronic disease management and rehabilitation, improving the overall utilization rate of medical resources, and reducing unnecessary referral costs and medical insurance expenditures.

[0146] 2. The accuracy of fit assessment has been significantly improved, enhancing the scientific basis of decision-making.

[0147] Multi-dimensional feature fusion capability: By constructing a heterogeneous network graph of the medical consortium that includes patient nodes, hospital nodes, patient-hospital matching edges, and hospital-hospital collaboration edges, it breaks through the limitations of the traditional binary matching of "patient condition-hospital qualification" and integrates key dimension features such as hospital qualification indicators, patient condition indicators, geographical accessibility indicators, and treatment continuity indicators to achieve comprehensive adaptation between individual patient needs and hospital resource supply.

[0148] Attention mechanism accurately captures association weights: The 2-layer GAT network automatically learns the association importance of two types of node pairs, namely patient-hospital and hospital-hospital, through a multi-attention head mechanism (e.g., the matching weight of acute myocardial infarction patients and chest pain center hospitals is higher, and the collaborative edge weight between primary hospitals and tertiary hospitals is dynamically adjusted with the frequency of referral). Compared with traditional machine learning models (such as logistic regression and random forest), it can reduce the mean absolute error (MAE) of fitness prediction.

[0149] Layered training enhances the model's generalization ability: First, the basic features of hospital nodes are pre-trained using MLP, then the graph structure parameters are optimized by fixing the features, and finally, end-to-end joint fine-tuning is performed. This effectively solves the problem of "few hospital samples and sparse collaborative relationships" in medical consortium data, improves the model's fitness prediction accuracy in small sample scenarios, and avoids overfitting.

[0150] 3. Referral decision-making efficiency has been greatly improved, reducing the waste of medical resources.

[0151] Automated decision-making replaces manual screening: Traditional referral decisions require medical staff to manually check information such as hospital qualifications, bed availability, and treatment capabilities, with an average single referral decision taking more than 30 minutes. This invention automatically outputs the compatibility score (0-10 points) between the patient and each candidate hospital through a model, and directly recommends the optimal referral hospital and route, reducing the single decision time to within 1 minute and improving efficiency by more than 95%.

[0152] Optimization of medical resource allocation: The model learns the attention weights of hospital-hospital collaboration edges, and prioritizes recommending referral paths with close collaboration within the medical consortium (such as referral from a tertiary hospital to a primary hospital for rehabilitation after diagnosis), reducing the proportion of cross-medical consortium referrals, increasing the referral rate within the medical consortium, and reducing the time and economic costs of patient referrals.

[0153] 4. The model is highly interpretable and meets the requirements for clinical application.

[0154] Attention weight visualization supports decision tracing: The node attention weights output by the model can directly explain "the core reason why a patient is suitable for a hospital" - such as "patient P001 and hospital H003 have a compatibility score of 8.5 points, of which the attention weight of the hospital's chest pain center qualification accounts for 60% and the response time weight accounts for 20%", which solves the pain point of the "black box" of traditional deep learning models and makes it easier for medical staff to understand and accept the decision results.

[0155] Layered training and loss function design ensure the stability of results: Through the layered training strategy of "pre-training-graph structure training-end-end optimization" and the weighted loss function of "main loss + balance loss", the problem of imbalance between the number of patient and hospital samples and large differences in the distribution of data in different levels of hospitals is effectively alleviated. It can reduce the standard deviation of the prediction results of the model under different disease types (acute / chronic / rehabilitation) and different medical consortium levels (community-secondary-tertiary), and the stability is significantly better than the model trained in one end-to-end.

[0156] 5. Improved collaboration efficiency within medical consortia facilitates the implementation of hierarchical medical services.

[0157] Facilitating the formation of a two-way referral closed loop: Based on the hospital-hospital collaboration edge features, the model not only supports "referral from primary care to higher-level hospitals", but also recommends "referral from higher-level hospitals to lower-level hospitals" according to the patient's needs at the recovery stage, helping medical consortia to form a hierarchical diagnosis and treatment closed loop.

[0158] Dynamically adapt to changes in medical consortium resources: When a medical consortium adds a hospital, adjusts its collaborative relationships, or updates hospital resources (such as adding an intensive care unit or expanding specialty services), it is only necessary to update the node features or edge relationships of the heterogeneous network graph. There is no need to retrain the model. It can quickly adapt to resource changes through fine-tuning, which can reduce the cost of model iteration and maintenance.

[0159] On the other hand, the present invention also provides a computer storage medium storing executable program code; the executable program code is used to execute any of the above-mentioned intelligent referral decision methods.

