Doctor operation dynamic authorization method and system based on AI intelligent analysis

Abstract

The invention provides a physician operation dynamic authorization method and system based on AI intelligent analysis, and the method comprises the steps: firstly obtaining the multi-modal historical data of a target physician, constructing a multi-dimensional capability portrait model of the physician through a multi-modal feature fusion and deep learning representation method, and granting a basic operation authority according to the model; meanwhile, the system obtains recent workload data of the doctor from a hospital information system, quantifies the current working state of the doctor into a doctor state model, and carries out risk quantitative modeling on a surgical task planned to be executed. And then, the system integrates the ability portrait model, the state model and the operation risk model, constructs a risk prediction model by utilizing machine learning, and predicts the risk level of the doctor currently executing the specific operation. And finally, triggering a multi-level early warning mechanism according to the predicted risk level and the daily behavior score data, and automatically generating a dynamic adjustment strategy of the doctor operation permission according to a preset intervention strategy rule base.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent data processing, and particularly relates to a method and system for dynamically authorizing a physician's operation based on AI intelligent analysis. BACKGROUND

[0002] The development of physician operation permission management technology is closely related to the evolution of medical safety management system. The traditional physician permission management originated from the hospital department system and professional division of labor, and has experienced a development process from experience to standardization. The early permission management mainly relied on the title, experience and experience of physicians, and adopted the "apprentice system" mode for skill inheritance and evaluation. With the promotion of medical standardization, countries gradually established a permission management system based on qualification certification, such as specialist system and operation grading management. In practical application, the existing permission management technology mainly embodies three modes: one is a static authorization mode based on qualification, that is, according to fixed indicators such as education, title, training experience, etc. of the physician, the corresponding operation permission is granted; two is a periodic adjustment mode based on regular review, which usually conducts a comprehensive evaluation on the physician once a year or every few years, and adjusts the permission according to the evaluation result; three is a passive adjustment mode based on major events, that is, when the physician has a medical accident or complaint, the permission review and adjustment are triggered.

[0003] These traditional modes have played a certain role in medical practice, but with the changes in the medical environment, their limitations have become increasingly prominent. Modern hospitals are facing challenges such as diversification of operation types, acceleration of technology updates, and increase of medical risks, and the traditional static management mode is difficult to adapt to this complex and dynamic environment. Especially in large general hospitals, the work intensity of physicians is high, the pressure is high, and the task is complex, and their state and performance may fluctuate significantly, and the static permission management is difficult to capture such changes. At the same time, with the development of medical informatization, hospitals have accumulated a large amount of data related to physician work, including electronic medical records, operation records, training archives, etc., which provides a data basis for realizing more refined permission management. However, how to use computers to quickly and accurately analyze these data to complete the objective evaluation and risk prediction of physician ability has become a new topic in the field of medical management. SUMMARY

[0004] The present application provides a method and system for dynamically authorizing a physician's operation based on AI intelligent analysis to solve the above technical problems.

[0005] In a first aspect, the present application provides a method for dynamically authorizing a physician's operation based on AI intelligent analysis, which comprises the following steps:

[0006] Obtaining multi-modal historical data related to the target physician;

[0007] construct a multi-dimensional capability portrait model of the target physician based on multi-modal historical data and by multi-modal feature fusion and deep learning representation methods, and grant the target physician a surgery permission according to the multi-dimensional capability portrait model;

[0008] obtain recent workload data of the target physician from a hospital information system, model a current working state of the target physician as a physician state model according to the recent workload data, and model inherent risks of a target surgery task planned to be performed by the target physician as a surgery risk model;

[0009] integrate the multi-dimensional capability portrait model, the physician state model and the surgery risk model, and construct a risk prediction model based on a machine learning model to predict a predicted risk level of the target physician in performing the target surgery task;

[0010] trigger a multi-level early warning mechanism according to the predicted risk level and daily behavior score data in the multi-modal historical data, and automatically generate a dynamic adjustment strategy of the surgery permission of the target physician according to a preset intervention strategy rule base.

