Internet hospital grading diagnosis and treatment and resource scheduling method and system

Abstract

The invention discloses an internet hospital hierarchical diagnosis and treatment and resource scheduling method and system, and the method comprises the following steps: responding to a diagnosis and treatment request sent by a user, and obtaining diagnosis and treatment data containing the chief complaint information of a patient; inputting the diagnosis and treatment data into an illness state evaluation model for feature extraction, and outputting a feature vector representing the illness state of the patient; the feature vector at least comprises a first dimension parameter representing an emergency degree of an illness state and a second dimension parameter representing a complexity degree of the illness state; determining a target resource node in a hierarchical medical resource database based on the feature vector; and generating a resource scheduling instruction according to the target resource node, and establishing a communication link between the user terminal and a server corresponding to the target resource node. According to the method, a two-dimensional evaluation system of urgency and complexity is introduced, and a dynamic threshold adjustment mechanism based on supply and demand prediction is combined, so that the utilization efficiency of medical resources and the accuracy of triage matching are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical resource allocation, and particularly relates to an internet hospital hierarchical diagnosis and treatment and resource scheduling method and system. BACKGROUND

[0002] With the deep integration of mobile Internet and medical technology, Internet hospitals have become an important way to alleviate the pressure of offline medical treatment and optimize the allocation of medical resources. In the operation of Internet hospitals, how to reasonably allocate a large number of patient consultation requests to limited medical resources to achieve hierarchical diagnosis and treatment and up-down linkage is a core challenge currently faced by the industry.

[0003] Most existing Internet medical platforms use patient self-registration or simple keyword triage mode. The former relies on the patient's self-judgment, which can easily lead to the occupation of specialist resources by light patients or the delay of treatment by severe patients due to wrong department registration. The latter, although introducing intelligent means, is often limited to single rule matching.

[0004] In order to solve the above problems, some AI-based pre-trial diversion schemes have appeared in the prior art. The Chinese patent application with publication number CN121096569A discloses a hierarchical diagnosis and treatment resource allocation system based on AI priority pre-trial mechanism. The system mainly includes a user input module, an AI intelligent pre-diagnosis module, a resource allocation module, and a referral guidance module. Specifically, the system receives the initial information of the patient (such as symptoms, medical history, and examination data) through an AI applet, uses the AI intelligent pre-diagnosis module to extract the disease index through multiple rounds of questioning, and inputs the index into a pre-set risk stratification model to determine the severity / urgency index (level number 1-5) of the patient. The core logic is that patients with a severity / urgency index greater than a disease threshold (such as level number 4) are determined as high-risk, and a special number source pool reserved for experts is opened for them; while patients with a lower index (such as level number 2) are guided to general outpatient or Internet platform. In addition, the system also combines a time control unit (release number source if not paid within a certain time) and a risk notification module, and introduces indication judgment rules (such as tumor size and stage) and critical value referral prompt rules in the risk stratification model.

[0005] However, the above prior art still has the following limitations in actual application: the evaluation dimension is single, and there is a lack of independent consideration of complexity, resulting in insufficient matching accuracy. The threshold setting is static and rigid, and cannot adapt to the dynamic tidal effect of medical resources. There is a lack of fine-grained sub-specialty matching based on vector space, making it difficult to accurately match specific patients to specific experts who best match their pathological characteristics through simple stratification index, and there is still room for improvement in the utilization efficiency of medical resources and the effect of diagnosis and treatment. SUMMARY

[0006] The application provides an internet hospital hierarchical diagnosis and treatment and resource scheduling method and system, aiming to solve the problems of single evaluation dimension, threshold setting unable to cope with dynamic load fluctuations, and lack of fine sub-specialty matching in the existing triage system, resulting in low utilization of medical resources and insufficient triage accuracy.

[0007] To solve the above technical problems, on the one hand, the application provides an internet hospital hierarchical diagnosis and treatment and resource scheduling method, comprising the following steps: In response to a diagnosis and treatment request sent by a user, diagnosis and treatment data containing patient complaint information are obtained; The diagnosis and treatment data are input into a disease assessment model for feature extraction, and a feature vector representing the patient's disease state is output; the feature vector at least includes a first dimension parameter representing the severity of the disease and a second dimension parameter representing the complexity of the disease; Based on the feature vector, a target resource node is determined in a hierarchical medical resource database; the hierarchical medical resource database contains multiple levels of resource nodes, and each resource node is associated with a level label and an ability attribute label; According to the target resource node, a resource scheduling instruction is generated, and a communication link is established between the user terminal and the server corresponding to the target resource node.

[0008] Preferably, the disease assessment model includes a natural language processing subnetwork and a multi-modal fusion subnetwork; The feature extraction specifically includes: The natural language processing subnetwork is used to perform semantic understanding and entity recognition on the text information in the diagnosis and treatment data, extract symptom keywords and time dimension features; The multi-modal fusion subnetwork is used to process image data or physiological parameter data in the diagnosis and treatment data to generate an auxiliary feature vector; The symptom keywords, time dimension features and auxiliary feature vectors are weighted and fused to output the first dimension parameter and the second dimension parameter.