[0160] On the other hand, the present invention also provides a terminal device, including a memory and a processor; the memory stores program code that can be executed by the processor; the program code is used to execute any of the above-mentioned intelligent referral decision methods.

[0161] For example, the program code can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the program code in the terminal device.

[0162] The terminal device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the terminal device may also include input / output devices, network access devices, buses, etc.

[0163] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0164] The memory can be an internal storage unit of the terminal device, such as a hard drive or RAM. The memory can also be an external storage device of the terminal device, such as a plug-in hard drive, SmartMediaCard (SMC), Secure Digital (SD) card, or FlashCard. Furthermore, the memory can include both internal and external storage units of the terminal device. The memory is used to store the program code and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output.

[0165] The aforementioned computer storage medium and terminal device are created based on the aforementioned intelligent referral decision-making method. Their technical functions and beneficial effects will not be elaborated here. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0166] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. An intelligent referral decision-making method for collaborative matching of patients and hospitals, characterized in that, include: Acquire patient and hospital data, label patient-hospital fit scores, and construct a triplet sample set integrating patient-hospital fit; the patient-hospital fit score is determined based on hospital qualification indicators, patient condition indicators, geographical accessibility indicators, and treatment continuity indicators. Based on the triplet sample set, patient data is mapped to patient node feature vectors and bound to patient nodes, and hospital data is mapped to hospital node feature vectors and bound to hospital nodes; a heterogeneous network graph of the medical consortium containing patient nodes, hospital nodes, patient-hospital matching edges, and hospital-hospital collaboration edges is constructed. The heterogeneous network graph of the medical consortium is input into a preset graph attention network. The attention weights of the nodes are calculated through the graph attention network, and the node features are weighted and aggregated to obtain the node aggregated features. The patient-hospital fit score is calculated and output based on node aggregation features; the graph attention network is a two-layer cascaded graph attention network, including a cascaded first-layer graph attention module, a second-layer graph attention module, and a fit calculation module. The first layer graph attention module takes the node feature vectors of the heterogeneous network graph of the medical consortium as input, calculates the attention weights between adjacent nodes through a multi-attention head mechanism, and aggregates them to obtain the first layer aggregated features. The second-layer graph attention module takes the aggregated features of the first layer as input and repeats the attention weight calculation and feature aggregation process to obtain the second-layer aggregated features. The fit calculation module calculates the patient-hospital fit score based on the second-layer aggregated features; Referral decisions are made based on patient-hospital fit scores.

2. The intelligent referral decision-making method according to claim 1, characterized in that, The steps for acquiring patient and hospital data and labeling patient-hospital fit scores include: Obtain information on hospital level, annual PCI procedure volume percentage, and response time to determine hospital qualification indicators; To obtain information on the types of illnesses the patient has, the severity of each illness, and the hospital's ability to treat each illness, and to determine the patient's condition indicators; Obtain the estimated arrival time and straight-line distance between the patient and the hospital, as well as the maximum acceptable time and maximum acceptable distance for patient transfer, to determine geographical accessibility indicators; To obtain information on the strength of the patient's past treatment relationship with the hospital, the convenience of follow-up, and the suitability of rehabilitation resources, and to determine indicators of treatment continuity; Based on hospital qualification indicators, patient condition indicators, geographical accessibility indicators, and treatment continuity indicators, a patient-hospital fit assessment model is constructed, and a weighted score is determined for the patient-hospital fit.

3. The intelligent referral decision-making method according to claim 1, characterized in that, The first layer graph attention module includes: a feature input unit, an attention weight calculation unit, and a weighted aggregation unit; The input end of the feature input unit is connected to the output end of the node feature vector of the heterogeneous network graph of the medical consortium, and is used to receive and verify the feature vectors of patient nodes and hospital nodes. The input end of the attention weight calculation unit is connected to the output end of the feature input unit. It has several parallel attention head branches built in. Each attention head branch establishes a calculation connection with all adjacent node pairs in the heterogeneous graph. It performs a linear transformation on the feature vectors of the source node and the target node through a linear transformation matrix. Then, it outputs the normalized node attention weights through the sequentially connected attention coefficient calculation subunit, LeakyReLU activation subunit, and softmax normalization subunit. The input of the weighted aggregation unit is connected to the output of the attention weight calculation unit. Based on the node attention weight, the features of neighboring nodes are weighted and summed to output the first layer of aggregated features.