[0011] Optionally, the obtaining of the multi-modal historical data related to the target physician comprises the following steps:

[0012] obtaining structured archive data of the target physician through a standardized data interface;

[0013] collecting surgery simulation historical data of the target physician from a surgery simulator, the surgery simulation historical data comprising surgery simulator operation data, physiological index data and eye movement pattern data of the target physician in historical surgery simulation operations;

[0014] performing data cleaning and data integration on the structured archive data and the surgery simulation historical data to form a multi-modal historical data set of the target physician.

[0015] Optionally, the constructing of the multi-dimensional capability portrait model of the target physician based on the multi-modal historical data and by the multi-modal feature fusion and deep learning representation methods comprises the following steps:

[0016] respectively performing feature extraction and independent coding on different modal data in the multi-modal historical data;

[0017] adopting a preset fusion strategy to fuse feature vectors obtained after independent coding of each modal data, to generate unified multi-modal fusion features;

[0018] inputting the multi-modal fusion features into a deep neural network, and training an evaluation model outputting quantitative scores of the target physician in pre-defined capability dimensions by a supervised learning manner, to constitute the multi-dimensional capability portrait model;

[0019] Applying model explainability techniques to analyze key input features that influence the evaluation results of the multidimensional capability profile model;

[0020] Monitoring dynamic changes of the key input features to obtain a dynamically updated multidimensional capability profile model.

[0021] Optionally, the predefined capability dimensions at least include technical skill dimension, cognitive skill dimension, non-technical skill dimension, and learning adaptability dimension.

[0022] Optionally, the step of obtaining the recent workload data of the target physician from the hospital information system, quantitatively modeling the current working state of the target physician into a physician state model according to the recent workload data, and quantitatively modeling the inherent risk of the target surgical task planned to be performed by the target physician into a surgery risk model includes the following steps:

[0023] Obtaining the recent workload data of the target physician from the hospital information system, and estimating the fatigue index of the target physician according to the recent workload data to form a quantitatively modeled physician state model of the current working state of the target physician;

[0024] Evaluating the surgery inherent complexity and patient-specific risk factors based on the surgery type, estimated duration, and patient basic conditions of the target surgical task planned to be performed by the target physician;

[0025] Combining the surgery inherent complexity and patient-specific risk factors, and constructing a surgery risk model of the target surgical task performed by the target physician by using a risk scoring model;

[0026] Standardizing the quantitatively modeled physician state model and the surgery risk model.

[0027] Optionally, the step of constructing a risk prediction model based on the machine learning model to predict the predicted risk level of the target physician in the current execution of the target surgical task includes the following steps:

[0028] Constructing a risk prediction model based on the machine learning model, and the input features of the risk prediction model at least include the multidimensional capability profile model, the physician state model, and the surgery risk model;

[0029] Training and verifying the risk prediction model by using historical surgery case data;

[0030] Performing performance evaluation and calibration on the trained risk prediction model;

[0031] Applying the risk prediction model to output a predicted risk level of the target physician in the current execution of the target surgical task.

[0032] Optionally, the triggering of the multi-level early warning mechanism according to the predicted risk level and the daily behavior score data in the multi-modal historical data to automatically generate a dynamic adjustment strategy for the target physician's operation permission according to the preset intervention strategy rule base comprises the following steps:

[0033] Setting a threshold value based on the predicted probability or risk level of adverse events and a triggering condition of the daily behavior score data in the multi-modal historical data to activate different levels of risk early warning;

[0034] Establishing a multi-level early warning mechanism including at least information prompt level, suggestion guidance level and warning review level;

[0035] According to the preset intervention strategy rule base and according to the risk warning in the early warning level of the multi-level early warning mechanism, a dynamic adjustment strategy for the target physician's operation permission is matched from the preset intervention measures.

[0036] Optionally, the method further comprises the following steps:

[0037] Establishing an automatic clinical outcome data and user feedback data data collection process;

[0038] Using the data collected by the data collection process to continuously monitor the performance of the multi-dimensional capability portrait model and the risk prediction model in real world data;

[0039] When the model performance of the multi-dimensional capability portrait model or the risk prediction model is lower than the preset threshold, the multi-dimensional capability portrait model and the risk prediction model are adjusted using the accumulated clinical outcome data and user feedback data.