[0009] Preferably, the generation logic of the first dimension parameter and the second dimension parameter is as follows: The first dimension parameter is obtained by comprehensive calculation based on the matching weight of the acute symptom library, the abnormal amplitude of the vital sign data, and the attenuation function of the symptom duration, and is used to represent the urgency of the patient's need for medical intervention; The second dimension parameter is obtained by comprehensive calculation based on the number of organ systems involved in the identified symptoms, the correlation degree between the past medical history and the current symptoms, and the hit rate of rare disease characteristic words, and is used to represent the diagnostic difficulty of the disease.

[0010] Preferably, the method for determining the target resource node is: constructing a third dimension feature vector containing patient preference factors, which are generated based on historical evaluation data of the user, historical doctor records, and preset attribute screening conditions; mapping the first dimension parameter, the second dimension parameter, and the third dimension feature vector to a multi-dimensional vector space; calculating the cosine similarity or Euclidean distance between the patient vector and each candidate resource node vector in the multi-dimensional vector space; selecting N resource nodes with the maximum cosine similarity or the minimum Euclidean distance as the recommended target resource nodes.

[0011] Preferably, the method for determining the target resource node is: when the first dimension parameter is lower than a preset emergency threshold and the second dimension parameter is lower than a preset complexity threshold, determining a first level resource node with an automatic response attribute as the target resource node; when the first dimension parameter or the second dimension parameter meets a preset intervention condition, selecting the target resource node from a second level or higher level resource node according to the matching degree calculation result between the feature vector and the ability attribute label of the candidate resource node.

[0012] Preferably, the method for obtaining the preset emergency threshold and complexity threshold is: monitoring the current load state of each level resource node in the hierarchical medical resource database in real time; the load state includes the number of online resources and the current average queuing time; when the current average queuing time of a resource node of a target level exceeds a warning value, dynamically increasing the admission threshold of the corresponding level, or dynamically decreasing the admission threshold of the adjacent level resource node, to guide part of the diagnosis and treatment request to flow to the alternative level.

[0013] Preferably, the method for obtaining the preset emergency threshold and complexity threshold is: based on a sliding time window, collecting historical diagnosis and treatment request flow data in a preset period, and using a time series prediction algorithm to calculate the expected value of the request inflow rate in a future first period; obtaining the number of online resources of the current each level resource node and the historical average diagnosis and treatment time, and calculating the expected value of the resource digestion rate of the system at the current time; constructing a system pressure index according to the ratio of the expected value of the request inflow rate to the expected value of the resource digestion rate :

[0014] wherein, This is the projected rate of request inflow during the first time period. This represents the projected resource consumption rate of the system at the current moment. The threshold adjustment coefficient corresponding to the system pressure index is calculated using a nonlinear mapping function. The nonlinear mapping function is configured such that: when When within the preset steady-state range, Keep constant; when When the steady-state range is exceeded, It changes in an S-shaped curve; The baseline emergency threshold and the baseline complex threshold are weighted and corrected using the threshold adjustment coefficient to obtain the current emergency threshold and the complex threshold.

[0015] Preferably, the method for obtaining the preset emergency threshold and complexity threshold is as follows: Construct a probabilistic graphical model that includes observed variables and latent variables; the observed variables include the system pressure index at the current moment. and historical resource allocation efficiency indicators The latent variable is the threshold adjustment coefficient. ; A variational distribution network is established to use observed variables as input and output the variational distribution parameters of latent variables; the variational distribution parameters include at least the mean and variance. Training or updating variational distribution networks online based on the principle of maximizing the lower bound of evidence makes the variational distribution... Approximating the true posterior distribution of latent variables ; The system stress index will be calculated in real time during the inference phase. The input is fed into the trained variational distribution network, and the current threshold adjustment coefficient is determined based on the mean of the output. The baseline emergency threshold and the baseline complex threshold are weighted and corrected using a threshold adjustment coefficient.

[0016] Preferably, the resource scheduling instruction further includes a pre-diagnosis summary generation instruction; In response to the pre-consultation summary generation instruction, based on the process data of the feature extraction, a structured medical condition summary text is automatically generated using a generative language model and sent to the server interface of the target resource node through the communication link.

[0017] Preferably, the method further includes the following steps: After the communication link ends, acquire the diagnosis and treatment results data and user evaluation data; If the user rating data is lower than the preset satisfaction threshold, or if the diagnosis and treatment result data indicates that a misdiagnosis or referral has occurred, a negative feedback label will be generated. The disease assessment model is adversarially trained or fine-tuned using multimodal diagnostic and treatment data with negative feedback labels to correct the output weights of the feature vectors of similar multimodal diagnostic and treatment data in subsequent applications.

