Service dispute risk early warning method and system based on multi-dimensional data features

Abstract

This application discloses a service dispute risk early warning method and system based on multidimensional data features, relating to the field of risk prevention and control. The method includes: acquiring multi-source heterogeneous data in real time through a risk early warning platform; constructing a multidimensional feature index matrix based on the multi-source heterogeneous data; analyzing the multidimensional feature index matrix using a preset machine learning model to generate patient risk scores and high-risk patient profiles; triggering an early warning signal when the risk score exceeds a preset threshold, and automatically matching corresponding intervention strategies according to the risk score level; generating intervention tasks based on the intervention strategies, continuously tracking the execution status of the intervention tasks, and generating feedback results and displaying them on the risk early warning platform after the execution status is completed. The feedback results are used to update the parameters of the machine learning model. This application can effectively improve the efficiency of medical service dispute risk early warning, reduce the incidence of service disputes, and promote the harmonious development of medical service quality and doctor-patient relationships.

Description

Technical Field

[0001] This application relates to the field of risk prevention and control, and in particular to a service dispute risk early warning method and system based on multi-dimensional data features. Background Technology

[0002] In current technology, medical institutions primarily address service disputes through reactive measures, resolving them after they occur via complaint handling, mediation, or litigation. Some institutions have also attempted to establish simple early warning mechanisms, such as identifying potential risks through patient satisfaction surveys and complaint record statistics. However, these methods rely heavily on a single data source and human experience, exhibiting significant limitations. Human judgment is highly subjective, inefficient, and unable to monitor the risk status of a large number of patients in real time. Often, problems are only discovered when disputes have already occurred or are about to erupt, missing the optimal intervention window.

[0003] In summary, the current early warning efficiency for medical service disputes is relatively low. This deficiency makes it difficult for medical institutions to identify high-risk patients in a timely manner and take effective preventive intervention measures, resulting in a persistently high incidence of service disputes, which seriously affects the quality of medical services and the harmonious development of doctor-patient relationships. Summary of the Invention

[0004] This application provides a service dispute risk early warning method and system based on multi-dimensional data features, which can effectively improve the efficiency of medical service dispute risk early warning, enabling medical institutions to promptly identify high-risk patients and take effective preventive intervention measures, thereby effectively reducing the incidence of service disputes and promoting the harmonious development of medical service quality and doctor-patient relationships.

[0005] To achieve the above objectives, the embodiments of this application adopt the following technical solutions: Firstly, a service dispute risk early warning method based on multi-dimensional data features is provided, which is applied to servers with a risk early warning platform deployed. This method includes: Real-time acquisition of multi-source heterogeneous data through a risk warning platform; Constructing a multidimensional feature index matrix based on multi-source heterogeneous data; By analyzing the multidimensional feature index matrix through a pre-set machine learning model, risk scores and high-risk patient profiles are generated. When the risk score exceeds a preset threshold, an early warning signal is triggered, and the corresponding intervention strategy is automatically matched according to the level of the risk score. Intervention tasks are generated based on the intervention strategy, and the execution status of the intervention tasks is continuously tracked. When the execution status is completed, feedback results are generated and displayed on the risk warning platform. The feedback results are used to update the parameters of the machine learning model.

[0006] In one possible implementation of the first aspect, the multi-source heterogeneous data includes structured in-hospital diagnostic and treatment data and unstructured out-of-hospital feedback data. A multi-dimensional feature index matrix is ​​constructed based on the multi-source heterogeneous data, including: Feature extraction is performed on structured diagnosis and treatment data within the hospital and unstructured feedback data outside the hospital to obtain a feature set, which includes multiple feature indicators; Acquire historical dispute data, and assign weights to each feature index in the feature set based on the historical dispute data to generate a multi-dimensional feature index matrix.

[0007] In another possible implementation of the first aspect, the method further includes, before feature extraction from the structured in-hospital diagnostic and treatment data and the unstructured out-of-hospital feedback data: Real-time acquisition of diagnosis and treatment data and electronic medical record data, and establishment of a spatiotemporal topology map of patient diagnosis and treatment behavior according to timestamps and spatial departments. The spatiotemporal topology map includes multiple behavior nodes and topological connection relationships connecting behavior nodes. Calculate the outlier centrality of each behavioral node in the spatiotemporal topology graph; Active nodes are identified based on outlier centrality, and active nodes are behavioral nodes whose outlier centrality exceeds a preset outlier threshold. Determine the behavioral characteristics of active nodes and use these characteristics as auxiliary input features.

[0008] In another possible implementation of the first aspect, the outlier centrality of each behavioral node in the spatiotemporal topology graph is calculated, including: Count the access frequency of each behavior node within a preset time window; Calculate the topological distance between each behavior node and other behavior nodes; Based on the access frequency and topological distance, the deviation of each behavioral node from the normal diagnosis and treatment path is calculated to obtain the outlier centrality.

[0009] In another possible implementation of the first aspect, after identifying active nodes based on outlier centrality, the method further includes: Retrieve medical records and patient interaction texts associated with active nodes; Medical treatment documents and patient interaction texts are converted into medical semantic vectors and patient semantic vectors, respectively, and then mapped to a unified semantic vector space. The semantic deviation value between the semantic vectors of the medical staff and the semantic vectors of the patient is calculated in the semantic vector space to quantify the degree of cognitive asymmetry between the two parties in the information transmission process.

[0010] In another possible implementation of the first aspect, the semantic deviation value between the medical semantic vector and the patient semantic vector is calculated in the semantic vector space, including: In the semantic vector space, the semantic overlap is calculated based on the semantic vectors of the medical staff and the semantic vectors of the patient. Based on the comparison between semantic overlap and preset coefficient, a semantic deviation value is determined. When the semantic overlap is lower than the preset coefficient, semantic loss is determined and the semantic deviation value is marked as a high-risk level.

[0011] In another possible implementation of the first aspect, after determining the semantic deviation value, the method further includes: The interpretation cost for doctors and the trust loss for patients are determined based on the semantic vectors of doctors and patients. A strategy matrix is ​​generated based on interpretation cost and trust loss, wherein the strategy matrix includes a set of medical strategies and a set of patient strategies. The strategy evolution trajectory is determined based on the strategy matrix, where the strategy evolution trajectory is used to indicate the likelihood of disputes occurring.

[0012] In another possible implementation of the first aspect, the policy evolution trajectory is determined based on the policy matrix, including: Based on the strategy matrix, determine the payoff functions for the medical staff when adopting a fully explained strategy and a simplified explained strategy, and determine the payoff functions for the patient when adopting a rational negotiation strategy and an irrational escalation strategy. The fitness difference is calculated based on the medical benefit function and the patient benefit function. The equilibrium point of the policy evolution trajectory is determined based on fitness differences, and the convergence direction of the policy evolution trajectory is determined based on the equilibrium point. Calculate the phase trajectory in the policy space based on the convergence direction; The strategy evolution trajectory is determined based on the convergence direction and phase trajectory.

[0013] Secondly, this application provides a server, comprising: The memory is configured to store instructions; and The processor is configured to retrieve instructions from memory and, when executing instructions, implement the aforementioned service dispute risk warning method based on multi-dimensional data features.

[0014] Thirdly, this application provides a risk warning platform, which includes a cloud server and a client. The client is used to send a request to the cloud server, and the cloud server is used to execute the operation steps of any of the above methods according to the request.

[0015] Through the aforementioned technical solution, a risk warning platform acquires multi-source heterogeneous data in real time and constructs a multi-dimensional feature index matrix, enabling a comprehensive and multi-dimensional assessment of patient risk status. Compared to traditional methods relying on a single data source, this solution integrates heterogeneous data from multiple systems such as electronic medical records, nursing records, billing systems, and patient feedback. It comprehensively portrays patient status characteristics from multiple dimensions, including medical care, nursing, costs, and communication, significantly improving the comprehensiveness and accuracy of risk identification and avoiding risk omissions due to incomplete information. A pre-set machine learning model intelligently analyzes the multi-dimensional feature index matrix, automatically generating patient risk scores and high-risk patient profiles. Compared to traditional manual experience-based judgment, machine learning models can learn risk patterns and rules from massive amounts of historical data, achieving objective and standardized risk assessment and eliminating interference from subjective human factors. Simultaneously, the generation of high-risk patient profiles provides crucial evidence for subsequent precise interventions, enabling medical institutions to clearly understand the characteristic attributes and risk sources of high-risk patients, laying the foundation for developing targeted intervention strategies. When the risk score exceeds a preset threshold, an early warning signal is automatically triggered, and corresponding intervention strategies are intelligently matched according to the risk level, generating specific intervention tasks. This achieves seamless integration from risk identification to intervention response, significantly shortening response time and enabling medical institutions to take preventative measures before disputes occur, shifting from passive response to proactive prevention. By continuously tracking the execution status of intervention tasks and generating feedback results, a closed-loop risk management mechanism is formed. Feedback results are used to update the parameters of the machine learning model, enabling the model to continuously learn and optimize, thereby continuously improving the accuracy of risk prediction and the effectiveness of intervention strategies. In summary, this technical solution achieves accurate early warning and effective prevention of medical service disputes, significantly reducing the incidence of disputes and improving the quality of medical services and patient satisfaction.

[0016] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a service dispute risk early warning method based on multidimensional data features, provided for an embodiment of this application; Figure 2 An architecture diagram of a risk warning platform provided in this application embodiment; Figure 3 This is a schematic diagram of a spatiotemporal topology graph provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0019] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0020] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0021] Figure 1 The illustration shows a flowchart of a service dispute risk warning method based on multi-dimensional data features according to an embodiment of this application. Figure 1 As shown in the figure, this application provides a service dispute risk warning method based on multi-dimensional data features, which is applied to a server with a risk warning platform deployed. The method may include the following steps.

[0022] S110. Obtain multi-source heterogeneous data in real time through the risk warning platform; S120. Construct a multi-dimensional feature index matrix based on multi-source heterogeneous data; S130. Analyze the multidimensional feature index matrix through a preset machine learning model to generate the patient's risk score and high-risk patient profile. S140. When the risk score exceeds the preset threshold, an early warning signal is triggered, and the corresponding intervention strategy is automatically matched according to the level of the risk score. S150. Generate intervention tasks based on the intervention strategy, continuously track the execution status of the intervention tasks, and generate feedback results and display them on the risk warning platform when the execution status is completed. The feedback results are used to update the parameters of the machine learning model.

[0023] In this embodiment, the risk warning platform is deployed on the server side. Figure 2 An architecture diagram of a risk warning platform provided in an embodiment of this application is shown. (Refer to...) Figure 2 The risk warning platform includes a data acquisition module, a feature construction module, a risk assessment module, an early warning triggering module, and a task management module. The data acquisition module processes step S110, the feature construction module processes step S120, the risk assessment module processes step S130, the early warning triggering module processes step S140, and the task management module processes step S150. The risk warning platform establishes data interfaces with multiple business systems of medical institutions, such as electronic medical record systems, nursing systems, billing systems, and patient feedback systems, to achieve real-time data acquisition and two-way communication.

[0024] In practice, the risk warning platform connects with multiple business systems within medical institutions through pre-configured data interfaces to achieve real-time collection of multi-source heterogeneous data. This multi-source heterogeneous data mainly includes two categories: structured in-hospital diagnostic and treatment data and unstructured feedback data from outside the hospital. Structured in-hospital diagnostic and treatment data originates from electronic medical record systems, medical order systems, laboratory and examination systems, nursing record systems, and billing systems, and includes structured data such as basic patient information, diagnostic information, treatment plans, medication records, examination and test results, nursing records, and detailed expenses. Unstructured feedback data from outside the hospital originates from patient satisfaction surveys, online consultation records, complaint and suggestion systems, and third-party evaluation platforms, and includes unstructured text data such as patients' written evaluations, emotional expressions, and communication records.

[0025] Data acquisition employs a streaming processing architecture, utilizing message queue technology for real-time data transmission. Specifically, each business system automatically pushes data to the message queue when it is generated or updated. The risk warning platform then retrieves the data from the message queue in real time and performs preliminary processing. For structured data, the platform verifies the integrity of the data format and the validity of fields, marks missing values, and performs preliminary identification of outliers. For unstructured text data, the platform performs text cleaning, removing irrelevant symbols and stop words, and performs word segmentation.

[0026] To ensure the real-time nature and completeness of data collection, the platform has implemented a data monitoring mechanism. If a data source fails to transmit data for a preset period, an alarm will be automatically triggered, prompting the administrator to check the data interface status. Through this step, the risk warning platform establishes a data collection network covering the entire patient care process and all touchpoints, providing a comprehensive and timely data foundation for subsequent risk assessments and effectively solving the problems of single data sources and delayed updates in traditional methods.

[0027] After acquiring multi-source heterogeneous data, the platform performs deep feature extraction on both structured in-hospital medical data and unstructured out-of-hospital feedback data. For structured medical data, feature extraction includes two levels: basic features and derived features. Basic features are extracted directly from the raw data, such as patient age, gender, disease diagnosis, length of hospital stay, total medical expenses, and out-of-pocket payment ratio. Derived features are obtained through calculation and combination of basic features, such as calculating patient visit frequency, average length of hospital stay, cost growth rate, and complexity of examination items.