[0040] In a second aspect, the present application also provides a physician operation dynamic authorization system based on AI intelligent analysis, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to realize the physician operation dynamic authorization method based on AI intelligent analysis in the first aspect.

[0041] In a third aspect, the present application also provides a computer readable storage medium, wherein the computer readable storage medium stores instructions which, when executed by a processor, cause the processor to be configured to execute the physician operation dynamic authorization method based on AI intelligent analysis according to the first aspect.

[0042] The present application has the following beneficial effects:

[0043] The present application solves the technical problems of strong subjectivity and poor timeliness of traditional physician ability evaluation by introducing advanced artificial intelligence technology and multi-modal data analysis methods, and using computers to quickly and accurately analyze physician-related data. Compared with the traditional evaluation method relying on expert experience and simple indicators, the present application can automatically process and analyze massive multi-dimensional data, including operation records, training experience, complication rate, patient feedback, etc., realizing comprehensive and objective evaluation of physician ability. The application of computer algorithms eliminates human bias, ensures the fairness and consistency of the evaluation process, and greatly improves the data processing efficiency, so that the evaluation work that originally takes several days or even weeks can be completed in a short time. Especially in risk prediction, the present application uses machine learning models to identify subtle patterns and correlations that humans cannot detect, and integrates multi-dimensional information such as physician ability, current state and operation characteristics, etc., to realize high-precision prediction of operation risk. This data-driven risk prediction capability enables hospital managers to take preventive measures before problems occur, from passive response to proactive prevention. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 FIG. 1 is a flowchart of a physician operation dynamic authorization method based on AI intelligent analysis according to an embodiment of the present application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be clearly described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0046] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are usually a class, not limited to the number of objects, for example, the first object can be one or more. In addition, "and / or" in the specification and claims means at least one of the connected objects, and the character " / ", generally represents a "or" relationship between the objects before and after.

[0047] Figure 1 FIG. 1 is a flowchart of a physician operation dynamic authorization method based on AI intelligent analysis according to an embodiment of the present application. It should be understood that although Figure 1The steps in the flowchart are displayed in sequence according to the arrows, but the steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated herein, the execution of the steps is not strictly limited in order, and the steps can be executed in other orders. Moreover, Figure 1 At least part of the steps in the flowchart can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of the sub-steps or stages is not necessarily sequential, but can be alternately executed with other steps or at least part of the sub-steps or stages of other steps. As shown in Figure 1 The physician surgery dynamic authorization method based on AI intelligent analysis disclosed in the present application specifically includes the following steps:

[0048] S101. Obtain multi-modal historical data related to the target physician.

[0049] Among them, the multi-modal historical data acquisition is the basic link of constructing the physician ability evaluation system, and a comprehensive physician portrait is formed by integrating data sources from different dimensions. In specific implementation, the basic information data of the physician is extracted from the hospital information management system, including educational background, professional qualification, working experience and other static characteristics; historical surgery data is collected from the surgery record system, covering surgery type, surgery duration, complication incidence, patient prognosis and other dynamic performance indicators; physician operation behavior data is obtained from the medical equipment monitoring system, such as surgery instrument use proficiency, operation precision and other skill performance parameters; continuing education records, skill assessment scores, peer review results and other ability evaluation data are extracted from the training and examination system. By establishing a unified data interface and standardized data format, these heterogeneous data sources are integrated and cleaned to form a comprehensive data set containing text, numerical, time series and other formats. This multi-dimensional data fusion method can fully reflect the professional ability, technical level and development track of the physician, provide rich and accurate data basis for subsequent ability modeling, and ensure the scientificity and reliability of the authorization decision.

[0050] S102. Based on the multi-modal historical data and through multi-modal feature fusion and deep learning representation method, a multi-dimensional ability portrait model of the target physician is constructed, and surgery permission is granted to the target physician according to the multi-dimensional ability portrait model.