[0018] A second aspect of the present invention also proposes an internet hospital hierarchical diagnosis and treatment and resource scheduling system, said system being used to implement the method as described in the first aspect of the present invention, comprising: The data acquisition module is used to respond to the diagnosis and treatment requests sent by users and obtain diagnosis and treatment data containing the patient's chief complaint information; The feature extraction module is used to input the diagnosis and treatment data into the disease assessment model for feature extraction and output a feature vector representing the patient's disease status; the feature vector includes at least a first dimension parameter representing the urgency of the disease and a second dimension parameter representing the complexity of the disease. The resource matching module is used to determine the target resource node in the hierarchical medical resource database based on the feature vector. The scheduling and execution module is used to generate resource scheduling instructions based on the target resource node and establish a communication link between the user terminal and the server corresponding to the target resource node.

[0019] Compared with the prior art, the present invention has the following technical effects: 1. The hierarchical diagnosis and treatment and resource allocation method proposed in this invention introduces a second-dimensional parameter characterizing the complexity of the illness, upgrading the illness assessment from a single-track system to a dual-track system. This enables the system to identify complex cases that, while not life-threatening, require intervention from senior specialists and accurately route them to specialist resources. This avoids mild cases crowding out specialist appointments and also prevents complex cases from being delayed in general outpatient clinics, significantly improving the clinical rationality of triage.

[0020] 2. The hierarchical diagnosis and treatment and resource scheduling method proposed in this invention dynamically adjusts the intervention threshold using a nonlinear function by monitoring the load status in real time or predicting future supply and demand pressure indices based on time series data. During peak traffic periods, the system automatically raises the expert access threshold to prioritize critically ill patients; during off-peak periods, it automatically lowers the threshold, allowing more patients with moderate conditions to be matched with expert resources. This peak-shaving and valley-filling mechanism avoids resource waste during off-peak hours and system paralysis during peak hours, ensuring the robustness and service continuity of the system.

[0021] 3. The hierarchical diagnosis and resource scheduling method proposed in this invention maps patient characteristics, including preference factors, and resource node capabilities to the same multi-dimensional vector space, and performs matching by calculating cosine similarity or Euclidean distance. This achieves more accurate profile matching, thereby significantly improving the initial diagnosis rate and the patient's medical experience.

[0022] 4. The hierarchical diagnosis and treatment and resource scheduling method proposed in this invention collects post-diagnosis evaluation data and referral / misdiagnosis records, and constructs negative feedback labels to conduct adversarial training or fine-tuning of the model. The system can learn from each incorrect match, and with the accumulation of running time, its ability to identify atypical symptoms and the accuracy of resource matching will continuously optimize itself, overcoming the shortcomings of traditional rule-based systems with lagging updates. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the method described in this invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0025] Example 1 The method for hierarchical diagnosis and treatment and resource allocation in internet hospitals includes the following steps one through four: Step 1: Respond to the user's medical request and obtain medical data containing the patient's chief complaint information.

[0026] In this step, the user initiates a consultation request through a client installed on their terminal device or a mini-program on another platform. When the system detects that the user clicks on a relevant control in the interface, the front-end application calls the API to initialize a new consultation session and generates a unique session ID (Session_ID). To ensure data integrity, the system can use multi-turn guided dialogue or structured forms.

[0027] The diagnostic data does not only contain a single text description; to support subsequent accurate triage, the data is configured as a multimodal data packet, specifically containing at least one or more of the following data types: Text-based complaint data, which is natural language text directly entered by the user, such as "left lower abdominal pain, lasting for three days, accompanied by nausea"; For voice feature data, if the user selects voice input, the microphone of the terminal device will collect the user's voice stream. The original audio file can be uploaded directly to extract acoustic features, or it can be converted into a text copy using speech recognition technology on the local machine or in the cloud and uploaded together. Image / video evidence: The interface provides an attachment upload function, allowing users to take photos of the affected area, upload photos of past paper test reports, or record a video of abnormal gait. Physiological parameter data: The system requests to read data from wearable devices authorized by the user via Bluetooth or other health interfaces. Specific fields include, but are not limited to, average heart rate, heart rate variability, blood oxygen saturation, sleep duration, etc. in the 24 hours prior to initiating the request. With user authorization, the system automatically retrieves the user's electronic medical record tags from the hospital's internal database using structured historical data.

[0028] Those skilled in the art will understand that before sending data to the server, the terminal device performs preprocessing and encapsulation steps to ensure that the data can be read by the disease assessment model. Specific preprocessing includes data cleaning and data encapsulation. Data cleaning removes illegal characters from the text and compresses or standardizes the format of uploaded images (e.g., converting them to JPG format and adjusting the resolution to 1024x1024). The terminal then encapsulates the cleaned and dispersed information into a standardized JSON data packet.