[0028] For unstructured feedback data, natural language processing techniques were used for feature extraction. First, a pre-trained BERT model was used to encode the text, converting it into a vector representation. Then, a sentiment analysis model was used to identify the patient's emotional tendency, extracting sentiment polarity scores and sentiment intensity values. Simultaneously, a keyword extraction algorithm was used to identify high-frequency words and key themes in the text, which were then used as feature indicators. After feature extraction, a feature set containing hundreds of feature indicators was obtained.

[0029] To improve the accuracy of risk prediction, weights need to be assigned to each feature indicator in the feature set. This weight assignment is based on historical dispute data. Specifically, all medical dispute cases that occurred within the past three years are collected, and the feature indicator values ​​for each patient in each case are extracted to construct a dispute sample dataset. The information gain algorithm is used to calculate the discriminative power of each feature indicator in predicting disputes; the higher the information gain value, the greater the contribution of that feature to dispute prediction, and the higher its assigned weight. The weight calculation formula is as follows: ,in This represents the weight of the i-th feature. Let represent the information gain value of the i-th feature, and n be the total number of features. Through weight allocation, the feature set is transformed into a multi-dimensional feature index matrix. Rows in the matrix represent different patient samples, columns represent different feature indices, and each element is the product of the feature value and its corresponding weight. This matrix integrates information from multiple dimensions, including medical care, nursing, costs, and communication, providing rich input features for machine learning models and significantly improving the comprehensiveness and accuracy of risk identification.

[0030] After constructing a multidimensional feature index matrix, the matrix is ​​intelligently analyzed using a pre-defined machine learning model. The machine learning model employs an ensemble learning architecture, combining the advantages of gradient boosting decision trees and deep neural networks.

[0031] Specifically, the model comprises three main components: a feature encoding layer, a risk assessment layer, and a profile generation layer. The feature encoding layer employs an autoencoder structure, consisting of an input layer, two hidden layers, and an output layer. The hidden layers have 128 and 64 neurons respectively, and ReLU is used as the activation function. This layer compresses the high-dimensional multidimensional feature index matrix into a low-dimensional dense vector, extracting key risk features. The risk assessment layer uses the XGBoost algorithm, setting the maximum tree depth to 6, the learning rate to 0.1, and the subsampling ratio to 0.8. Based on the encoded feature vector, this layer outputs a patient risk score through ensemble prediction of multiple decision trees. The risk score ranges from 0 to 100, with higher scores indicating a greater risk of dispute. During model training, historical dispute data is used as positive samples, and patient data without disputes is used as negative samples, with a sample ratio of 1:5. The cross-entropy loss function is used to measure the difference between the predicted results and the true labels. The loss function is: , where m is the sample size. For real labels, To predict probabilities, the optimizer uses Adam with an initial learning rate of 0.001, decaying by 0.85 every 20 epochs. An early stopping strategy is employed during training: training stops when the validation set loss fails to decrease for 10 consecutive epochs to prevent overfitting. The profiling layer extracts salient features from high-risk patients whose risk scores exceed a threshold, generating high-risk patient profiles. These profiles include typical attributes across multiple dimensions, such as demographics, disease characteristics, cost characteristics, and communication characteristics, for example, "middle-aged woman, chronic disease patient, high out-of-pocket expenses, and repeated expressions of dissatisfaction." This step enables a quantitative assessment of each patient's risk status and a precise characterization of high-risk group characteristics, providing a scientific basis for subsequent early warning triggering and intervention strategy development.

[0032] When a patient's risk score calculated by the machine learning model exceeds a preset threshold, the risk warning platform automatically triggers an alert. The preset threshold is set based on the medical institution's historical dispute rate and risk tolerance, typically 60 points. The alert signals include three risk levels: a yellow alert (moderate risk) is triggered when the risk score is between 60 and 75 points; an orange alert (higher risk) is triggered when the risk score is between 75 and 90 points; and a red alert (extremely high risk) is triggered when the risk score exceeds 90 points. Different intervention strategies correspond to different risk levels.

[0033] For patients with a yellow alert, the corresponding intervention strategy is to strengthen communication and care. Specific measures include assigning the responsible nurse to proactively communicate with the patient, understand their concerns and needs, and provide medical guidance and health education. For patients with an orange alert, the corresponding intervention strategy is to focus on key areas and coordinate between doctors and patients. Specific measures include forwarding the patient's information to the attending physician and department head, requiring the physician to pay special attention to the patient during ward rounds, explaining the treatment plan and expected outcomes in detail, and assigning a patient relations coordinator to communicate thoroughly with the patient and their family to resolve conflicts promptly. For patients with a red alert, the corresponding intervention strategy is to activate the emergency response mechanism. Specific measures include immediately notifying the medical department and the doctor-patient relations office, organizing a multidisciplinary expert consultation to assess the rationality of the treatment plan, adjusting the treatment plan if necessary, and assigning an experienced doctor-patient communication specialist to communicate one-on-one with the patient and their family, fully listening to their needs and providing psychological support.

[0034] The matching of intervention strategies employs a combination of rule engines and case-based reasoning. The rule engine, based on a predefined rule base, automatically matches corresponding intervention measures according to risk level and patient profile characteristics. Case-based reasoning retrieves cases with similar characteristics to the current patient from a historical database of successful intervention cases, extracting effective intervention measures used in those cases as a reference. This process automates and intelligently integrates risk identification with intervention response, significantly shortening response time and enabling medical institutions to take timely and targeted measures at the nascent stage of disputes, effectively preventing risk escalation.

[0035] Based on the matched intervention strategy, the risk warning platform automatically generates specific intervention tasks and assigns them to the corresponding responsible persons. Intervention tasks include elements such as task type, task content, responsible person, and deadline. For example, for a task requiring enhanced communication and care, the task content is "to communicate face-to-face with the patient to understand their comprehension and concerns regarding the treatment plan and provide detailed explanations," the responsible person is the nurse in charge, and the deadline is within 24 hours of task generation.

[0036] Once a task is generated, the platform's push notification function sends the task information to the responsible person's work terminal in real time, including both desktop and mobile applications. Upon receiving the task, the responsible person is required to complete it within the specified time and record the task's execution status on the platform. The platform continuously tracks and intervenes in the task's execution status, which includes four categories: pending, in progress, completed, and overdue. For tasks in the pending status, the platform will send a reminder notification before the deadline. For overdue tasks, the platform will automatically escalate the process, pushing the task information to the supervisor to urge completion.

[0037] After the responsible person completes the task, they need to fill in the task execution record on the platform, including detailed information such as communication time, communication content, patient feedback, and problem resolution status. The platform will perform correlation analysis between the execution record and the patient's subsequent medical data to evaluate the actual effectiveness of the intervention. When the task execution status changes to "completed," the platform automatically generates feedback results. The feedback results include key information such as whether the intervention was effective, the trend of changes in the patient's risk score, and whether any disputes occurred. For cases with effective interventions, their characteristics and intervention measures are extracted and added to the success case library. For cases with ineffective interventions or where disputes still occur, in-depth analysis is conducted to identify shortcomings in the early warning model. The feedback results are used to update the parameters of the machine learning model.

[0038] Specifically, new case data is added to the training set, the model is retrained, and feature weights and model parameters are adjusted, enabling the model to continuously learn and optimize from practice. Model updates employ incremental learning, with parameters updated monthly to ensure optimal performance. This process forms a complete closed loop from risk warning and intervention execution to effect feedback, achieving continuous improvement and optimization of the risk management mechanism and significantly enhancing the accuracy of warnings and the effectiveness of interventions.

[0039] This embodiment acquires multi-source heterogeneous data in real time through a risk warning platform and constructs a multi-dimensional feature index matrix, achieving a comprehensive and three-dimensional assessment of patient risk status. Compared with traditional methods relying on a single data source, this solution integrates data from multiple systems such as electronic medical records, nursing records, billing systems, and patient feedback, comprehensively depicting patient status characteristics from multiple dimensions including medical treatment, nursing, costs, and communication. This significantly improves the comprehensiveness and accuracy of risk identification and avoids risk omissions due to incomplete information. An integrated learning architecture-based machine learning model intelligently analyzes the multi-dimensional feature index matrix, automatically generating patient risk scores and high-risk patient profiles. Compared to traditional human experience-based judgment, the machine learning model can learn risk patterns and rules from massive historical data, achieving objective and standardized risk assessment and eliminating interference from subjective human factors. The generation of high-risk patient profiles provides an important basis for subsequent precise intervention, enabling medical institutions to clearly understand the characteristic attributes and risk sources of high-risk patients. When the risk score exceeds a preset threshold, an early warning signal is automatically triggered, and corresponding intervention strategies are intelligently matched according to the risk level, generating specific intervention tasks. This achieves seamless integration from risk identification to intervention response, significantly shortening response time and enabling medical institutions to take preventative measures before disputes occur, shifting from passive response to proactive prevention. By continuously tracking the execution status of intervention tasks and generating feedback results, a closed-loop risk management mechanism is formed. Feedback results are used to update the parameters of the machine learning model, enabling the model to continuously learn and optimize, thereby continuously improving the accuracy of risk prediction and the effectiveness of intervention strategies. This technical solution not only achieves accurate early warning and effective prevention of medical service disputes but also promotes continuous improvement in the quality of medical services through a data-driven approach, effectively reducing the incidence of disputes, improving doctor-patient relationships, increasing patient satisfaction, and enhancing the operational efficiency of medical institutions.

[0040] In one embodiment of this invention, the multi-source heterogeneous data includes structured in-hospital diagnostic and treatment data and unstructured out-of-hospital feedback data. Constructing a multi-dimensional feature index matrix based on the multi-source heterogeneous data includes the following steps: S210. Extract features from the structured diagnosis and treatment data within the hospital and the unstructured feedback data outside the hospital to obtain a feature set, which includes multiple feature indicators. S220. Obtain historical dispute data, and assign weights to each feature indicator in the feature set based on the historical dispute data to generate a multi-dimensional feature indicator matrix.

[0041] In this embodiment, the feature set may include patient-dimensional data feature indicators, treatment outcome-dimensional data feature indicators, and hospital environment-dimensional data feature indicators.

[0042] In this embodiment, during actual implementation, when extracting features from structured medical data within the hospital, patient-dimensional data feature indicators are first extracted from the electronic medical record system, medical order system, and laboratory examination system. These features include basic characteristics such as the patient's age, gender, occupation, education level, number of previous visits, and disease severity. Simultaneously, by calculating derived features such as the patient's visit interval, follow-up visit frequency, and treatment adherence, a patient behavioral profile is constructed.

[0043] Extracting data feature indicators for treatment outcomes from medical records, including diagnostic accuracy, number of treatment plan changes, surgical success rate, complication rate, and deviation of length of hospital stay, among which the deviation of length of hospital stay is determined by a formula. calculate, This refers to the actual number of days of hospitalization. The standard length of hospital stay was used. Hospital environment-related data features were extracted from the billing system and patient satisfaction surveys, including cost transparency, waiting time, ward conditions rating, and healthcare worker attitude rating.

[0044] For unstructured feedback data from outside the hospital, a BERT pre-trained model was used for text encoding, converting patient evaluation texts into 768-dimensional vector representations. A sentiment analysis model was used to identify emotional polarity in the text, extracting sentiment feature indicators such as emotion scores, anger levels, and anxiety levels. The TF-IDF algorithm was used to extract high-frequency keywords, identifying the core issues of concern to patients, such as the frequency of occurrence of keywords like "cost," "attitude," and "effect," as feature indicators. After feature extraction, a feature set of over 150 feature indicators was obtained, encompassing patient, treatment outcome, and hospital environment dimensions. This step, by converting multi-source heterogeneous data into structured feature indicators, solved the problem of inconsistent formats across different data sources, making direct use in model analysis difficult, and laid a data foundation for subsequent weight allocation and risk assessment.

[0045] After obtaining the feature set, historical dispute data was extracted from the medical institution's dispute management system, including all confirmed medical dispute cases from the past three years, totaling over 500 cases. For each dispute case, the corresponding patient's characteristic index values ​​before the dispute occurred were traced back to construct a dispute sample feature dataset. Simultaneously, data from 2500 patients who did not experience disputes were randomly selected as a control sample, forming a training dataset with a positive-to-negative sample ratio of 1:5.

[0046] The information gain algorithm is used to calculate the discriminative power of each feature index in distinguishing between disputes. The information gain value reflects the effectiveness of that feature in separating disputed and non-disputed samples. The calculation formula is as follows: ,in The entropy of the dispute For a given feature Conditional entropy under given conditions. A larger information gain value indicates a greater contribution of that feature to dispute prediction. Normalization based on the information gain value yields the weight coefficients for each feature: The experimental results showed that the "number of treatment plan changes" and "incidence of complications" had the highest weights in the diagnosis and treatment outcome dimension, at 0.12 and 0.11 respectively. The "number of past complaints" in the patient dimension had a weight of 0.09, and the "medical staff attitude rating" in the hospital environment dimension had a weight of 0.08.