[0051] The multi-dimensional capability portrait model converts multi-modal data into quantitative representations of physician capabilities through deep learning techniques, enabling precise assessment of physician professional levels. During implementation, first, feature extraction is performed on historical data from different modalities. Convolutional neural networks are used to process image-based data such as surgical video screenshots, recurrent neural networks are used to analyze time-series data such as surgical operation trajectories, and natural language processing techniques are used to analyze textual data such as medical records and evaluation feedback. Then, attention mechanisms are used to fuse multi-modal features. Next, a multi-layer perceptron network is constructed to map the fused features to capability scores in multiple dimensions such as technical ability, clinical experience, risk control, and team collaboration, with each dimension's score ranging from 0 to 100. The final capability portrait model can dynamically reflect a physician's competency level in different professional fields, providing quantitative basis for precise granting of surgical privileges and effectively avoiding the limitations of traditional subjective assessment.

[0052] S103. Obtain the target physician's recent workload data from the hospital information system, and model the target physician's current work state as a physician state model based on the recent workload data, and model the inherent risk of the target surgical task planned to be performed by the target physician as a surgical risk model.

[0053] The physician state model and the surgical risk model respectively quantify the current work state of the physician and the inherent risk of the surgical task, providing a real-time evaluation basis for dynamic authorization. Physician state modeling analyzes recent workload data, including continuous working time, surgery frequency, on-call arrangement, rest interval, and other factors, and uses a fatigue accumulation function to calculate a fatigue index. At the same time, combined with physiological monitoring data such as heart rate variability, sleep quality, and other indicators, a comprehensive state assessment model is constructed. Surgical risk modeling is based on factors such as surgery type, patient condition complexity, and surgery environment, and uses a risk assessment matrix to quantify the inherent risk of surgery. These two models can reflect the physician's professional state and the challenge level of the surgical task in real time, providing key input parameters for risk prediction.

[0054] S104. Integrate the multi-dimensional capability portrait model, the physician state model, and the surgical risk model, and use a machine learning-based model to build a risk prediction model to predict the target physician's predicted risk level in performing the target surgical task.

[0055] The risk prediction model uses a machine learning algorithm to integrate the outputs of three models: a multidimensional competency profile, physician status, and surgical risk. It predicts the overall risk level of a physician performing a specific surgical task. This model is constructed using an ensemble learning approach, combining multiple algorithms such as random forests, gradient boosting trees, and neural networks. The parameters represent the competency profile score, physician status index, and surgical risk score, respectively. Model training uses actual risk events from historical surgical data as labels, and cross-validation optimizes model parameters to ensure predictive accuracy. Predictions are categorized into three risk levels: low (0-30 points), medium (31-70 points), and high (71-100 points). The model also incorporates an uncertainty quantification mechanism, estimating prediction confidence using a Bayesian approach. When the confidence falls below a threshold, a manual review process is triggered. This predictive model provides accurate risk assessments before surgery, helping medical management departments make informed authorization decisions and effectively reducing the probability of medical risk events.

[0056] S105. Trigger a multi-level warning mechanism based on the predicted risk level and daily behavior score data in the multimodal historical data to automatically generate a dynamic adjustment strategy for the target physician's surgical authority based on a preset intervention strategy rule library.

[0057] Among them, the multi-level warning mechanism automatically triggers corresponding intervention measures and generates personalized authority adjustment strategies based on the risk prediction level and the physician's daily behavior score data. The system sets a three-level warning threshold: when the predicted risk is low and the behavior score is good, the existing authorization status is maintained; when the risk level is medium or the behavior score is abnormal, a second-level warning is triggered, requiring an additional preoperative assessment link or arranging guidance from a senior physician; when the prediction is high risk or the behavior score is seriously low, a first-level warning is activated, suspending the relevant surgical authority and requiring mandatory training. The preset intervention strategy rule library contains a variety of response plans, such as skills training, psychological counseling, workload adjustment, team configuration optimization, etc. Strategy generation uses a decision tree algorithm to automatically match the most appropriate intervention measures according to the specific risk factor combination.

[0058] In one embodiment, obtaining multimodal historical data related to a target physician includes the following steps:

[0059] Obtain the target physician's structured archival data through a standardized data interface;

[0060] Collecting the target physician's surgical simulation history data from the surgical simulator, the surgical simulation history data including the target physician's historical surgical simulation operation data, physiological index data, and eye movement pattern data;

[0061] The structured archive data and the surgery simulator historical data are data cleaned and data integrated to form a multi-modal historical data set of the target physician.