[0029] The following is an example of the data packet structure constructed in this embodiment: { "request_id":"REQ_20251231_001", "user_id":"U_987654", "timestamp":1704067200, "payload":{ "text_content":"Feeling tightness in my chest and shortness of breath", "audio_url":"https: / / oss.sample.com / audio / 001.mp3", "images":[ "https: / / oss.sample.com / img / ecg_report.jpg" ], "vitals":{ "heart_rate":110, "spo2":94 }, "history_tags":["Diabetes", "Smoker"] }, "device_info":"iPhone 15,iOS 18.0" } Furthermore, considering the privacy of medical data, the terminal establishes a secure connection with the server via HTTPS protocol (based on TLS 1.3) and sends the aforementioned JSON data packet to the gateway interface. Upon receiving the HTTP POST request, the server's communication module first performs authentication. After successful authentication, it parses the JSON data packet, extracting and saving the text, image links, and physiological parameters, preparing them for input into the subsequent "disease assessment model" for feature extraction.

[0030] Step Two: Input the diagnostic data into the disease assessment model for feature extraction, and output a feature vector representing the patient's condition status. The feature vector includes at least a first-dimensional parameter representing the urgency of the condition and a second-dimensional parameter representing the complexity of the condition. This feature extraction is performed in the server-side AI inference engine.

[0031] The disease assessment model is constructed as a multi-branch deep neural network architecture, which mainly consists of two parallel processing branches: a Natural Language Processing (NLP) sub-network and a multimodal fusion sub-network, as well as a fully connected fusion layer at the end. The Natural Language Processing sub-network can be built based on BERT (Bidirectional Encoder Representations from Transformers) or Bi-LSTM+CRF (Bidirectional Long Short-Term Memory Network + CRF) architecture.

[0032] The feature extraction specifically includes the following steps S21 to S23: S21: The natural language processing sub-network is used to perform semantic understanding and entity recognition on the text information in the diagnostic data, extracting symptom keywords and time-dimensional features. For example, for the text "severe chest pain lasting half an hour", the model uses the Named Entity Recognition (NER) layer to locate severe chest pain as a symptom keyword and half an hour as a time-dimensional feature; then, it further maps severe chest pain to the standard term "acute chest pain" in the medical knowledge graph and converts it into a high-dimensional vector representation.

[0033] S22: The multimodal fusion sub-network is used to process image data or physiological parameter data in the diagnostic data to generate an auxiliary feature vector. This sub-network includes a convolutional neural network (CNN) for image processing and a recurrent neural network (RNN) for processing numerical sequences. If the input contains an image of the affected area, the convolutional neural network branch extracts texture, color, and shape features; if the input contains physiological parameters such as heart rate and blood oxygenation, the normalization layer converts them into numerical features; these features are then concatenated to generate an auxiliary feature vector.

[0034] S23: The symptom keywords, time-dimensional features, and auxiliary feature vectors are weighted and fused to output the first-dimensional parameter and the second-dimensional parameter. Specifically, the system inputs the semantic vector output by NLP and the auxiliary feature vector output by multimodal processing into the attention layer, automatically assigning weights to different modalities (for example, when rashes are mentioned, image features are given higher weights), and finally outputs two normalized values ​​through the regression layer: the first-dimensional parameter and the second-dimensional parameter.

[0035] To ensure the interpretability and accuracy of the output parameters, the model's internal or post-processing modules strictly adhere to the following logic for calculations: The first dimension parameter is calculated based on the matching weights of the acute symptom database, the abnormal amplitude of vital sign data, and the decay function of symptom duration, and is used to characterize the urgency of the patient's need for medical intervention. The following is a specific algorithm example of this embodiment.

[0036] Let the first dimension parameter be :

[0037] in, The matching weight for the acute symptom database is as follows: The system has a built-in acute symptom database (such as entries related to myocardial infarction, stroke, and respiratory failure). If the extracted keywords match entries in the database, then... Take the higher value (e.g., 0.9), otherwise take the lower value; This refers to the abnormal range of vital signs data; for example, the greater the deviation of real-time heart rate from the standard resting heart rate, the larger this value. This is a decay function for symptom duration: for acute illnesses such as anaphylactic shock, the shorter the onset time, the more... The smaller the value, the higher the urgency, and the larger the function value. These are the corresponding preset weighting coefficients. In a preferred embodiment, This ensures that the calculated first-dimensional parameters are normalized within a preset range.

[0038] The second dimension parameter is calculated based on the number of human organ systems involved in the identified symptoms, the correlation between past medical history and current symptoms, and the hit rate of rare disease feature words, and is used to characterize the diagnostic difficulty of the condition. The following is a specific algorithm example of this embodiment.

[0039] Let the second dimension parameter be :

[0040] in, To determine the number of human organ systems involved in the identified symptoms: the model uses knowledge graph reasoning to determine if the symptoms involve the circulatory system, respiratory system, and nervous system. This indicates extremely high complexity; The correlation between past medical history and current symptoms: If the patient has a history of diabetes and the current chief complaint is foot ulcers, the correlation is high, suggesting possible complications and increased complexity. Hit rate of rare disease characteristic words: If uncommon professional terms or atypical symptom combinations appear in the text, this value increases, indicating that expert intervention is needed; These are the corresponding preset weight coefficients.