[0047] The original value of each feature indicator in the feature set is multiplied by its corresponding weight to construct a multidimensional feature indicator matrix. The rows of the matrix represent different patient samples, and the columns represent the weighted feature indicators. The matrix has a dimension of m×150, where m is the number of patient samples. This step, through weight allocation based on historical dispute data, quantifies and distinguishes the importance of feature indicators, highlighting key features strongly correlated with the occurrence of disputes. This effectively solves the problem of insufficient prediction accuracy caused by equal weighting of all features in traditional methods, significantly improving the discriminative ability of the risk assessment model.

[0048] This embodiment utilizes deep feature extraction from structured in-hospital diagnostic and treatment data and unstructured out-of-hospital feedback data to construct a multi-level feature set encompassing patient, treatment outcome, and hospital environment dimensions, enabling a multi-faceted characterization of patient risk status. Compared to traditional methods that focus only on a single dimension of data, this approach comprehensively captures risk factors from multiple perspectives, including individual patient characteristics, treatment process quality, and hospital service environment, avoiding risk omissions due to a single perspective and significantly improving the comprehensiveness of risk identification. The mechanism of weighting feature indicators based on historical dispute data allows the risk assessment model to automatically identify and strengthen key features strongly correlated with the occurrence of disputes, while weakening the interference of irrelevant features. Compared to traditional equal-weighting or manual experience-based weighting methods, this approach achieves data-driven objective weight allocation through an information gain algorithm, eliminating the bias of subjective judgment and improving the scientific nature and accuracy of weight allocation. The generated multi-dimensional feature index matrix provides structured and standardized input data for machine learning models, effectively solving the technical challenge of inconsistent formats of multi-source heterogeneous data that are difficult to directly apply to model training. This technical solution not only improves the predictive accuracy of the risk warning model, but also provides medical institutions with interpretable risk factor analysis results. This enables managers to clearly understand which factors have the greatest impact on dispute risk, thereby improving service quality in a targeted manner. It achieves effective transformation from data to decision-making, promoting the continuous improvement of medical service quality and the harmonious development of doctor-patient relationships.

[0049] In one embodiment of this invention, before feature extraction from the structured in-hospital diagnostic and treatment data and the unstructured out-of-hospital feedback data, the method further includes the following steps: S310. Real-time acquisition of diagnosis and treatment data and electronic medical record data, and establishment of a spatiotemporal topology map of patient diagnosis and treatment behavior according to timestamps and spatial departments. The spatiotemporal topology map includes multiple behavior nodes and topological connection relationships connecting behavior nodes. S320. Calculate the outlier centrality of each behavior node in the spatiotemporal topology graph; S330. Identify active nodes based on outlier centrality. Active nodes are behavioral nodes whose outlier centrality exceeds a preset outlier threshold. S340. Determine the behavioral characteristics of active nodes and use the behavioral characteristics of active nodes as auxiliary input features.

[0050] Behavioral nodes can include registration nodes, laboratory testing nodes, department transfer nodes, payment nodes, and complaint nodes.

[0051] In this embodiment, the risk warning platform obtains real-time diagnostic and treatment data and electronic medical record data from the hospital information system through a data interface. The diagnostic and treatment data includes patient registration records, outpatient visit records, inpatient records, examination and testing records, payment records, complaint records, etc., with each record containing a timestamp and department location information. The electronic medical record data includes progress notes, medical orders, nursing records, etc., and similarly contains information in both time and space dimensions.

[0052] Based on this data, a spatiotemporal topological map of the diagnosis and treatment behavior is constructed for each patient according to the chronological order of timestamps and the distribution of departments in the space, such as... Figure 3 As shown, the spatiotemporal topology graph is represented by a directed graph structure. The nodes in the graph represent various behavioral events that occur during a patient's medical treatment process, including registration nodes, laboratory testing nodes, department transfer nodes, payment nodes, and complaint nodes. Each behavioral node contains four key attributes: node type, timestamp, spatial location (department code), and behavioral status (normal / abnormal). The topological connections between nodes are represented by directed edges, where the direction of the edge indicates the temporal order of the behavior, and the weight of the edge indicates the time interval between two behaviors. For example, the path from registration to laboratory testing is represented as: registration node → laboratory testing node, with the edge weight being the time difference between the two behaviors. For complex medical treatment paths, such as when a patient is transferred between multiple departments, the topology graph will form a complex network structure containing multiple branches.

[0053] In the specific construction process, all patient behavior records are first extracted and sorted by timestamp. Then, corresponding behavior nodes are created sequentially, and directed connections are established between nodes based on their chronological order. For behaviors occurring in different departments within the same time period, parallel branch paths are created. This step, by constructing a spatiotemporal topology graph, transforms the patient's discrete medical behavior into a structured graph data representation. This allows for the systematic analysis and quantification of the patient's medical trajectory and behavioral patterns, laying the data structure foundation for subsequent identification of abnormal behavior patterns and solving the problem that traditional methods struggle to comprehensively capture the spatiotemporal characteristics of patient medical behavior.

[0054] After constructing the spatiotemporal topology graph, it is necessary to calculate the outlier centrality of each behavioral node in the graph to quantify the degree of deviation of that node from the normal treatment path. The calculation of outlier centrality considers two dimensions: the frequency of node access and the topological distance. First, the access frequency of each behavioral node within a preset time window is counted. The time window is set to the most recent 30 days. For each node type, the number of times the patient accesses that type of node within 30 days is counted. For example, the number of times the patient makes an appointment, undergoes tests, or files a complaint within 30 days is counted. Access frequency reflects the intensity of a patient's use of a certain type of medical service; abnormally high or low frequencies may indicate potential risks.

[0055] Secondly, the topological distance between each behavioral node and other behavioral nodes is calculated. Topological distance is defined as the minimum number of edges required to reach another node from one node, calculated using a breadth-first search algorithm. For each node, the average topological distance to all other nodes in the graph is calculated; a larger distance indicates a more isolated node in the topological structure. Based on the visit frequency and topological distance, the deviation of each behavioral node from the normal treatment path is calculated, yielding the outlier centrality. The calculation formula is: ,in Let i be the outlier centrality of node i. Let i be the frequency of visits to node i. and These represent the average access frequency and standard deviation for this type of node, respectively. Let i be the average topological distance to node i. The average topological distance across all nodes. and These are weighting coefficients, typically set to 0.6 and 0.4.

[0056] The normal medical treatment pathway is analyzed by examining the spatiotemporal topology of patients with no historical disputes, extracting typical path patterns as a benchmark. For example, a typical outpatient medical path includes: registration, consultation, laboratory tests, medication dispensing, and payment, with time intervals and topological distances between nodes within normal ranges. This step, by quantifying the outlier centrality of behavioral nodes, enables the automatic identification of abnormal patient behavior, solving the problem of traditional methods relying on human experience and difficulty in timely detecting abnormal behavior patterns. It provides objective quantitative indicators for accurately identifying high-risk patients.

[0057] After calculating the outlier centrality of each behavioral node in the spatiotemporal topology graph, active nodes are identified based on the outlier centrality. Active nodes are defined as behavioral nodes whose outlier centrality exceeds a preset outlier threshold. These nodes represent abnormal patterns in patient medical behavior and may indicate potential dispute risks. The preset outlier threshold is determined by analyzing the spatiotemporal topology graphs of patients in historical dispute cases, statistically analyzing the outlier centrality distribution of behavioral nodes of disputed patients, and selecting a threshold that effectively distinguishes between disputed and non-disputed patients. The outlier threshold is typically set to 1.5.

[0058] During the identification process, all behavioral nodes in the spatiotemporal topology graph are traversed, and the outlier centrality is compared with a preset threshold. If... (in If a node is identified as an active node (with a preset outlier threshold), then that node is marked as active. Different types of behavioral nodes have different meanings when identified as active nodes. An active registration node typically indicates that the patient frequently registers for appointments or changes departments multiple times within a short period, potentially reflecting dissatisfaction with the treatment plan or "doctor shopping" behavior. An active laboratory testing node may indicate that the patient has undergone tests beyond the standard scope, suggesting over-testing or a complex condition. An active departmental transfer node indicates that the patient frequently transfers between multiple departments, potentially reflecting unclear diagnosis or poor interdepartmental collaboration. An active payment node typically indicates that the patient has objections to the fees and has repeatedly visited the cashier to inquire or complain. An active complaint node directly indicates that the patient has developed dissatisfaction and is a strong signal of potential disputes.

[0059] By identifying these active nodes, abnormal situations during a patient's medical treatment can be detected in a timely manner, providing crucial clues for subsequent risk warnings. This step, through automatic identification of active nodes, enables real-time monitoring and early warning of abnormal patient behavior, solving the problem that traditional methods cannot promptly detect abnormal behavioral patterns during patient treatment. This allows medical institutions to take intervention measures at the nascent stage of potential disputes, significantly improving the timeliness of risk warnings.

[0060] After identifying active nodes, it is necessary to determine their behavioral characteristics and incorporate them as auxiliary input features into the multidimensional feature index matrix to enhance the predictive ability of the risk assessment model. The behavioral characteristics of active nodes include four dimensions: node type, node quantity, node distribution, and node sequence pattern. The node type feature records which types of behavioral nodes are identified as active nodes, using one-hot encoding. For example, if a complaint node and a transfer node are active nodes, the corresponding position is encoded as 1, and other positions are 0. The node quantity feature counts the total number of active nodes; a higher number indicates more frequent abnormal patient behavior and a higher risk of disputes.

[0061] The node distribution feature analysis examines the distribution of active nodes in the spatiotemporal topology graph, including the temporal distribution density and spatial distribution concentration of active nodes. Temporal distribution density is measured by calculating the standard deviation of the time intervals between active node occurrences; a smaller standard deviation indicates a more concentrated temporal pattern of abnormal behavior, potentially reflecting a concentrated problem within a specific time period. Spatial distribution concentration is measured by calculating the number of departments involved in the active nodes; fewer departments indicate the problem is concentrated in a specific department, while more departments indicate the problem involves multiple departments. The node sequence pattern feature extracts the order of occurrence of active nodes in the spatiotemporal topology graph and represents it using sequence encoding. For example, if the sequence of active node occurrences is "registration, laboratory tests, complaint," it is encoded as a corresponding sequence vector. Sequence patterns can reflect the evolution of a patient's journey from initial medical visit to developing dissatisfaction, providing important information for understanding the mechanism of dispute formation.

[0062] These behavioral features are converted into numerical feature vectors and used as auxiliary input features, which are then merged with the existing multidimensional feature index matrix. Specifically, new columns are added to the feature matrix to store the behavioral feature values ​​of active nodes. For example, columns for "number of active nodes," "complaint node identifier," "department transfer node identifier," "temporal distribution density," and "spatial distribution concentration" are added. These auxiliary features, together with the original patient-dimensional, treatment outcome-dimensional, and hospital environment-dimensional features, constitute an enhanced multidimensional feature index matrix, providing richer input information for the machine learning model. This step, by extracting the behavioral features of active nodes and using them as auxiliary input features, enriches the feature dimensions of the risk assessment model, enabling the model to capture the dynamic evolution patterns of patient medical behavior. This solves the problem that traditional static features cannot reflect the changing trends of patient behavior, significantly improving the accuracy and sensitivity of risk prediction, and allowing medical institutions to identify high-risk patients earlier.

[0063] This embodiment constructs a spatiotemporal topology graph of patient treatment behavior, transforming discrete behavioral events during the medical process into structured graph data representations, thus achieving systematic modeling of patient treatment trajectories and behavioral patterns. Compared to traditional methods that focus only on individual behavioral events, the spatiotemporal topology graph can comprehensively depict the patient's treatment path and behavioral sequence, revealing the temporal and spatial relationships between behavioral events, providing a data foundation for in-depth analysis of patient behavioral patterns. By calculating the outlier centrality of behavioral nodes and identifying active nodes, automated and quantitative identification of abnormal patient treatment behaviors is achieved. Compared to traditional methods relying on human experience, this approach, based on objective calculations of access frequency and topological distance, can accurately identify abnormal behavioral patterns deviating from the normal treatment path, eliminating the bias of subjective judgment and improving the accuracy and consistency of abnormal behavior identification. Extracting the behavioral features of active nodes and using them as auxiliary input features enriches the feature dimensions of the risk assessment model, enabling the model to simultaneously consider the patient's static attribute features and dynamic behavioral features. Static features reflect the patient's basic state, while dynamic behavioral features reveal abnormal patterns and risk evolution trends during the patient's treatment process; the combination of the two significantly enhances the model's predictive ability. This technical solution not only improves the accuracy and timeliness of risk warnings, but also provides medical institutions with a traceable chain of behavioral evidence, enabling managers to clearly understand when, where, and which aspects of abnormal behavior of high-risk patients occurred, thereby conducting targeted investigations and interventions, effectively reducing the incidence of disputes, improving the quality of medical services and patient satisfaction, and promoting the harmonious development of doctor-patient relationships.

[0064] In one embodiment of this invention, calculating the outlier centrality of each behavioral node in the spatiotemporal topology graph includes the following steps: S410. Count the access frequency of each behavior node within a preset time window; S420. Calculate the topological distance between each behavior node and other behavior nodes; S430. Based on the access frequency and topological distance, calculate the degree of deviation of each behavioral node from the normal diagnosis and treatment path to obtain the outlier centrality.