4. The intelligent referral decision-making method according to claim 3, characterized in that, The second-layer graph attention module includes: an aggregated input unit, an attention weight recalculation unit, and a weighted aggregated output unit; The input end of the aggregation input unit is unidirectionally connected to the output end of the weighted aggregation unit in the first layer graph attention module, and is used to receive the first layer aggregation features and complete the dimension verification. The input of the attention weight recalculation unit is connected to the output of the aggregation input unit. The topology, number of attention heads, and connection relationship of the internal sub-units of the attention weight recalculation unit are the same as those of the attention weight calculation unit of the first layer graph attention module. It is used to repeatedly perform attention weight calculation on the first layer aggregated features. The input of the weighted aggregation output unit is connected to the output of the attention weight recalculation unit to output the second layer of aggregated features.

5. The intelligent referral decision-making method according to claim 4, characterized in that, Adaptability calculation module: It includes a feature concatenation unit, a fully connected mapping unit, and a score output unit; The input of the feature splicing unit is connected in two ways to the patient node aggregation feature end and the hospital node aggregation feature end output by the second layer graph attention module, which are used to fuse the aggregation features of the two types of nodes, namely patients and hospitals. The input of the fully connected mapping unit is connected to the output of the feature concatenation unit, and the feature dimension transformation is completed through multiple levels of fully connected layers; The input of the scoring output unit is connected to the output of the fully connected mapping unit, and is used to map the transformed features into patient-hospital fit scores and output them.

6. The intelligent referral decision-making method according to any one of claims 1 to 5, characterized in that, Note that during the training of the network, the total loss function is determined by a weighted average of the main loss function and the balanced loss function; where: The main loss function, used for ranking by fusing graph structure information, is expressed as: ； in, For ranking loss, The number of training samples, As a balance factor, For patients ; For the hospital , For the hospital , For the predicted fit between patient i and hospital j, The predicted fit between patient i and hospital k. The graph structure regularization coefficient is used. The number of nodes in the graph. For nodes The neighborhood group, For nodes and nodes Attention weights , for nodes ,node eigenvectors; The balancing function, used to integrate the balancing losses of resource load, is expressed as: ； in, To balance the losses, For the total number of hospitals, The number of patients allocated to hospital h, For the capacity h of the hospital, Target resource utilization rate; The total loss function, weighted, is as follows: ； in, For the total loss, For ranking loss, To balance the losses, This is the loss weighting coefficient.

7. The intelligent referral decision-making method according to claim 6, characterized in that, The training process of the graph attention network includes: pre-training of hospital node features: The pre-trained data is input into the multilayer perceptron, and the model parameters are optimized using the main loss function. During training, the BatchNorm layer is used to accelerate convergence. If the hospital resource load difference coefficient is greater than a set threshold, the balanced loss function is fused to construct the total loss function. After training, the feature vectors output by the intermediate layers of the MLP are extracted as the initial features of the hospital nodes. The initial features of hospital nodes are bound to the hospital nodes in the heterogeneous network graph of the medical consortium, serving as fixed basic features for subsequent training.

8. The intelligent referral decision-making method according to claim 7, characterized in that, The training process for graph attention networks also includes: graph structure training: The heterogeneous network graph of the medical consortium is input into the graph attention network. The labeled patient-hospital fit scores and inter-hospital referral priority ranking are used as supervision. The error between the predicted ranking and the actual ranking is calculated using the main loss function. At the same time, real-time resource load data of hospitals is introduced to construct a balanced loss function that integrates resource load. Based on the gradient backpropagation of the total loss function, only the parameters of the attention weight calculation unit are updated. The association weights of patient-hospital matching edges and hospital-hospital collaboration edges in the heterogeneous network graph of the medical consortium are learned, as well as the fit adjustment rules under different hospital resource loads. The preliminarily trained graph attention network is output.

9. The intelligent referral decision-making method according to claim 8, characterized in that, The training process for graph attention networks also includes: end-to-end training: Unlock the MLP parameters corresponding to the hospital node initialization features frozen during the pre-training phase, as well as the parameters of all modules in the graph attention network; optimize the parameters of the entire network with a small learning rate; retain the attention weight parameters learned during the graph structure training phase during training; the main loss function still adopts the main loss function that integrates graph structure information; add a resource load dynamic adjustment term to the balancing loss function; and integrate L2 regularization loss to suppress overfitting until the iteration termination condition is reached, at which point training stops and the current optimal model parameters are saved as the final training model parameters.

10. A computer storage medium, characterized in that, It stores executable program code; the executable program code is used to execute the intelligent referral decision method according to any one of claims 1 to 9.

Images