[0041] In one embodiment of this application, the above-mentioned weighting coefficient ( Based on the Delphi method or the analytic hierarchy process, fixed values ​​are pre-set and stored in the system parameter table by a group of medical experts for different departments.

[0042] In another preferred embodiment of this application, the above-mentioned weighting coefficient ( The coefficients are not fixed values, but rather learnable parameters in the fully connected layers of the disease assessment model. During the training phase of the model, supervised learning is performed using historical medical record data labeled with the actual grading results (i.e., Ground Truth, such as the urgency and complexity levels actually determined by doctors). The coefficients are iteratively updated through the backpropagation algorithm until the error between the feature vector output by the model and the actual grading results is minimized.

[0043] Step 3: Based on the feature vector, determine the target resource node in the hierarchical medical resource database; the hierarchical medical resource database contains resource nodes at multiple levels, and each resource node is associated with a level label and a capability attribute label.

[0044] Data structure example: Each node (Node_ID) in the database must contain at least the following fields: Tier_Tag: such as L0 (AI Assistant), L1 (Health Manager), L2 (General Practitioner), L3 (Specialist).

[0045] Capability_Tags: e.g., ['Cardiology', 'Pediatrics', 'Expert_in_Arrhythmia'].

[0046] Vector_Embedding (Capability Vector): A pre-trained feature vector representing the doctor's capabilities.

[0047] In an implementation scenario that emphasizes personalization and accuracy, the system executes the following logic: To achieve patient-centered scheduling, the system first constructs a third-dimensional feature vector containing patient preference factors. These patient preference factors are generated based on the user's historical evaluation data, historical doctor records, and preset attribute filtering conditions. Specifically, the system first queries a user profile database and converts user preferences into a numerical vector based on the user profile. .

[0048] Next, the system maps the first-dimensional parameter, the second-dimensional parameter, and the third-dimensional feature vector to a multi-dimensional vector space. For example: Construct a unified doctor-patient matching space, patient vector Constructed as:

[0049] in, Indicates preference factor, These represent the weight coefficients of the corresponding preference factors.

[0050] Calculate the cosine similarity or Euclidean distance between the patient vector and each candidate resource node vector in the multidimensional vector space; the following is an example of calculating cosine similarity.

[0051]

[0052] in, This represents the vector of the resource node (such as the doctor vector), and its purpose is to quantify the geometric proximity between patient needs and doctor capabilities.

[0053] The N resource nodes with the highest cosine similarity or the smallest Euclidean distance are selected as the recommended target resource nodes; for example, the top 3 doctors are selected to recommend to patients, or the request is directly routed to the online doctor with the highest Sim value.

[0054] In another implementation scenario that emphasizes efficiency and the principle of hierarchical medical treatment, the method for determining the target resource node is as follows: When the first dimension parameter is lower than a preset urgency threshold and the second dimension parameter is lower than a preset complexity threshold, the first-level resource node with automatic response attribute is identified as the target resource node. For example, when and If the system determines that the discomfort is minor, it will directly establish a conversation with the AI ​​intelligent assistant (Level 0) or online health manager (Level 1) without consuming doctor resources.

[0055] When either the first dimension parameter or the second dimension parameter meets the preset intervention conditions, the target resource node is selected from the resource nodes at the second or higher level based on the matching degree calculation result between the feature vector and the capability attribute labels of the candidate resource node. For example, when or When the conditions for expert intervention are triggered, the system locks the L3 level expert database and performs tag matching.

[0056] To address the peaks and troughs in medical requests, this embodiment does not use a fixed threshold but instead employs a dynamic threshold. This embodiment proposes three methods to dynamically obtain the emergency threshold and the complex threshold.

[0057] Method 1: The preset emergency threshold and complexity threshold are obtained as follows: The system monitors the current load status of resource nodes at each level of the hierarchical medical resource database in real time. This load status includes the number of online resources and the current average queuing time. When the current average queuing time of a resource node at a target level exceeds a warning threshold, the system dynamically increases the admission threshold for the corresponding level or dynamically decreases the admission threshold for resource nodes at adjacent levels to guide some treatment requests to alternative levels. For example, if a Level 2 general practitioner waits for more than 30 minutes, the system automatically increases the complexity threshold for Level 2 admission from 4 to 5 points, forcing patients with mild symptoms (4-5 points) to be diverted to Level 1 health managers, thus alleviating congestion.

[0058] Method 2, the preset emergency threshold and complexity threshold are obtained as follows: Based on a sliding time window, historical medical request traffic data within a preset time period is collected, and a time series prediction algorithm is used to calculate the expected request inflow rate in the first future time period. Using the ARIMA (Autoregressive Integrated Moving Average) model, based on the data of the past 30 minutes, the expected request inflow rate for the next 10 minutes is predicted. For example, it is expected that 20 new requests will be added per minute in the next 10 minutes.