[0065] In this embodiment, the risk warning platform first sets a preset time window, typically the most recent 30 days, to capture patients' recent medical behavior patterns. The selection of the time window is based on the formation cycle analysis of medical disputes; research shows that risk signals for most medical disputes gradually emerge within 30 days before the event. For each behavioral node in the spatiotemporal topology graph, the platform counts the frequency of access to that node within the time window.

[0066] The statistical process is conducted separately for each node type. For the registration node, the total number of registrations by the patient within 30 days and the distribution of registrations across different departments are counted. For example, if a patient registers 8 times within 30 days (3 times in internal medicine, 2 times in surgery, and 3 times in orthopedics), then the total access frequency of the registration node is 8. The departmental distribution entropy can be calculated using the formula... Calculation, where Let k represent the proportion of patients registering in the i-th department, and k be the total number of departments involved. For the laboratory testing node, the frequency of various examination and testing items is statistically analyzed, including blood routine tests, biochemical tests, imaging tests, etc., and the diversity index of examination items is calculated. For the department transfer node, the number of department transfers and the complexity of the transfer path are statistically analyzed. For the payment node, the number of payments, the amount of each payment, and the number of refunds are statistically analyzed. For the complaint node, the number of complaints, the complaint channels (in person, by phone, online), and the types of complaint content are statistically analyzed.

[0067] The frequency of visits is statistically analyzed using a sliding window mechanism, updated daily to ensure data real-time performance. The platform compares the statistical results with the frequency distribution of visits from normal patients, which is modeled using historical data from patients without disputes. For example, the average number of appointments for a normal outpatient over 30 days is 1.5, with a standard deviation of 0.8. If a patient's appointment frequency is 8, it significantly exceeds the normal range, and the abnormality of this visit frequency can be quantified. Where F is the actual frequency, Average frequency The standard deviation is used. This step quantifies the frequency characteristics of patient medical behavior by statistically analyzing the frequency of visits to behavioral nodes within a time window. This solves the problem that traditional methods struggle to objectively assess whether the frequency of patient visits is abnormal, providing a quantitative basis for identifying high-risk behavioral patterns such as excessive medical visits and frequent complaints. This shifts risk identification from subjective judgment to data-driven approaches.

[0068] After counting the frequency of visits, it is necessary to calculate the topological distance between each behavioral node and other behavioral nodes in the spatiotemporal topology graph to evaluate the location characteristics of the nodes in the medical treatment path. Topological distance is defined as the minimum number of edges required to reach another node from one node, reflecting the degree of correlation between two behavioral events in the medical treatment process. The calculation process uses a breadth-first search algorithm, starting from the target node and traversing its neighboring nodes layer by layer until all reachable nodes are reached. For each behavioral node, the topological distance to all other nodes in the graph is calculated, and the average topological distance is calculated. ,in Let be the shortest path length from node i to node j, and n be the total number of nodes in the graph. The larger the average topological distance, the more isolated the node is in the topological structure, and the weaker its connection with other nodes.

[0069] Meanwhile, the topological centrality of the compute nodes is measured using the proximity centrality index, as shown in the formula: A higher proximity centrality indicates a more central position of the node in the medical treatment path. In a normal treatment path, the topological distances between various behavioral nodes exhibit a regular distribution. For example, the registration node is usually adjacent to the consultation node (topological distance 1), the consultation node and the laboratory node are typically 1-2 apart, and the laboratory node and the medication pickup node are 2-3 apart. The platform constructs a benchmark model of topological distances for normal treatment paths, extracting typical path patterns and their topological distance distributions by analyzing the spatiotemporal topological graphs of patients without historical disputes.

[0070] If a patient's behavioral node topological distance deviates significantly from the baseline model—for example, the topological distance between the complaint node and the registration node is 1 (indicating the patient filed a complaint immediately after registration), or the topological distances between the department transfer node and multiple department nodes are very small (indicating the patient frequently transferred departments within a short period)—then it indicates an abnormal patient's medical path. The deviation in topological distance is calculated using the formula... Calculation, where The average topological distance of the normal path. The standard deviation is given. This step quantifies the positional characteristics of patient medical behavior within the process structure by calculating the topological distance of nodes. This solves the problem of traditional methods focusing only on individual behavioral events and ignoring the relationships between them, enabling risk assessment to capture abnormal structural patterns in the medical path and improving the ability to identify complex risk scenarios.

[0071] After obtaining access frequency and topological distance, the deviation of each behavioral node from the normal treatment path is calculated based on these two dimensions, yielding outlier centrality. Outlier centrality comprehensively reflects the degree of anomalousness of a node in both frequency and structure dimensions, and is a key indicator for identifying high-risk behavioral nodes. The calculation formula is as follows: ,in Let be the outlier centrality of node i, and the first term be the standardized deviation of the visit frequency. Let i be the frequency of visits to node i. and These are the average visit frequency and standard deviation for this type of node, respectively. The second term is the standardized value of the topological distance. Let i be the average topological distance to node i. The third term represents the average topological distance across all nodes. The inherent risk weights for each node type differ, with complaint nodes having the highest inherent risk weight (1.0), followed by referral nodes (0.8), payment nodes (0.6), laboratory testing nodes (0.4), and registration nodes (0.3). Weighting coefficients are used for these risk factors. , , The weights were set to 0.4, 0.3, and 0.3 respectively, and the optimal weight combination was determined through regression analysis of historical dispute cases.

[0072] In calculating outlier centrality, a time decay factor for nodes is also considered; nodes closer to the current time have higher weights. The time decay factor uses an exponential decay function. Where t is the number of days since the node occurred. The attenuation coefficient is set to 0.05. The final outlier centrality is... This calculation method accurately identifies behavioral nodes that deviate from normal patterns across multiple dimensions, including access frequency, topology, and node type. For example, if a patient's complaint node has been accessed 3 times in 30 days (normal is 0.1 times), and its topological distance from the registration node is 1 (indicating a complaint was filed immediately after registration), then the outlier centrality of this complaint node will be significantly higher than the threshold, and it will be identified as an active node. This step, by comprehensively calculating outlier centrality based on access frequency and topological distance, achieves a multi-dimensional quantitative assessment of abnormal patient behavior, solving the misjudgment and omission problems caused by the single-dimensional assessment of traditional methods. It significantly improves the accuracy and robustness of abnormal behavior identification, enabling the risk warning system to more accurately capture dispute risk signals.

[0073] This embodiment achieves quantitative analysis of patient medical behavior frequency characteristics by statistically analyzing the access frequency of behavioral nodes within a preset time window, enabling the system to objectively identify abnormal behavioral patterns such as excessive medical visits and frequent complaints. Compared to traditional methods relying on the subjective perception of medical staff, this solution, based on objective data statistics, eliminates the bias of human judgment and improves the consistency and reliability of abnormal behavior identification. By calculating the topological distance between behavioral nodes, the structural characteristics of the patient's medical path are revealed, enabling the system to identify abnormal jumps and path deviations in the medical process. Compared to traditional methods that only focus on individual behavioral events, topological distance analysis can capture the correlation between behavioral events and discover risk patterns hidden in complex medical paths, significantly improving the depth and breadth of risk identification. By combining access frequency and topological distance to calculate outlier centrality, a multi-dimensional comprehensive assessment of abnormal patient behavior is achieved, avoiding the one-sidedness of single-dimensional assessment. This technical solution not only improves the accuracy of abnormal behavior identification, but also provides medical institutions with interpretable risk quantification indicators, enabling managers to clearly understand the specific manifestations and severity of patients' abnormal behavior, thereby formulating targeted intervention strategies, effectively reducing the incidence of disputes, improving the quality of medical services and patient satisfaction, and promoting the harmonious development of doctor-patient relationships.

[0074] In one embodiment of this invention, after identifying active nodes based on outlier centrality, the method further includes the following steps: S510, Obtain medical treatment documents and patient interaction texts associated with the active node; S520. Convert the medical treatment documents and patient interaction text into medical semantic vectors and patient semantic vectors respectively, and map the medical semantic vectors and patient semantic vectors to a unified semantic vector space. S530. Calculate the semantic deviation value between the semantic vector of the medical staff and the semantic vector of the patient in the semantic vector space to quantify the degree of cognitive asymmetry between the two parties in the information transmission process.

[0075] This includes mapping the semantic vectors of the medical staff and the semantic vectors of the patient to a unified semantic vector space, including: The medical diagnosis and treatment documents are input into a pre-trained medical domain semantic mapping model to generate an initial semantic vector for the medical documents. The semantic mapping model includes a medical terminology encoder and an everyday language encoder. The patient's interactive text is input into a pre-trained medical domain semantic mapping model to generate the patient's initial semantic vector. By performing coordinate transformation on the initial semantic vectors of the medical staff and the patient through a cross-domain alignment layer, the mapped semantic vectors of the medical staff and the patient are obtained and located in the same semantic vector space.

[0076] In this embodiment, after identifying active nodes, the risk warning platform needs to conduct in-depth analysis of the cognitive differences between doctors and patients during information transmission, which is one of the important root causes of medical disputes. The platform first obtains the medical treatment documents and patient interaction texts associated with the active nodes. The medical treatment documents include formal medical documents such as medical records, diagnostic reports, treatment plan descriptions, surgical informed consent forms, medical orders, and nursing records corresponding to the time period of the active node. These documents typically use standardized medical terminology and contain a large number of professional terms such as disease names, drug names, examination items, and medical indicators.

[0077] The patient-interaction texts come from multiple sources, including verbal records of patients during their consultations, questionnaires completed by patients and their families, text exchanges on online consultation platforms, complaints in complaint and suggestion systems, and evaluation texts from third-party evaluation platforms. These texts use everyday spoken language and often contain emotional vocabulary and non-professional descriptions.

[0078] The platform preprocesses the acquired text using natural language processing technology, including removing irrelevant symbols, correcting typos, and word segmentation. For medical treatment documents, it focuses on extracting key information paragraphs such as disease diagnosis, treatment plans, risk disclosures, and prognostic descriptions. For patient interaction texts, it focuses on extracting key content such as the patient's understanding of their condition, expectations for treatment, evaluation of medical services, and expressed doubts and dissatisfaction. This step, by acquiring raw text data from both doctors and patients, provides a data foundation for subsequent analysis of cognitive differences in doctor-patient communication, solving the problem of traditional methods lacking objective assessment criteria for doctor-patient communication quality, and enabling dispute risk assessment to delve into the semantic level of doctor-patient communication.

[0079] After obtaining the texts from both the doctor and patient, it is necessary to convert the doctor's medical records and the patient's interactive texts into semantic vector representations and map them to a unified semantic vector space for quantitative comparison. This process uses a pre-trained medical domain semantic mapping model, which is specifically designed for doctor-patient communication scenarios and includes two parallel encoding branches: a medical terminology encoder and an everyday language encoder.

[0080] The medical terminology encoder, based on the BERT architecture and pre-trained on large-scale medical literature and clinical case data, accurately understands the semantics of medical terminology. The encoder comprises a 12-layer Transformer structure with 768 hidden layers and 12 attention heads. For the input medical treatment document, word segmentation and medical entity recognition are performed first, identifying medical entities such as disease names, drug names, and examination items. The text sequence is then input into the medical terminology encoder. The encoder captures the semantic relationships between words in the text through a multi-layer self-attention mechanism, ultimately generating a 768-dimensional initial semantic vector at the output layer. This vector comprehensively represents the medical semantic information in the treatment document, with different dimensions corresponding to different semantic features; for example, some dimensions may correspond to disease severity, while others may correspond to the complexity of the treatment plan.

[0081] The everyday language encoder is also based on the BERT architecture, but it is pre-trained on everyday language data such as patient forums, doctor-patient dialogue records, and patient reviews, enabling it to understand non-professional spoken expressions and emotional vocabulary. For the input patient-patient interaction text, the encoder first performs sentiment analysis and intent recognition to identify the patient's emotional tendencies and core concerns. Then, the text sequence is input into the everyday language encoder to generate a 768-dimensional initial semantic vector for the patient. This vector represents the patient's understanding and feelings about medical information, including multi-dimensional information such as the patient's cognitive level, emotional state, and expectations.

[0082] Because the medical terminology encoder and the everyday language encoder are trained on different datasets, the semantic vectors they generate reside in different vector spaces and cannot be directly compared. Therefore, a cross-domain alignment layer is needed to map the two vectors to a unified semantic vector space. The cross-domain alignment layer employs a linear transformation plus non-linear activation structure, containing two fully connected layers. The first layer maps the 768-dimensional vector to 512 dimensions using ReLU activation, and the second layer maps the 512-dimensional vector to a unified 256-dimensional semantic space using Tanh activation. The transformation matrix is ​​obtained through alignment learning. The training data contains the correspondence between medical terms and everyday language, such as "myocardial infarction" corresponding to "heart attack," and "hypertension" corresponding to "high blood pressure." The goal of alignment learning is to minimize the vector distance between semantically similar medical terms and everyday language in the unified space. The loss function is: ,in and Let i be the vector of medical terminology and the vector of everyday language corresponding to the i-th pair. and This is the transformation function for the cross-domain alignment layer.