[0062] In the present embodiment, the standardized data interface automatically extracts physician archive information from the hospital information system through a unified data exchange protocol, ensuring the integrity and consistency of data acquisition. The implementation principle is based on the RESTful API architecture, establishing a standardized data request format and response mode, and transmitting data through JSON or XML format. In specific implementation, first, a unified data field mapping table is defined to standardize the physician information fields in different hospital systems into a unified format, such as unifying different expressions such as "work years", "employment time", "practice years" into the "practice_years" field. Then, a data verification mechanism is established to check the integrity of the acquired data, and the verification formula is , wherein is the number of filled fields, is the total number of mandatory fields. The data interface also integrates real-time synchronization function, automatically updates the local database when the physician archive information changes. The structured archive data acquired includes basic information (name, age, gender), education background (education, graduation school, professional direction), professional qualification (practice certificate, specialist certification, technical title), work experience (department distribution, job change, performance appraisal) and other dimensions. This standardized acquisition method eliminates the problem of data silos and provides high-quality basic data support for subsequent analysis.

[0063] The surgery simulator data acquisition records the physician's operation performance and physiological response in the virtual surgery environment through multi-sensor fusion technology, and builds the objective evaluation basis of the physician's skill level. The implementation principle adopts a distributed data acquisition architecture, with force feedback sensors, position trackers, pressure sensors and other devices built-in the surgery simulator to monitor the physician's operation behavior in real time, while equipped with physiological monitoring devices and eye tracking instruments to record the physician's physiological state and attention distribution. The operation data acquisition includes the movement trajectory of surgical instruments, operation force, action accuracy and other parameters, which are converted into standardized space coordinates through the coordinate transformation formula , wherein R is the rotation matrix and T is the translation vector. The physiological index data covers heart rate, blood pressure, skin conductivity, electromyographic signals and other parameters, and uses a sliding window algorithm to calculate the trend of physiological indicators. The eye movement pattern data captures the pupil movement trajectory through a high-speed camera, records the gaze point distribution, saccade path, gaze duration and other parameters. The data acquisition frequency is set to 1000Hz to ensure that the subtle changes in the physician's operation are captured.

[0064] The data cleaning and integration process converts heterogeneous data sources into a unified format multi-modal dataset through advanced data preprocessing techniques, laying a solid foundation for subsequent analysis. The implementation principle adopts a multi-stage data processing pipeline. First, data quality assessment is performed to identify missing values, outliers, and duplicate records. Missing value processing uses multiple imputation methods. According to the data type, select the appropriate imputation strategy. For numerical data, use mean imputation or regression imputation. For categorical data, use mode imputation. Outlier detection uses the boxplot method and Z-score standardization. The detection formula is where μ is the mean and σ is the standard deviation. When |Z| > 3, it is determined to be an outlier. In the data integration stage, a unified timestamp alignment mechanism is established to synchronize data of different collection frequencies to a unified time axis. Linear interpolation method is used to fill in the time gap. In the feature engineering stage, key feature indicators such as operation stability index, physiological stress intensity, visual attention concentration degree, etc. are extracted. The final multi-modal dataset includes structured table data, time series data, image data, etc.

[0065] In one embodiment, based on multi-modal historical data and through multi-modal feature fusion and deep learning representation method, a multi-dimensional capability portrait model of the target physician is constructed, including the following steps:

[0066] Feature extraction and independent coding are performed on different modal data in the multi-modal historical data;

[0067] The feature vectors obtained after independent coding of each modal data are fused using a preset fusion strategy to generate unified multi-modal fusion features;

[0068] The multi-modal fusion features are input into a deep neural network, and an evaluation model that outputs the target physician's quantitative score on the pre-defined capability dimension is trained through supervised learning, forming a multi-dimensional capability portrait model;

[0069] Model explainability techniques are applied to analyze key input features that affect the evaluation results of the multi-dimensional capability portrait model;

[0070] The dynamic changes of the key input features are monitored to obtain a dynamically updated multi-dimensional capability portrait model.