[0059] Obtain the number of online resource nodes at each level and the historical average consultation time, and calculate the estimated resource consumption rate of the system at the current moment; for example: estimated resource consumption rate = number of online doctors / average consultation time.

[0060] A system stress index is constructed based on the ratio of the projected request inflow rate to the projected resource consumption rate. :

[0061] In the formula, This is the projected rate of request inflow during the first time period. This represents the projected resource consumption rate of the system at the current moment. The threshold adjustment coefficient corresponding to the system pressure index is calculated using a nonlinear mapping function. The nonlinear mapping function is configured such that: when When within the preset steady-state range, Keep constant; when When the steady-state range is exceeded, It exhibits an S-shaped curve variation; the S-shaped curve can be achieved using a variant of the Sigmoid function:

[0062] In the formula, the steady-state interval refers to... The state. When When significantly greater than 1, Rapidly rising and tending to saturate; when When significantly less than 1, Decrease; K is the preset maximum gain limit; k is the preset sensitivity coefficient; The steady-state critical point is a preset constant that represents the inflection point of the S-curve. It is usually set to 1.0 or slightly greater than 1.0, indicating that the system allows for slight queuing.

[0063] The baseline emergency threshold and baseline complex threshold are weighted and corrected using the threshold adjustment coefficient to obtain the current emergency threshold and complex threshold. The final threshold is based on the baseline threshold and... The product of.

[0064] Method 3, the preset emergency threshold and complexity threshold are obtained as follows: Construct a probabilistic graphical model that includes observed variables and latent variables; the observed variables include the system pressure index at the current moment. and historical resource allocation efficiency indicators The latent variable is the threshold adjustment coefficient. ; A variational distribution network is established to use observed variables as input and output the variational distribution parameters of latent variables; the variational distribution parameters include at least the mean and variance. Training or updating variational distribution networks online based on the principle of maximizing the lower bound of evidence makes the variational distribution... Approximating the true posterior distribution of latent variables ; The system stress index will be calculated in real time during the inference phase. The input is fed into the trained variational distribution network, and the current threshold adjustment coefficient is determined based on the mean of the output. The baseline emergency threshold and the baseline complex threshold are weighted and corrected using a threshold adjustment coefficient.

[0065] In specific implementation, this embodiment first defines a generation process: assuming the current resource scheduling efficiency (For example, whether the queue length is zero, or whether the doctor idle rate is reasonable) is determined by system pressure. and threshold adjustment coefficient It was decided jointly. Among them, These are considered latent variables. The goal is to solve for the posterior distribution. That is: when the current pressure is observed and expected performance In this case, What value should be taken?

[0066] Since the true posterior distribution is difficult to calculate directly, this embodiment constructs a variational distribution network. This network can be a multilayer perceptron or a recurrent neural network. The input to the variational distribution network is the current system stress index. And the past The state sequence within the time window. The output is an adjustment coefficient. The Gaussian distribution it follows has the following parameters: mean and diagonal covariance .

[0067]

[0068] In the formula, the NeuralNet function represents the processing of the variational distribution network.

[0069] To train the network, this embodiment optimizes the parameters by maximizing the Evidence Lower Bound (ELBO):

[0070] in, To optimize the objective function, which is the maximization objective during the system training process. Maximization This is equivalent to minimizing the difference between the variational distribution and the true posterior distribution, while maximizing the model's ability to interpret the observed data; the first term For reconstruction terms, used to encourage sampling. It can bring good scheduling efficiency , Indicates variational distribution The expected value of the second term; The KL divergence (Kullback-Leibler Divergence) is the variational posterior distribution. With prior distribution The KL divergence is used to constrain the predicted distribution. Do not deviate from the prior distribution. Too far.

[0071] Finally, online inference and reparameterized sampling are performed. In actual operation, when system pressure is monitored... Time: will Inputting the variational network yields... and To avoid system oscillations caused by randomness, this embodiment preferably uses the mean directly. As the current threshold adjustment coefficient In other embodiments, variance may also be utilized. As a confidence index; if The smaller value indicates that the model is very deterministic and can be applied directly and quickly. ;like A large value (exceeding the preset uncertainty threshold) indicates that an unprecedented extreme traffic pattern has been encountered, triggering a conservative strategy in the system, such as switching back to manual rules or imposing restrictions. The range of change can be adjusted to improve system security.

[0072] Step Four: Once the system determines the optimal target resource node, the server-side scheduling and control module executes this step. Specifically, the system generates a resource scheduling instruction based on the target resource node, establishing a communication link between the user terminal and the server corresponding to the target resource node. The resource scheduling instruction includes the target node's IP address, port number, and session authentication token. The system utilizes the WebRTC or WebSocket protocol to establish a low-latency, bidirectional data channel between the patient's terminal and the doctor's workstation software. This link supports the synchronous transmission of text, high-definition voice, and real-time video streams.