[0083] After cross-domain alignment, the initial semantic vectors of the medical staff and the patient are mapped to the same 256-dimensional semantic vector space, resulting in mapped semantic vectors for both the medical staff and the patient. In this unified space, each dimension of the vector represents a specific semantic attribute; different understandings of the same medical information by the medical staff and the patient will manifest as differences in the vectors along those dimensions. This step, by mapping medical professional language and everyday spoken language to a unified semantic vector space, solves the problem of semantic comparison difficulties caused by the use of different language systems by medical staff and patients. It achieves a quantitative representation of cognitive differences between medical staff and patients, providing a technical means for accurately identifying communication barriers between doctors and patients, and enabling risk warning to penetrate to the semantic cognitive level.

[0084] After mapping the semantic vectors of the medical staff and the patient to a unified semantic vector space, it is necessary to calculate the semantic deviation value between the two vectors to quantify the degree of cognitive asymmetry between the two parties during the information transmission process. The semantic deviation value reflects the gap between the medical information conveyed by the doctor and the medical information understood by the patient, and is an important risk factor leading to medical disputes.

[0085] The calculation process first calculates the semantic overlap between the medical and patient semantic vectors in the semantic vector space. The semantic overlap is measured using cosine similarity, and the formula is: ,in For medical semantic vectors, For the semantic vector of the patient, Represents the vector dot product. This represents the Euclidean norm of the vectors. The cosine similarity ranges from -1 to 1. A value closer to 1 indicates that the two vectors are more aligned in direction, meaning that the doctor and patient have a closer understanding of the medical information. A value closer to 0 or a negative value indicates a greater difference in understanding. In medical communication scenarios, a preset coefficient of 0.7 is typically set; that is, when the semantic overlap is below 0.7, a significant cognitive difference between the doctor and patient is considered to exist.

[0086] The semantic deviation value is determined based on the comparison between the semantic overlap degree and the preset coefficient. The formula for calculating the semantic deviation value is as follows: When the semantic overlap is within the normal range (≥0.7), the semantic deviation value is a simple linear transformation; when the semantic overlap is at a moderate level... When the semantic overlap is very low (<0.3), the deviation value doubles, indicating a significant communication risk; when the semantic overlap is very low (<0.3), the deviation value triples, indicating a serious loss of semantic meaning and a fundamental difference in understanding of medical information between doctors and patients.

[0087] Specifically, when the semantic overlap is below a preset coefficient of 0.7, semantic loss is identified, and the semantic deviation value is marked as high-risk. High-risk semantic deviation indicates that the patient may have completely misunderstood the doctor's diagnostic intentions, or that the doctor has failed to effectively convey key medical information, significantly increasing the probability of disputes. The platform further analyzes the specific dimensions of semantic deviation, identifying specific aspects of cognitive differences by comparing the semantic vectors of the medical staff and the patient across various dimensions. For example, a large difference in the dimension of disease severity may indicate that the doctor believes the condition is serious, but the patient has not fully recognized the risks; a large difference in the dimension of treatment necessity may indicate that the patient has doubts about the necessity of the treatment plan; a large difference in the dimension of cost expectation may indicate a significant gap between the patient's psychological expectations of medical costs and the actual costs. The specific analysis results of these dimensions serve as supplementary information, helping medical institutions understand the root causes of communication problems and develop targeted improvement measures. This step quantifies the degree of cognitive asymmetry between doctors and patients by calculating semantic deviation values, solving the problem that traditional methods cannot objectively assess the quality of doctor-patient communication. It enables precise measurement of the effectiveness of doctor-patient information transmission, allowing the risk warning system to promptly detect communication barriers and take intervention measures before disputes arise, significantly reducing the risk of disputes caused by poor communication.

[0088] This embodiment achieves comprehensive collection of doctor-patient communication content by acquiring medical treatment documents and patient interaction texts associated with active nodes, providing a data foundation for in-depth analysis of the quality of doctor-patient information transmission. Compared with traditional methods that rely solely on patient satisfaction surveys or complaint records, this scheme directly extracts information from the original texts of both doctors and patients, reflecting the actual situation of doctor-patient communication more realistically and comprehensively, avoiding distortion and omissions from secondary information. A pre-trained medical domain semantic mapping model transforms medical professional language and everyday speech into semantic vectors, and maps them to a unified semantic vector space through a cross-domain alignment layer, achieving semantic alignment across different language systems. Compared with traditional keyword matching or simple text similarity calculation methods, this scheme can deeply understand the semantic connotations of medical terminology and everyday language, accurately capturing the true meaning expressed by both doctors and patients, and solving the problem of semantic comparison difficulties caused by language differences. By calculating the semantic deviation value between the medical semantic vector and the patient semantic vector, a quantitative assessment of the degree of cognitive asymmetry between doctors and patients is achieved. This quantitative indicator can objectively reflect the gap in information understanding between doctors and patients, providing a scientific basis for identifying doctor-patient communication barriers. When the semantic deviation value exceeds a threshold, the system automatically triggers a high-risk warning, prompting medical institutions to strengthen communication with patients to ensure that patients fully understand the treatment plan and related risks. This technical solution not only improves the accuracy of risk warnings but also provides medical institutions with specific directions for improving doctor-patient communication. It enables medical staff to understand where patients have misunderstandings, thereby adjusting communication strategies and expressions, effectively reducing the incidence of disputes caused by poor information transmission, improving the quality of doctor-patient communication and patient satisfaction, and promoting the harmonious development of doctor-patient relationships.

[0089] In one embodiment of this invention, calculating the semantic deviation value between the medical semantic vector and the patient semantic vector in the semantic vector space includes the following steps: S610. In the semantic vector space, calculate the semantic overlap degree based on the semantic vectors of the medical staff and the semantic vectors of the patient. S620. Based on the comparison result between the semantic overlap degree and the preset coefficient, determine the semantic deviation value. When the semantic overlap degree is lower than the preset coefficient, it is determined that semantic loss has occurred, and the semantic deviation value is marked as a high-risk level.

[0090] In this embodiment, the risk warning platform calculates the semantic overlap between the medical and patient semantic vectors in a unified semantic vector space to quantify the consistency of understanding of medical information between the two parties. The semantic overlap is calculated using a cosine similarity algorithm. This algorithm assesses the directional similarity of vectors by measuring the cosine of the angle between two vectors, is unaffected by vector length, and accurately reflects the similarity of semantic content. The calculation formula is as follows: ,in This is a 256-dimensional semantic vector for medical prescriptions. It is a 256-dimensional semantic vector of the patient. and These are the component values ​​of the two vectors in the i-th dimension.

[0091] The cosine similarity value ranges from -1 to 1. A value of 1 indicates that the two vectors are completely in the same direction, meaning that the doctor and patient have a completely consistent understanding of the medical information. A value of 0 indicates that the two vectors are orthogonal, meaning that there is a significant difference in understanding between the doctor and patient. A negative value indicates that the two vectors are in opposite directions, meaning that there is a fundamental conflict in understanding between the doctor and patient. In medical communication scenarios, analysis of historical dispute cases reveals that when the semantic overlap is higher than 0.7, the doctor and patient can usually achieve effective communication, and the dispute rate is low. When the semantic overlap is between 0.3 and 0.7, there are some differences in understanding between the doctor and patient, requiring strengthened communication. When the semantic overlap is lower than 0.3, there is a serious deviation in understanding between the doctor and patient, and the risk of dispute is extremely high.

[0092] The platform also performs multi-dimensional decomposition and analysis of semantic overlap, grouping the 256-dimensional semantic vector according to semantic categories, such as disease cognition dimension (first 64 dimensions), treatment plan dimension (65-128 dimensions), risk expectation dimension (129-192 dimensions), and cost perception dimension (193-256 dimensions). It calculates the local semantic overlap for each dimension to identify specific aspects of differences in doctor-patient perception. This step, by calculating semantic overlap, achieves precise quantification of the consistency of doctor-patient understanding, solving the problem that traditional methods cannot objectively assess the effectiveness of doctor-patient communication. It provides a reliable numerical indicator for identifying differences in doctor-patient perception, enabling risk assessment to shift from subjective judgment to data-driven objective analysis.

[0093] After calculating the semantic overlap, the semantic deviation value is determined and a risk level is assigned based on the comparison between the semantic overlap value and a preset coefficient. The preset coefficient is set to 0.7, a threshold determined based on statistical analysis of numerous historical medical dispute cases, which effectively distinguishes between normal communication and communication with inherent risks. The semantic deviation value is calculated using a piecewise function, applying different calculation strategies according to different intervals of semantic overlap. .

[0094] When the semantic overlap is greater than or equal to 0.7, it indicates that the understanding between doctors and patients is basically consistent. The semantic deviation value is a simple linear transformation, ranging from 0 to 0.3, and is marked as low risk, indicated by green. When the semantic overlap is between 0.3 and 0.7, it indicates that there is a significant difference in understanding between doctors and patients. The semantic deviation value is multiplied by a magnification factor of 1.5, ranging from 0.45 to 1.05, and is marked as medium risk, indicated by yellow, suggesting the need to strengthen doctor-patient communication. When the semantic overlap is less than 0.3, semantic loss is determined, indicating a fundamental difference in understanding of medical information between doctors and patients. The semantic deviation value is multiplied by a magnification factor of 2, ranging from 1.4 to 1.7, and is marked as high risk, indicated by red.

[0095] High-risk semantic biases indicate that key medical information conveyed by doctors has not been correctly understood by patients, or that patients have seriously misunderstood the medical information, which can easily lead to medical disputes. The platform automatically generates risk warning reports, detailing the specific manifestations of semantic biases, such as "a significant difference between the patient's perception of the severity of the disease and the doctor's diagnosis," "the patient's insufficient understanding of the necessity of the treatment plan," and "a large gap between the patient's expected and actual medical costs." The reports are sent to the attending physician, department head, and patient relations coordinator, requiring immediate intervention measures, including arranging in-depth communication between the doctor and patient, restating the treatment plan in plain language, and providing written treatment instructions and cost lists, to ensure the patient fully understands the medical information and eliminates cognitive biases. This step, by calculating semantic bias values ​​in segments and marking risk levels, achieves a refined assessment of the severity of cognitive differences between doctors and patients, overcoming the limitations of traditional methods that cannot distinguish between different levels of communication problems. It enables risk warnings to automatically match appropriate intervention strategies based on the degree of bias, significantly improving the targeting and effectiveness of risk management.

[0096] This embodiment achieves a precise quantitative assessment of the consistency of doctor-patient information understanding by calculating the semantic overlap between the medical and patient semantic vectors in a unified semantic vector space. Compared with traditional methods that rely on patient satisfaction ratings or subjective feelings, this scheme, based on the mathematical calculation of semantic vectors, can objectively and accurately reflect the gap in information understanding between doctors and patients, eliminating the bias and uncertainty of subjective assessment. By determining the semantic deviation value based on the comparison result of semantic overlap with a preset coefficient, and using a piecewise function to differentiate and quantify different degrees of cognitive differences, a refined classification of risk levels is achieved. Compared with the traditional dichotomous judgment (risky or no risk), this scheme can distinguish between low, medium, and high risk levels, providing medical institutions with a more detailed basis for risk management, enabling intervention measures to match the risk level, and avoiding the problems of resource waste and insufficient intervention. The mechanism of determining semantic loss and marking it as a high-risk level when the semantic overlap is lower than the preset coefficient enables automatic identification and early warning of serious doctor-patient communication barriers. This technical solution not only improves the accuracy and timeliness of risk identification, but also provides medical institutions with actionable intervention guidance, enabling medical staff to promptly identify and correct patients' cognitive biases, ensuring the effective transmission of medical information, significantly reducing the incidence of disputes caused by poor communication, improving the quality of doctor-patient communication and patient satisfaction, and promoting the harmonious development of doctor-patient relationships.

[0097] In one embodiment of this example, after determining the semantic deviation value, the method further includes the following steps: S710. Determine the interpretation cost of the medical staff and the trust loss of the patient based on the semantic vectors of the medical staff and the semantic vectors of the patient. S720. Generate a strategy matrix based on explanation cost and trust loss, wherein the strategy matrix includes a set of medical strategies and a set of patient strategies. S730. Determine the strategy evolution trajectory based on the strategy matrix, wherein the strategy evolution trajectory is used to indicate the likelihood of a dispute.

[0098] Among them, determining the interpretation cost for doctors and the trust loss for patients based on the semantic vectors of doctors and patients includes: Extract the density and expression complexity features of professional terms from the semantic vector of medical prescriptions; Based on the density and complexity of technical terms, the additional explanatory work required by the medical staff to achieve effective communication is calculated. The additional workload of explanation is converted into time cost and cognitive cost, resulting in the explanation cost for the medical staff; Extract the comprehension level index and emotional stability index from the patient's semantic vector; Based on the comprehension level index and the emotional stability index, combined with the semantic deviation value, the decrease in the patient's trust in the treatment plan is calculated, and the patient's trust loss is obtained.

[0099] A strategy matrix is ​​generated based on interpretation cost and trust loss, including: Construct a two-dimensional strategy space, where the first dimension represents the strategy choice of the medical staff and the second dimension represents the strategy choice of the patient. Based on the cost of explanation, determine the cost difference between the medical staff's full explanation strategy and the brief explanation strategy; Based on trust loss, determine the difference in benefits for the patient from adopting a rational negotiation strategy and an irrational escalation strategy; Fill the corresponding cells in the two-dimensional strategy space with the cost difference and revenue difference to generate the strategy matrix; Calculate the Nash equilibrium point for each policy combination in the policy matrix and mark the stable policy combinations.