[0071] In this embodiment, feature extraction and independent coding of multi-modal data convert different types of raw data into computer-processable numerical feature vectors through targeted algorithms, laying a foundation for subsequent fusion analysis. The implementation principle is based on deep learning representation learning technology. Special encoder architectures are used for different modal data. For structured archive data, a multi-layer perceptron is used for feature encoding. Categorical variables are converted to vector form through one-hot encoding, and numerical variables are standardized and input into a fully connected layer. The encoding formula is where σ is an activation function, W and b are weight matrix and bias vector respectively. For the surgical operation time series data, long short-term memory network (LSTM) is used to extract time-dependent features to capture the dynamic change patterns of operation skills. For the physiological indicator data, one-dimensional convolutional neural network is used to extract local feature patterns to identify the characteristic waveforms of physiological responses. For the eye movement trajectory data, the spatial relationship between fixation points is modeled by graph convolution network. The output dimension of the encoder for each modality is uniformly set to 256 dimensions to ensure the consistency of feature expression. This independent encoding mode for each modality fully utilizes the intrinsic characteristics of each type of data, and the extracted feature vectors can effectively retain the key information in the original data, providing high-quality feature representation for multi-modal fusion.

[0072] Multi-modal feature fusion integrates independent feature vectors of different modalities into a unified fusion representation through attention mechanism and feature interaction strategy, realizing effective integration and complementation of information. The implementation principle adopts a hierarchical fusion architecture, which first performs early fusion at the feature level and then performs late fusion at the decision level. In the early fusion stage, a multi-head attention mechanism is used to calculate the correlation weight between different modalities. Through cross-attention mechanism, different modalities pay attention to each other, enhancing the semantic association between features. Then, a gated fusion unit is used to dynamically adjust the contribution weight of each modality feature, and the fusion formula is where is the gating weight of the i-th modality, is the corresponding feature vector, and represents element-wise multiplication. The gating weight is learned and can adaptively adjust the importance of each modality according to the specific task. This fusion strategy effectively solves the problem of information imbalance and feature redundancy between modalities, and the generated fusion features have stronger expression ability and generalization performance.

[0073] The multi-dimensional capability portrait model maps the multi-modal fusion features to multiple evaluation dimensions of physician capability through a deep neural network, realizing quantitative evaluation of physician professional level. The implementation principle adopts a multi-task learning framework, and the network structure includes a shared feature extraction layer and multiple task-specific output branches. The shared layer uses a deep network with residual connection to extract general feature representation, and the calculation formula is where is the hidden state of the l-th layer. The task-specific branch designs an output layer for each of the technical skill dimension, cognitive skill dimension, non-technical skill dimension, and learning adaptability dimension in the pre-defined capability dimension, and the score range of each dimension is 0-100 points. The training process uses supervised learning, using historical evaluation data as labels, and the loss function combines regression loss and regularization term, with the formula , where m is the number of samples, k is the number of ability dimensions, and λ is the regularization coefficient. The model was trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. The trained model accurately reflects the physician's true level across all ability dimensions, providing a reliable quantitative basis for surgical authorization.

[0074] Model interpretability analysis uses feature importance assessment and attention visualization technology to identify key factors that affect physician competence assessment, thereby improving the transparency and credibility of model decisions. The implementation principle combines SHAP (SHapley Additive exPlanations) value analysis and gradient activation mapping technology to quantify the contribution of each input feature to the final evaluation result. SHAP value calculation is based on game theory principles and evaluates feature importance by calculating the marginal contribution of a feature in all possible feature subsets. The calculation formula is: , where F is the full feature set, S is the feature subset, and f is the model prediction function. An integrated gradient method is also used to analyze the sensitivity of features to prediction results. Feature contributions are determined by calculating the integral path of the input feature gradient. Visual analysis generates feature importance ranking graphs and heat maps, visually demonstrating which features in each modality's data play a key role in capability assessment.