[0073] To improve patient reception efficiency, the system does not simply connect the two parties, but rather optimizes the initial consultation experience through AI intervention. In this embodiment, the resource scheduling instructions also include pre-consultation summary generation instructions; In response to the pre-consultation summary generation instruction, based on the process data of the feature extraction, a structured medical condition summary text is automatically generated using a generative language model and sent to the server interface of the target resource node through the communication link.

[0074] The data source includes the feature extraction process data extracted in step two, such as identified symptom entities, time entities, abnormal sign values, and risk labels. This structured data is then input into a pre-trained generative language model. The model uses prompting engineering templates to reorganize the fragmented data into a clinical summary conforming to the format of chief complaint-present illness-past history-suggestion. This summary text is prioritized and pushed to the doctor's workstation's floating window during the ringing phase before the video call is connected.

[0075] In this embodiment, the method further includes step five: After the communication link ends, the system acquires treatment outcome data and user evaluation data. The data includes user evaluation data and treatment outcome data. User evaluation data includes patient star ratings (1-5 stars) and text comments on the service provided via the app. Treatment outcome data includes the final diagnosis code and treatment method (prescription / hospitalization / referral) entered by the doctor into the system. The system also pays special attention to referral request records, as these often indicate that the initial triage was not accurate enough.

[0076] If the user rating data is lower than the preset satisfaction threshold, or if the diagnosis and treatment results data indicate that a misdiagnosis or referral has occurred, a negative feedback label is generated; a specific logic example is as follows.

[0077] Threshold judgment: If the user rating is ≤2 stars, the system automatically marks it as a low satisfaction sample; Error correction judgment: If the system initially assigns a patient to the gastroenterology department, but the doctor diagnoses the patient with inferior wall myocardial infarction and urgently transfers the patient to the cardiology department, the system identifies this dispatch as a department matching error and generates a misdiagnosis label.

[0078] The disease assessment model is adversarially trained or fine-tuned using multimodal diagnostic and treatment data with negative feedback labels to correct the output weights of feature vectors for subsequent applications of similar multimodal diagnostic and treatment data. Specific implementation method: The misdiagnosed medical data is added to the training set as a difficult sample. During model fine-tuning, the weight of the loss function is increased for this sample. The system adjusts the weight parameters in the neural network using the backpropagation algorithm. The system can also automatically generate adversarial examples with minor perturbations to test and improve the model's robustness to ambiguous cases, ensuring that the model does not develop significant classification bias due to minor descriptive differences.

[0079] Example 2 This embodiment is an internet hospital hierarchical diagnosis and treatment and resource scheduling system. The system is used to implement the method described in Embodiment 1, including: The data acquisition module is used to respond to the diagnosis and treatment requests sent by users and obtain diagnosis and treatment data containing the patient's chief complaint information; The feature extraction module is used to input the diagnosis and treatment data into the disease assessment model for feature extraction and output a feature vector representing the patient's disease status; the feature vector includes at least a first dimension parameter representing the urgency of the disease and a second dimension parameter representing the complexity of the disease. The resource matching module is used to determine the target resource node in the hierarchical medical resource database based on the feature vector. The scheduling and execution module is used to generate resource scheduling instructions based on the target resource node and establish a communication link between the user terminal and the server corresponding to the target resource node.

[0080] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for hierarchical diagnosis and treatment and resource allocation in internet hospitals, characterized in that, Includes the following steps: Responding to user requests for medical treatment, it obtains medical data containing the patient's chief complaint information; The diagnostic and treatment data are input into the disease assessment model for feature extraction, and a feature vector representing the patient's disease status is output. The feature vector includes at least a first-dimensional parameter representing the urgency of the illness and a second-dimensional parameter representing the complexity of the illness. Based on the feature vector, the target resource node is determined in the hierarchical medical resource database; the hierarchical medical resource database contains resource nodes at multiple levels, and each resource node is associated with a level label and a capability attribute label. Based on the target resource node, a resource scheduling instruction is generated, and a communication link is established between the user terminal and the server corresponding to the target resource node.

2. The method according to claim 1, characterized in that, The disease assessment model includes a natural language processing subnetwork and a multimodal fusion subnetwork; The feature extraction specifically includes: The natural language processing subnetwork is used to perform semantic understanding and entity recognition on the text information in the diagnosis and treatment data, and to extract symptom keywords and time dimension features. The multimodal fusion sub-network is used to process image data or physiological parameter data in diagnostic data to generate auxiliary feature vectors; The symptom keywords, time-dimensional features, and auxiliary feature vectors are weighted and fused to output the first dimension parameter and the second dimension parameter.

3. The method according to claim 1, characterized in that, The generation logic for the first dimension parameter and the second dimension parameter is as follows: The first dimension parameter is calculated based on the matching weight of the acute symptom database, the abnormal amplitude of vital sign data, and the decay function of symptom duration, and is used to characterize the urgency of the patient's need for medical intervention. The second dimension parameter is calculated based on the number of human organ systems involved in the identified symptoms, the correlation between past medical history and current symptoms, and the hit rate of rare disease characteristic words, and is used to characterize the diagnostic difficulty of the disease.