[0100] In this embodiment, after determining the semantic deviation value, the risk warning platform needs to further analyze the cost-benefit characteristics of both doctors and patients in the communication game to predict the evolution trend of disputes. This process first determines the explanation cost for doctors and the trust loss for patients based on the semantic vectors of doctors and patients.

[0101] To calculate the cost of medical interpretation, the platform first extracts the density and complexity features of medical terms from the medical terminology vector. The density of medical terms is calculated by statistically analyzing the proportion of medical terminology in the medical records, using a medical dictionary matching method to identify medical terms, including disease names, drug names, examination items, and medical indicators. The density calculation formula is as follows: ,in For the number of technical terms, This represents the total word count of the document. A higher density of technical terms indicates a more specialized document, but also a greater difficulty for patients to understand. Expression complexity is calculated by analyzing the complexity of sentence structure, using syntactic tree depth and the number of clauses as metrics. The complexity calculation formula is: ,in For the average syntax tree depth, The average number of clauses. and These are the weighting coefficients, set to 0.6 and 0.4 respectively.

[0102] Based on the density and complexity of technical terms, the additional explanatory workload required by medical staff to achieve effective communication is calculated. This additional explanatory workload reflects how much extra time and effort doctors need to invest in translating technical information into expressions that patients can understand. The calculation formula is as follows: ,in This is the semantic deviation value. The conversion factor is set to 2.5. A larger semantic deviation value indicates a greater gap in understanding between doctors and patients, and a greater amount of explanation work the doctor needs to invest. This additional explanation work is converted into time cost and cognitive cost to obtain the doctor's explanation cost. The time cost is calculated based on the doctor's average hourly wage. Assuming the average hourly wage is 300 yuan and each unit of explanation work requires 5 minutes, the time cost is: Cognitive cost reflects the psychological burden and communication difficulty experienced by doctors during the explanation process. It is quantified using cognitive load theory, with the following formula: The total cost of explanation for the medical staff is: .

[0103] To calculate patient trust loss, the platform first extracts comprehension and emotional stability indicators from the patient's semantic vector. The comprehension indicator is calculated by analyzing the accuracy of the patient's restatement of medical information and the relevance of their questions in the patient's interactive text. A higher comprehension level is indicated by the patient's ability to accurately restate the diagnosis and treatment plan, and by asking targeted questions. Comprehension level is rated from 0 to 1, calculated based on the degree of matching with a standard comprehension template. The emotional stability indicator uses a sentiment analysis model to identify the degree of emotional fluctuation in the patient's text, including the intensity and frequency of negative emotions such as anxiety, anger, and fear. Emotional stability is also rated from 0 to 1, with higher scores indicating more stable emotions.

[0104] Based on comprehension level indicators and emotional stability indicators, combined with semantic deviation values, the decrease in patient trust in the treatment plan was calculated. The decrease in trust reflects the degree of loss of patient trust due to information comprehension difficulties and emotional fluctuations. The calculation formula is as follows: ,in To understand the level of indicators, As an indicator of emotional stability, This is the semantic deviation value. The trust sensitivity coefficient is set to 1.8. Trust loss increases significantly when patients have low comprehension, are emotionally unstable, and exhibit large semantic deviations. (Patient's trust loss) The value ranges from 0 to 3, with higher values ​​indicating a greater loss of trust in the medical institution and a higher likelihood of irrational behavior by the patient. This step quantifies the explanation costs for the medical staff and the loss of trust for the patient, transforming implicit factors in doctor-patient communication into calculable numerical indicators. This solves the problem that traditional methods cannot assess the cost-benefit structure in doctor-patient games, providing a theoretical basis for predicting doctor-patient behavioral strategy choices and dispute evolution trends, and enabling risk warning to shift from static assessment to dynamic game analysis.

[0105] After determining the explanation costs for the medical staff and the trust losses for the patient, a strategy matrix is ​​generated based on these parameters to model the strategic game relationship between doctors and patients during the communication process. The strategy matrix construction process first establishes a two-dimensional strategy space. The first dimension represents the medical staff's strategy choices, including two options: a thorough explanation strategy and a concise explanation strategy. A thorough explanation strategy involves the doctor investing sufficient time and effort to explain the treatment plan, risks, and prognosis in clear and understandable language, ensuring the patient fully understands. A concise explanation strategy involves the doctor providing only basic information, using professional terminology, and explaining in a shorter time. The second dimension represents the patient's strategy choices, including two options: a rational negotiation strategy and an irrational escalation strategy. A rational negotiation strategy involves the patient communicating with the doctor through normal channels when encountering problems, rationally expressing their demands, and seeking solutions. An irrational escalation strategy involves the patient resorting to extreme measures such as complaints, disturbances, or lawsuits, attempting to obtain compensation or vent their dissatisfaction through pressure.

[0106] Based on the cost of explanation, determine the cost difference between the healthcare provider's full explanation strategy and the limited explanation strategy. When adopting the full explanation strategy, the healthcare provider bears the full explanation cost. This includes time and cognitive costs. When a simplified explanation strategy is adopted, the explanation cost for the medical staff is reduced to [missing information]. However, this will increase the potential costs of handling subsequent disputes. If the patient adopts an irrational escalation strategy, the medical institution will need to bear the costs of dispute resolution. This includes mediation costs, compensation costs, reputational damage, etc., and is usually set as follows: When the medical staff adopts a strategy of providing sufficient explanation, it can significantly reduce the probability of the patient resorting to irrational escalation strategies. Let the reduction coefficient be . .

[0107] Based on trust loss, the difference in payoffs for patients using rational negotiation and irrational escalation strategies is determined. The patient's payoff function considers two factors: satisfaction with problem-solving and the cost of action. When adopting a rational negotiation strategy, if the medical staff provides sufficient explanation, the patient's satisfaction is higher, and the payoff is... If the medical staff adopts a simplified explanation strategy, patient satisfaction will be lower, and the benefit will be... When irrational escalation strategies are adopted, the patient incurs time, energy, and psychological stress, with costs amounting to [amount missing]. However, additional compensation may be received, with an expected compensation value of [value missing]. Therefore, the net income is .

[0108] The cost and revenue differences are filled into the corresponding cells of the two-dimensional strategy space to generate a strategy matrix. The strategy matrix is ​​a 2×2 revenue matrix, and each cell contains a revenue pair (U) between the doctor and patient. m U pWhen the medical staff adopts a strategy of full explanation and the patient adopts a strategy of rational negotiation, the medical staff benefits from the following: The patient's benefit is When the medical staff adopts a strategy of full explanation and the patient adopts an irrational escalation strategy, the medical staff's benefit is... The patient's benefit is When the medical staff adopts a simplified explanation strategy and the patient adopts a rational negotiation strategy, the medical staff benefits from [the following]. The patient's benefit is When the medical staff adopts a simplified explanation strategy and the patient adopts an irrational escalation strategy, the medical staff's benefit is... The patient's benefit is .

[0109] Calculate the Nash equilibrium point for each strategy combination in the strategy matrix. A Nash equilibrium is a strategy combination where, given the opponent's strategy, neither party can unilaterally change their strategy to gain a higher payoff. In the calculation, first, the patient's strategy is fixed, and the payoffs of the doctor under the two strategies are compared to find the doctor's optimal response strategy; then, the doctor's strategy is fixed, and the payoffs of the patient under the two strategies are compared to find the patient's optimal response strategy. If a strategy combination is the intersection of the optimal response strategies of both parties, then that combination is the Nash equilibrium point. Calculations show that when trust loss is low... At this point, the strategy combination (full explanation, rational negotiation) represents the Nash equilibrium, where both doctors and patients tend to cooperate and communicate. However, when trust depletion is high... At certain times, a strategy combination (simply put, irrational escalation) may become a Nash equilibrium, at which point the doctor and patient are in conflict. Stable strategy combinations, i.e., the strategy combinations corresponding to the Nash equilibrium, are labeled as a basis for predicting doctor-patient behavior. This step, by constructing a strategy matrix and calculating the Nash equilibrium, achieves mathematical modeling of the doctor-patient game relationship, solving the problem that traditional methods cannot predict the strategy choices of doctor-patient behavior. It enables risk warning to analyze the evolution direction of doctor-patient interaction based on game theory principles, providing theoretical support for formulating effective intervention strategies.

[0110] After generating the strategy matrix, the strategy evolution trajectory is determined based on the strategy matrix to predict the dynamic evolution of the doctor-patient relationship and the likelihood of disputes. The strategy evolution trajectory is constructed based on evolutionary game theory, taking into account the dynamic adjustment process of strategy choices made by both doctors and patients.

[0111] First, based on the strategy matrix, determine the payoff functions for the medical staff when adopting a fully explained strategy and a simplified explanation strategy, and the payoff functions for the patient when adopting a rational negotiation strategy and an irrational escalation strategy. The expected payoff for the medical staff when adopting a fully explained strategy is: , where y is the probability that the patient adopts a rational negotiation strategy. The expected payoff for the medical staff adopting a simplified explanation strategy is: The expected benefit of the patient adopting a rational negotiation strategy is: , where x is the probability that the medical staff adopts a fully explanatory strategy. The expected payoff for the patient adopting an irrational escalation strategy is: .

[0112] Fitness differences are calculated based on the payoff functions of both the healthcare provider and the patient. Fitness differences reflect the degree of advantage of one strategy relative to another and are the core driving force behind strategy evolution. The fitness difference for the healthcare provider is as follows: The differences in adaptability among patients are as follows: According to the replication dynamics equation, the rate of change of strategy selection probabilities for both the medical staff and the patient over time is: , When the fitness difference is positive, the probability of choosing the corresponding strategy increases; when the fitness difference is negative, the probability of choosing the corresponding strategy decreases.

[0113] The equilibrium point of the policy evolution trajectory is determined based on fitness differences. The equilibrium point is the state where the policy selection probability no longer changes. and The equilibrium points are determined by solving the system of equations. Multiple equilibrium points can be obtained, including four corner points (0,0), (0,1), (1,0), and (1,1), as well as potential internal equilibrium points. The corner point (1,1) corresponds to the strategy combination (fully explained, rationally negotiated), which is an ideal cooperative equilibrium; the corner point (0,0) corresponds to the strategy combination (briefly explained, irrationally escalated), which is the worst conflict equilibrium. The convergence direction of the strategy evolution trajectory is determined based on the equilibrium points. The stability of the equilibrium point is judged by calculating the eigenvalues ​​of the Jacobian matrix. If all real parts of the eigenvalues ​​are negative, the equilibrium point is stable, and the system will converge to that point; if all real parts of the eigenvalues ​​are positive, the equilibrium point is unstable, and the system will move away from that point; if some eigenvalues ​​are positive and some are negative, the equilibrium point is a saddle point.

[0114] Based on the convergence direction, the phase trajectory in the policy space is calculated. The phase trajectory describes how the policy choice probabilities of both the doctor and the patient evolve over time, starting from an arbitrary initial state. A numerical simulation method is used, setting an initial state (x0, y0), and iteratively calculating the policy probabilities (x, y0) at subsequent time steps according to the replicating dynamic equation. t y t The evolution trajectory curves are plotted. By analyzing the evolution trajectories of a large number of initial states, the attraction domain can be identified, which is the set of initial states that converge to the same equilibrium point.

[0115] The strategy evolution trajectory is determined based on the convergence direction and phase trajectory. If the system's evolution trajectory converges to the cooperative equilibrium point (1, 1), it indicates that both the doctor and patient will ultimately choose the cooperative strategy, and the probability of dispute is low. If the system's evolution trajectory converges to the conflict equilibrium point (0, 0), it indicates that both the doctor and patient will ultimately fall into conflict, and the probability of dispute is extremely high. If the system has multiple stable equilibrium points, the final convergence point depends on the initial state. In this case, timely intervention is needed to change the initial state or evolution path of the system and guide the system towards the cooperative equilibrium. The platform predicts the future evolution direction based on the current strategy selection probabilities of both the doctor and patient, and calculates the time and probability of reaching the conflict equilibrium point as a quantitative indicator of dispute risk. This step, by constructing the strategy evolution trajectory and analyzing the convergence direction, enables the prediction of the dynamic evolution process of the doctor-patient relationship. It solves the problem that traditional methods can only perform static risk assessment and cannot predict risk evolution trends. This allows risk warnings to identify the deterioration trend of the doctor-patient relationship in advance, providing a window of opportunity for timely intervention and significantly improving the foresight and effectiveness of dispute prevention.

[0116] This embodiment quantifies the implicit cost-benefit factors in doctor-patient communication by determining the interpretive costs for doctors and the trust losses for patients based on the semantic vectors of both sides. It transforms the subjective feelings in the doctor-patient game into objective numerical indicators. Compared to traditional methods that only focus on explicit behaviors, this approach deeply analyzes the intrinsic motivations and constraints of both doctors and patients, revealing the underlying factors influencing their behavioral choices and providing a new perspective on understanding the formation mechanism of doctor-patient conflicts. By constructing a strategy matrix and calculating the Nash equilibrium point, it achieves mathematical modeling of the doctor-patient game relationship and identification of stable strategy combinations. Compared to traditional empirical judgment methods, this approach can accurately predict the strategy choices of both doctors and patients in different situations, providing a theoretical basis for developing targeted intervention strategies. By determining the strategy evolution trajectory and analyzing the convergence direction, it achieves the prediction of the dynamic evolution process of the doctor-patient relationship and the quantitative assessment of the probability of disputes. This technical solution can not only identify the current risk state but also predict future evolution trends, enabling medical institutions to take intervention measures before the doctor-patient relationship deteriorates, eliminating risks in their infancy. This technical solution significantly improves the foresight and scientific rigor of risk warning, provides medical institutions with decision support for optimizing doctor-patient communication strategies from a cost-benefit perspective, effectively reduces the incidence of disputes, improves the quality of medical services and patient satisfaction, and promotes the harmonious development of doctor-patient relationships.

[0117] In one embodiment of this invention, determining the policy evolution trajectory based on the policy matrix includes the following steps: S810. Determine the payoff functions for the medical staff when adopting a fully explained strategy and a simplified explained strategy, and the payoff functions for the patient when adopting a rational negotiation strategy and an irrational escalation strategy, based on the strategy matrix. S820. Calculate fitness differences based on medical benefit functions and patient benefit functions; S830. Determine the equilibrium point of the policy evolution trajectory based on fitness differences, and determine the convergence direction of the policy evolution trajectory based on the equilibrium point. S840. Calculate the phase trajectory in the policy space based on the convergence direction; S850: Determine the strategy evolution trajectory based on the convergence direction and phase trajectory.

[0118] Following S850, the following steps may also be included: Based on the endpoint of the strategy evolution trajectory, the final evolution state of the doctor-patient relationship is determined. When the endpoint converges to an irrational escalation strategy combination, the risk of dispute is determined to be high-risk.

[0119] In another embodiment of this application, the following steps may be included after S850: The system determines whether a cognitive bias dispute warning is triggered based on the strategy evolution trajectory, and when a cognitive bias dispute warning is triggered, it determines the expected bias. The cognitive load is calculated based on the expected deviation, where the cognitive load is used to offset the expected deviation. Based on the cognitive load, supplementary explanatory materials are retrieved from a pre-set medical knowledge base and pushed to the corresponding medical staff to guide them in carrying out targeted risk mitigation interventions.

[0120] Specifically, cognitive loading is calculated based on expected deviations, including: Determine the cognitive gap between the current state of doctor-patient communication and the target stable state based on expected deviations; The amount of information compensation that can eliminate the cognitive gap is calculated based on the cognitive gap. The cognitive load is determined based on the information compensation amount and the preset information transmission efficiency coefficient.

[0121] Based on cognitive load, supplementary explanatory materials are retrieved from a pre-defined medical knowledge base, including: Analyze cognitive load to determine missing knowledge dimensions and levels of understanding; A knowledge graph index is established in the medical knowledge base, which includes medical concept nodes and concept relationships. Based on the missing knowledge dimensions and understanding levels, semantic matching retrieval is performed in the knowledge graph index, and the data content with the highest relevance to the semantic matching retrieval results is selected as supplementary explanatory materials.

[0122] In this embodiment, the risk warning platform determines the payoff functions for both the doctor and patient under different strategy choices based on the strategy matrix, providing a mathematical foundation for subsequent evolutionary game analysis. The doctor's payoff function considers two strategy choices: a fully explanatory strategy and a simplified explanatory strategy. When the doctor adopts the fully explanatory strategy, it needs to invest the full explanatory cost. This cost includes time and cognitive costs. Meanwhile, the medical provider's benefit also depends on the patient's strategy choice. Let y be the probability that the patient adopts a rational negotiation strategy, and 1-y be the probability that they adopt an irrational escalation strategy. When the patient adopts a rational negotiation strategy, the medical provider does not incur additional dispute resolution costs; when the patient adopts an irrational escalation strategy, the medical provider will incur dispute resolution costs. However, due to the adoption of a thorough explanation strategy, the probability of disputes decreased, and the actual cost was [missing information]. ,in The risk reduction factor is typically set to 0.3. Therefore, the expected return function for the medical staff adopting a fully explained strategy is: .

[0123] When the medical staff adopts a simplified explanation strategy, the explanation cost is reduced to However, the risk of disputes increases significantly. When patients adopt irrational escalation strategies, the medical institution needs to bear the full cost of dispute resolution. Therefore, the expected return function of the medical staff's simplified explanation strategy is: .

[0124] The patient's payoff function also considers two strategy choices: rational negotiation and irrational escalation. Let x be the probability that the doctor will adopt a full explanation strategy, and 1-x be the probability that they will adopt a simplified explanation strategy. When the patient adopts a rational negotiation strategy, the payoff mainly comes from the satisfaction with the problem-solving process, which is related to the sufficiency of the doctor's explanation and the patient's trust loss. When the doctor adopts a full explanation strategy, the patient's satisfaction is higher, with a base payoff of 10, minus the trust loss. The net income after that is When the medical staff adopts a simplified explanation strategy, patient satisfaction is lower, with a base benefit of 5 and a net benefit of [missing value]. Therefore, the expected payoff function for the patient adopting a rational negotiation strategy is: .

[0125] When patients adopt irrational escalation strategies, they need to bear the costs of their actions. This includes time costs, energy costs, and psychological stress, typically set at 8. However, patients may obtain additional compensation through complaints, lawsuits, etc., with the expected compensation value being [value missing]. It is typically set to 15. Simultaneously, trust erosion reduces the psychological burden on the patient to engage in irrational behavior, thus the net benefit is... The expected payoff function for the patient adopting an irrational escalation strategy is: This payoff function is independent of the doctor's strategy choice because the direct payoff for the patient to adopt an irrational escalation strategy is the same regardless of the doctor's strategy. However, the doctor's strategy affects the probability of the patient's success in adopting that strategy. This step, by establishing payoff functions for both doctors and patients, achieves a quantitative model of the doctor-patient game relationship. It solves the problem that traditional methods cannot accurately describe the motivations and decision-making logic of doctor-patient behavior, providing a mathematical tool for predicting doctor-patient strategy choices and relationship evolution, and enabling risk warnings to be scientifically predicted based on a rigorous theoretical framework.

[0126] After determining the payoff functions for both the doctor and patient, fitness differences are calculated based on these functions to quantify the relative advantages between different strategies. Fitness difference is a core concept in evolutionary game theory, reflecting the survival advantage of one strategy relative to another, and is the fundamental driving force behind strategy evolution. The fitness difference for the doctor is defined as the difference between the expected payoff of the fully explained strategy and the expected payoff of the poorly explained strategy: After simplification, we get: This expression suggests that the difference in medical adaptation depends on the cost of interpretation. Dispute resolution costs The probability, y, of adopting a rational negotiation strategy with the patient. In such cases, a full explanation strategy is preferred over a brief explanation strategy, and medical staff tend to choose a full explanation; when In such cases, a simplified explanation strategy is preferred over a full explanation strategy, and medical professionals tend to choose simplified explanations. This is achieved through solving... The critical probability can be obtained as follows: When the patient is more likely to adopt a rational negotiation strategy... In cases where the situation is complex, a more thorough explanation strategy is preferable for the medical staff; conversely, a more concise explanation strategy is preferable for the medical staff.

[0127] The patient's fitness difference is defined as the difference between the expected payoff of a rational negotiation strategy and the expected payoff of an irrational escalation strategy: After simplification, we get: This expression suggests that the patient's fitness variation depends on the probability x of the physician adopting a fully explanatory strategy and the patient's trust loss. .when In such cases, rational negotiation strategies are superior to irrational escalation strategies, and the patient tends to choose rational negotiation; when In such cases, irrational escalation strategies are preferred over rational negotiation strategies, and the patient tends to choose irrational escalation. This can be addressed by solving... The critical probability can be obtained as follows: When the medical staff adopts a strategy of full explanation, the probability is higher than In such cases, it is better for the patient to choose a rational negotiation strategy; conversely, it is better for the patient to choose an irrational escalation strategy.

[0128] The calculation of fitness differences reveals the interdependence between doctor-patient strategy choices. The doctor's optimal strategy depends on the patient's strategy choice probability, and vice versa, forming a complex dynamic game relationship. The platform monitors the strategy choice probabilities of both doctors and patients in real time, calculates the current fitness difference, and predicts the direction and speed of strategy evolution. This step quantifies the relative advantages between different strategies by calculating fitness differences, solving the problem that traditional methods cannot assess the driving force of strategy choice. It provides a theoretical basis for predicting the dynamic changes in doctor-patient behavior, enabling risk warnings to identify critical points and turning points in strategy evolution and promptly detect signs of deteriorating doctor-patient relationships.

[0129] After calculating the fitness differences, the equilibrium point of the strategy evolution trajectory is determined based on these differences, and the convergence direction of the trajectory is determined according to the equilibrium point. The equilibrium point refers to the state where the strategy selection probability no longer changes; it is a stable or critical state of system evolution. According to the replication dynamics equation, the rates of change of the strategy selection probability for the medical staff and the patient over time are respectively: , The equilibrium point satisfies the condition. and By solving this system of equations, five equilibrium points can be obtained: four corner points (0,0), (0,1), (1,0), and (1,1) and one internal equilibrium point. .

[0130] The corner point (0, 0) corresponds to the strategy combination (brief explanation, irrational escalation), representing the worst conflict equilibrium. This indicates that both the doctor and patient choose a non-cooperative strategy, inevitably leading to a dispute. The corner point (0, 1) corresponds to the strategy combination (brief explanation, rational negotiation), indicating that the doctor chooses a brief explanation while the patient remains rational. This state is usually unstable because a brief explanation increases the patient's dissatisfaction. The corner point (1, 0) corresponds to the strategy combination (full explanation, irrational escalation), indicating that the doctor provides a full explanation but the patient still engages in irrational behavior. This situation is less common because a full explanation usually reduces the patient's irrational tendencies. The corner point (1, 1) corresponds to the strategy combination (full explanation, rational negotiation), representing the ideal cooperative equilibrium. This indicates that both the doctor and patient choose a cooperative strategy, minimizing the risk of dispute. Internal Equilibrium Point It is a mixed strategy equilibrium, which means that both doctors and patients randomly choose strategies with a certain probability. This equilibrium point is usually an unstable saddle point.

[0131] To determine the stability and convergence direction of each equilibrium point, it is necessary to calculate the Jacobian matrix of the equilibrium point and analyze its eigenvalues. The Jacobian matrix is: At the equilibrium point, calculate the eigenvalues ​​of the Jacobian matrix. If all real parts of the eigenvalues ​​are negative, the equilibrium point is a stable point and the system will converge to that point. If all real parts of the eigenvalues ​​are positive, the equilibrium point is an unstable point and the system will move away from that point. If the eigenvalues ​​are both positive and negative, the equilibrium point is a saddle point, and the system converges in some directions and diverges in others.

[0132] Calculations revealed that the corner point (1, 1) is a stable attractor under most parameter conditions, indicating that the system naturally converges to a cooperative equilibrium when both doctors and patients tend to cooperate. The corner point (0, 0) exhibits higher trust loss. And the cost of explanation is high. This can sometimes become a stable attractor, meaning that when patient trust is severely eroded and the medical staff's willingness to explain is insufficient, the system will fall into a conflict equilibrium. Internal equilibrium point. Typically, saddle points act as watersheds. Systems initially positioned above a saddle point converge to a cooperative equilibrium, while those below converge to a conflict equilibrium. Stability analysis of the equilibrium point determines the convergence direction of the strategy evolution trajectory, providing a basis for predicting the final state of the doctor-patient relationship. This step, by determining the equilibrium point and convergence direction, enables the prediction of the endpoint of the doctor-patient relationship's evolution, solving the problem that traditional methods cannot determine the final direction of the doctor-patient relationship. It allows risk warnings to identify in advance whether the doctor-patient relationship will deteriorate into a dispute, providing a clear objective and direction for developing intervention strategies.

[0133] After determining the convergence direction, the phase trajectory in the policy space is calculated based on the convergence direction to visualize the dynamic evolution of the doctor-patient relationship. The phase trajectory refers to the trajectory in the two-dimensional policy space starting from any initial state. Starting point, the probability of strategy choices for both doctors and patients. The trajectory curve evolves over time. The calculation process employs a numerical simulation method, based on iterative calculations using a replicated dynamic equation. A time step is set. From the initial state Begin by calculating the policy probability for the next time step based on the replication dynamics equation: , Repeat the iterations until the system converges to a certain equilibrium point or reaches the preset maximum number of iterations.

[0134] To comprehensively understand the system's evolutionary behavior, the platform sets multiple different initial states, covering various regions of the policy space, and calculates phase trajectories for each. By analyzing a large number of phase trajectories, attraction domains can be identified, which are the sets of initial states that converge to the same equilibrium point. The boundaries of attraction domains are formed by stable manifolds with saddle points, serving as watersheds in system evolution. Systems whose initial states are located within the attraction domain of cooperative equilibrium will eventually converge to cooperative equilibrium, while systems whose initial states are located within the attraction domain of conflict equilibrium will eventually converge to conflict equilibrium. The platform maps the current actual state of the doctor-patient relationship onto the policy space, determines which attraction domain it falls within, and thus predicts the final evolutionary direction of the doctor-patient relationship. Simultaneously, it calculates the distance from the current state to the boundary of the attraction domain; this distance reflects the system's stability, with smaller distances indicating a greater susceptibility to disturbances and a shift to another equilibrium. This step, through the calculation of phase trajectories, enables visualization and quantitative analysis of the doctor-patient relationship's evolutionary process, solving the problem of traditional methods' inability to dynamically track changes in the doctor-patient relationship. It allows risk warnings to monitor the evolutionary state of the doctor-patient relationship in real time, promptly detecting signs of the system approaching a critical point, and providing a scientific basis for selecting intervention timing.