4. The method according to claim 1, characterized in that, The method for determining the target resource node is as follows: A third-dimensional feature vector containing patient preference factors is constructed; the patient preference factors are generated based on the user's historical evaluation data, historical doctor records, and preset attribute filtering conditions. Map the first dimension parameter, the second dimension parameter, and the third dimension feature vector to a multidimensional vector space; Calculate the cosine similarity or Euclidean distance between the patient vector and each candidate resource node vector in the multidimensional vector space; The N resource nodes with the highest cosine similarity or the smallest Euclidean distance are selected as recommended target resource nodes.

5. The method according to claim 1, characterized in that, The method for determining the target resource node is as follows: When the first dimension parameter is lower than the preset emergency threshold and the second dimension parameter is lower than the preset complexity threshold, the first-level resource node with automatic response attribute is determined as the target resource node. When the first dimension parameter or the second dimension parameter meets the preset intervention conditions, the target resource node is selected from the resource nodes of the second level or higher based on the matching degree calculation result between the feature vector and the capability attribute label of the candidate resource node.

6. The method according to claim 5, characterized in that, The preset emergency threshold and complex threshold are obtained as follows: The current load status of each level of resource nodes in the hierarchical medical resource database is monitored in real time; the load status includes the number of online resources and the current average queuing time. When the current average queuing time of a resource node at a certain target level exceeds the warning value, the admission threshold of the corresponding level is dynamically increased, or the admission threshold of resource nodes at adjacent levels is dynamically decreased, so as to guide some medical requests to be transferred to an alternative level.

7. The method according to claim 5, characterized in that, The preset emergency threshold and complex threshold are obtained as follows: Based on a sliding time window, historical medical request traffic data within a preset time period is collected, and a time series prediction algorithm is used to calculate the expected request inflow rate in the first future time period. Obtain the current number of online resource nodes at each level and the historical average treatment time, and calculate the estimated resource consumption rate of the system at the current moment; A system stress index is constructed based on the ratio of the projected request inflow rate to the projected resource consumption rate. : In the formula, This is the projected rate of request inflow during the first time period. This represents the projected resource consumption rate of the system at the current moment. The threshold adjustment coefficient corresponding to the system pressure index is calculated using a nonlinear mapping function. The nonlinear mapping function is configured such that: when When within the preset steady-state range, Keep constant; when When the steady-state range is exceeded, It changes in an S-shaped curve; The baseline emergency threshold and the baseline complex threshold are weighted and corrected using the threshold adjustment coefficient to obtain the current emergency threshold and the complex threshold.

8. The method according to claim 5, characterized in that, The preset emergency threshold and complex threshold are obtained as follows: Construct a probabilistic graphical model that includes observed variables and latent variables; the observed variables include the system pressure index at the current moment. and historical resource allocation efficiency indicators ; The latent variable is the threshold adjustment coefficient. ; A variational distribution network is established to use observed variables as input and output the variational distribution parameters of latent variables; the variational distribution parameters include at least the mean and variance. Training or updating variational distribution networks online based on the principle of maximizing the lower bound of evidence makes the variational distribution... Approximating the true posterior distribution of latent variables ; The system stress index will be calculated in real time during the inference phase. The input is fed into the trained variational distribution network, and the current threshold adjustment coefficient is determined based on the mean of the output. The baseline emergency threshold and the baseline complex threshold are weighted and corrected using a threshold adjustment coefficient.

9. The method according to claim 1, characterized in that, The method further includes the following steps: After the communication link ends, acquire the diagnosis and treatment results data and user evaluation data; If the user rating data is lower than the preset satisfaction threshold, or if the diagnosis and treatment result data indicates that a misdiagnosis or referral has occurred, a negative feedback label will be generated. The disease assessment model is adversarially trained or fine-tuned using multimodal diagnostic and treatment data with negative feedback labels to correct the output weights of the feature vectors of similar multimodal diagnostic and treatment data in subsequent applications.

10. An internet hospital hierarchical diagnosis and treatment and resource scheduling system, characterized in that, The system is used to implement the method as described in any one of claims 1-9, comprising: The data acquisition module is used to respond to the diagnosis and treatment requests sent by users and obtain diagnosis and treatment data containing the patient's chief complaint information; The feature extraction module is used to input the diagnosis and treatment data into the disease assessment model for feature extraction and output a feature vector representing the patient's disease status; the feature vector includes at least a first dimension parameter representing the urgency of the disease and a second dimension parameter representing the complexity of the disease. The resource matching module is used to determine the target resource node in the hierarchical medical resource database based on the feature vector. The scheduling and execution module is used to generate resource scheduling instructions based on the target resource node and establish a communication link between the user terminal and the server corresponding to the target resource node.

Images