[0135] After calculating the phase trajectory, the strategy evolution trajectory is determined based on the convergence direction and the phase trajectory, comprehensively assessing the evolutionary trend of the doctor-patient relationship and the likelihood of disputes. The strategy evolution trajectory includes not only the evolutionary path from the current state to the equilibrium point, but also key information such as evolution speed, arrival time, and stability. Evolution speed is reflected by the magnitude of fitness differences; the greater the fitness difference, the faster the strategy adjustment and the faster the system evolution. Arrival time is estimated through numerical simulation, representing the time required to evolve from the current state to the equilibrium point, reflecting the urgency of the dispute risk. Stability is measured by the distance from the current state to the boundary of the attraction domain; a larger distance indicates a more stable system and less susceptible to external disturbances that could alter its evolutionary direction.

[0136] The platform will generate a risk assessment report based on the strategy evolution trajectory. The report includes the following: the strategy probability of the current doctor-patient relationship status. The platform predicts the evolutionary endpoint (cooperative or conflict equilibrium), the estimated time to reach the endpoint, the stability score of the evolutionary path, and key risk inflection points. If the predicted evolutionary endpoint is a conflict equilibrium, the platform automatically triggers a high-risk warning and calculates the required intervention intensity. The intervention intensity is determined based on the distance from the current state to the boundary of the attraction domain; the closer the distance, the greater the required intervention intensity. The platform also simulates the impact of different intervention measures on the evolutionary trajectory; for example, increasing the medical staff's explanatory input can increase the x-value, while improving the patient's trust level can decrease it. This value can change the direction of the system's evolution, guiding the system towards cooperative equilibrium.

[0137] After determining the strategy evolution trajectory, the final evolutionary state of the doctor-patient relationship is judged based on the endpoint of the trajectory. When the endpoint converges to an irrational escalation strategy combination, i.e., the corner point (0,0) or its vicinity, the dispute risk is determined to be high-risk. A high-risk dispute indicates that the doctor-patient relationship has entered an irreversible deterioration phase; without strong intervention, a dispute is inevitable. The platform immediately activates its emergency response mechanism, notifying the medical department, patient relations office, and hospital management to organize a multidisciplinary expert team for consultation and intervention. This includes comprehensive measures such as reassessing treatment plans, arranging high-level doctor-patient communication, and providing psychological counseling, aiming to reverse the evolutionary trend and prevent disputes. This step, by determining the strategy evolution trajectory and judging the final evolutionary state, achieves accurate prediction and hierarchical management of dispute risk, solving the problem that traditional methods cannot quantify the likelihood of disputes. It enables risk warnings to provide clear risk levels and intervention recommendations, significantly improving the targeting and effectiveness of dispute prevention.

[0138] After determining the strategy evolution trajectory, the platform can also determine whether to trigger a cognitive bias dispute warning based on the strategy evolution trajectory. The triggering conditions for a cognitive bias dispute warning include: the strategy evolution trajectory pointing towards conflict equilibrium, the semantic bias value exceeding a high-risk threshold, and the patient's trust loss exceeding a critical value. When the triggering conditions are met, the platform will determine the expected bias, that is, the gap between the patient's expectation of the medical outcome and the actual possible outcome. The expected bias is calculated by analyzing the expectation dimension in the patient's semantic vector and the prognostic assessment dimension in the medical staff's semantic vector, using the following formula: ,in For the patient's expectations, This is the prognostic assessment value for the medical staff.

[0139] Cognitive loading is calculated based on the expected deviation. This cognitive loading is used to offset the expected deviation, adjusting the patient's expectations to a reasonable range. The calculation process first determines the cognitive gap between the current doctor-patient communication state and the target stable state based on the expected deviation. The cognitive gap is: The greater the trust loss, the greater the cognitive gap. Then, based on the cognitive gap, the amount of information compensation needed to eliminate it is calculated. The information compensation amount is: Where k is the information demand coefficient, determined based on the patient's education level and comprehension ability, typically set between 1.2 and 2.0. Finally, the cognitive load is determined based on the information compensation amount and the preset information transmission efficiency coefficient. The information transmission efficiency coefficient reflects the effectiveness of information delivery by medical staff, typically set to 0.7, indicating that 70% of the delivered information can be effectively absorbed by the patient. The cognitive load is: ,in This is the information transmission efficiency coefficient.

[0140] Based on the cognitive load, supplementary explanatory materials are retrieved from a pre-defined medical knowledge base and pushed to the corresponding medical staff. The retrieval process first analyzes the cognitive load to determine the missing knowledge dimensions and understanding levels. Knowledge dimensions include disease knowledge, treatment knowledge, risk knowledge, and cost knowledge, while understanding levels include basic cognition, in-depth understanding, and application ability. Then, a knowledge graph index is built in the medical knowledge base, including medical concept nodes and concept relationships; for example, there are relationships between the "diabetes" node and the "insulin" node, and the "complications" node. Based on the missing knowledge dimensions and understanding levels, semantic matching retrieval is performed in the knowledge graph index, using a graph traversal algorithm to find the knowledge nodes most relevant to the patient's cognitive deficiencies. The content with the highest relevance to the semantic matching retrieval results is selected as supplementary explanatory materials, including popular science articles, illustrative explanations, video tutorials, and other formats. The platform pushes the supplementary explanatory materials to the corresponding medical staff and provides usage suggestions, guiding them to use these materials to communicate with patients in a targeted manner, complete risk mitigation interventions, eliminate cognitive biases, and reduce the risk of disputes. This extended step, by calculating cognitive load and providing targeted supplementary information, achieves closed-loop management from risk warning to intervention implementation. It solves the problem that traditional methods can only provide warnings but lack specific intervention guidance, enabling medical staff to obtain scientific communication tools and resource support, significantly improving the effectiveness of intervention measures and effectively reducing the incidence of disputes.

[0141] This embodiment achieves precise mathematical modeling of the doctor-patient game relationship by determining the payoff functions of both parties based on the strategy matrix, transforming complex doctor-patient interactions into a computable mathematical problem. Compared to traditional methods relying on experience-based judgment, this scheme, based on the rigorous framework of game theory, can accurately describe the decision-making logic and behavioral motivations of both parties, providing a scientific basis for predicting the evolution of the doctor-patient relationship. By calculating fitness differences and determining equilibrium points and convergence directions, it enables the prediction of the evolutionary trend of the doctor-patient relationship and the judgment of its final state. Compared to traditional static risk assessment methods, this scheme can dynamically track the changes in the doctor-patient relationship, identify signs of deterioration in advance, and provide a window of opportunity for timely intervention. By calculating phase trajectories and determining strategy evolution trajectories, it achieves visualization and quantitative analysis of the doctor-patient relationship evolution process, enabling medical institutions to intuitively understand the current state, evolutionary direction, and time of reaching the critical point of the doctor-patient relationship, providing clear objectives and timing for formulating intervention strategies. By triggering cognitive bias dispute early warnings and calculating cognitive load, a seamless connection is achieved from risk warning to intervention implementation. This provides healthcare professionals with targeted supplementary information and communication guidance, significantly improving the scientific rigor and effectiveness of intervention measures. This technical solution not only enhances the accuracy and foresight of dispute risk warnings but also provides medical institutions with systematic risk management tools and intervention methods, effectively reducing the incidence of disputes, improving the quality of doctor-patient communication and patient satisfaction, and promoting harmonious doctor-patient relationships.

[0142] This application embodiment also provides a server, including: The memory is configured to store instructions; and The processor is configured to retrieve instructions from memory and, when executing instructions, implement the aforementioned service dispute risk warning method based on multi-dimensional data features.

[0143] This application also provides a risk warning platform, which includes a cloud server and a client. The client is used to send a request to the cloud server, and the cloud server is used to execute the operation steps of any of the above methods according to the request.

[0144] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0145] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0146] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0147] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0148] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0149] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0150] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0151] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0152] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A service dispute risk early warning method based on multi-dimensional data features, characterized in that, Applied to servers where a risk warning platform is deployed, the method includes: The risk warning platform acquires multi-source heterogeneous data in real time, including structured diagnosis and treatment data within the hospital and unstructured feedback data outside the hospital. Real-time acquisition of diagnosis and treatment data and electronic medical record data, and establishment of a spatiotemporal topology map of patient diagnosis and treatment behavior according to timestamps and spatial departments. The spatiotemporal topology map includes multiple behavior nodes and topological connection relationships connecting behavior nodes. Calculate the outlier centrality of each behavioral node in the spatiotemporal topology graph; Active nodes are identified based on outlier centrality, and active nodes are behavioral nodes whose outlier centrality exceeds a preset outlier threshold. Retrieve medical records and patient interaction texts associated with active nodes; Medical treatment documents and patient interaction texts are converted into medical semantic vectors and patient semantic vectors, respectively, and then mapped to a unified semantic vector space. In the semantic vector space, the semantic overlap is calculated based on the semantic vectors of the medical staff and the semantic vectors of the patient. Based on the comparison between semantic overlap and preset coefficient, a semantic deviation value is determined to quantify the degree of cognitive asymmetry between doctors and patients in the information transmission process. When the semantic overlap is lower than the preset coefficient, semantic loss is determined and the semantic deviation value is marked as a high-risk level. Extract the density and expression complexity features of professional terms from the semantic vector of medical prescriptions; Based on the density and complexity of technical terms, the additional explanatory work required by the medical staff to achieve effective communication is calculated. The additional workload of explanation is converted into time cost and cognitive cost, resulting in the explanation cost for the medical staff; Extract the comprehension level index and emotional stability index from the patient's semantic vector; Based on the comprehension level index and the emotional stability index, combined with the semantic deviation value, the decrease in the patient's trust in the treatment plan is calculated, and the patient's trust loss is obtained. Construct a two-dimensional strategy space, where the first dimension represents the strategy choice of the medical staff and the second dimension represents the strategy choice of the patient. Based on the cost of explanation, determine the cost difference between the medical staff's full explanation strategy and the brief explanation strategy; Based on trust loss, determine the difference in benefits for the patient from adopting a rational negotiation strategy and an irrational escalation strategy; Fill the corresponding cells in the two-dimensional strategy space with the cost difference and revenue difference to generate the strategy matrix; Calculate the Nash equilibrium point for each policy combination in the policy matrix and mark the stable policy combinations; The strategy evolution trajectory is determined based on the strategy matrix, where the strategy evolution trajectory is used to indicate the likelihood of a dispute occurring; Determine the behavioral characteristics of active nodes and use these characteristics as auxiliary input features. Feature extraction is performed on structured diagnosis and treatment data within the hospital and unstructured feedback data outside the hospital to obtain a feature set, which includes multiple feature indicators; Acquire historical dispute data, and assign weights to each feature index in the feature set based on the historical dispute data to generate a multi-dimensional feature index matrix; By analyzing the multidimensional feature index matrix through a pre-set machine learning model, risk scores and high-risk patient profiles are generated. When the risk score exceeds a preset threshold, an early warning signal is triggered, and the corresponding intervention strategy is automatically matched according to the level of the risk score. Intervention tasks are generated based on the intervention strategy, and the execution status of the intervention tasks is continuously tracked. When the execution status is completed, feedback results are generated and displayed on the risk warning platform. The feedback results are used to update the parameters of the machine learning model.

2. The method according to claim 1, characterized in that, Calculate the outlier centrality of each behavioral node in the spatiotemporal topology graph, including: Count the access frequency of each behavior node within a preset time window; Calculate the topological distance between each behavior node and other behavior nodes; Based on the access frequency and topological distance, the deviation of each behavioral node from the normal diagnosis and treatment path is calculated to obtain the outlier centrality.

3. The method according to claim 1, characterized in that, Determine the policy evolution trajectory based on the policy matrix, including: Based on the strategy matrix, determine the payoff functions for the medical staff when adopting a fully explained strategy and a simplified explained strategy, and determine the payoff functions for the patient when adopting a rational negotiation strategy and an irrational escalation strategy. The fitness difference is calculated based on the medical benefit function and the patient benefit function. The equilibrium point of the policy evolution trajectory is determined based on fitness differences, and the convergence direction of the policy evolution trajectory is determined based on the equilibrium point. Calculate the phase trajectory in the policy space based on the convergence direction; The strategy evolution trajectory is determined based on the convergence direction and phase trajectory.

4. A server, characterized in that, include: The memory is configured to store instructions; as well as A processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the service dispute risk warning method based on multidimensional data features according to any one of claims 1 to 3.

5. A risk early warning platform, characterized in that, The risk warning platform includes a cloud server and a client. The client is used to send a request to the cloud server, and the cloud server is used to execute the operation steps of any one of the methods described in claims 1-3 according to the request.

Images