[0075] The dynamic update mechanism continuously monitors the changing trends of key input features and adjusts model parameters in real time to maintain the accuracy and timeliness of the evaluation results. The implementation principle is based on online learning and concept drift detection technology to establish a feature change monitoring system and a model adaptive update mechanism. Feature monitoring uses a sliding window statistical method to calculate the statistical characteristic changes of key features within the time window. The change detection formula is , where μ and σ are the new and old values ​​of the mean and standard deviation respectively. When the degree of change exceeds the preset threshold, the model update process is triggered. The model update adopts an incremental learning strategy, using the newly acquired data to fine-tune the existing model. The update formula is , where α is the learning rate, The gradient of new data. To prevent catastrophic forgetting, elastic weight consolidation technology is used to protect important parameters from excessive modification. The update frequency is adaptively adjusted based on the degree of data change, generally set to weekly or monthly. This continuous learning mechanism ensures that the capability profile model always keeps pace with the physician's actual capability level, providing a reliable real-time assessment basis for dynamic authorization decisions.

[0076] In one embodiment, obtaining recent workload data of a target physician from a hospital information system, quantitatively modeling the target physician's current work status into a physician status model based on the recent workload data, and quantitatively modeling the inherent risk of a target surgical task planned to be performed by the target physician into a surgical risk model includes the following steps:

[0077] obtaining recent workload data of the target physician from a hospital information system, and estimating a fatigue index of the target physician according to the recent workload data to form a physician state model quantified from a current working state of the target physician;

[0078] evaluating inherent complexity of the surgery and patient-specific risk factors based on a surgery type, an estimated duration, and a patient basic condition of a target surgery task planned to be performed by the target physician;

[0079] combining the inherent complexity of the surgery and the patient-specific risk factors, and constructing a surgery risk model of the target physician performing the target surgery task by using a risk scoring model;

[0080] standardizing the quantified physician state model and the surgery risk model.

[0081] In the embodiment, the physician state model evaluates the current physical and mental state of the physician by quantitatively analyzing the workload data of the physician, and provides an objective state evaluation basis for the surgery authorization. The implementation principle is based on the fatigue accumulation theory and the biological rhythm model. The recent work record data of the physician is automatically extracted from the hospital information system, including daily working hours, consecutive working days, night shift frequency, surgery table, on-duty arrangement and other key indicators. The fatigue index calculation adopts a multi-factor weighted model, and the formula is wherein is a work intensity fatigue factor, which is calculated by the consecutive working hours and the work density; is a sleep deficiency fatigue factor, which is evaluated based on the on-duty frequency and the rest interval; is a psychological stress fatigue factor, which is determined in combination with the surgery complexity and the work load fluctuation; is a recovery ability factor, which considers the age, constitution and historical recovery mode of the physician. The weight coefficients are determined by historical data training, and the typical values are w1=0.4, w2=0.3, w3=0.2, w4=0.1. The model also integrates a biological rhythm adjustment mechanism to adjust the fatigue weight in different time periods according to the theory of circadian rhythm. The final output of the physician state score ranges from 0 to 100 points, and the higher the score, the better the state, which can accurately reflect the real-time working state and professional ability of the physician.

[0082] The surgery complexity and patient risk assessment determine the inherent difficulty of the surgery task and the special risk of the patient individual by multi-dimensional analysis, which provides basic data for subsequent risk modeling. The implementation principle adopts a hierarchical evaluation framework. Firstly, the inherent complexity of the surgery is quantitatively evaluated, considering factors such as technical difficulty of the surgery type, complexity of anatomical structure, and operation precision requirement. The surgery complexity scoring formula is wherein is a surgery type difficulty coefficient, which is determined according to the surgery classification standard; To score the anatomical complexity, the surgical site and the number of organs involved are considered; To score the fine motor skills requirement, the precision of the operation is reflected. The patient-specific risk assessment considers the patient's age, underlying diseases, previous surgery history, drug allergy history, nutritional status, etc. The improved ASA classification system and Charlson comorbidity index are used for quantification. Through standardized assessment processes, the risk assessment of different types of surgery and patients is comparable, laying the foundation for accurate risk prediction.

[0083] The surgical risk model integrates the inherent complexity of surgery and patient-specific risk factors to predict the risk level that may occur during surgery using machine learning algorithms. The implementation principle is based on the ensemble learning method, combining multiple algorithms such as logistic regression, random forest, and gradient boosting tree to build a risk scoring model. The input features of the model include the aforementioned surgical complexity score, patient risk factors, and surgical environment factors such as equipment status, team configuration, etc. The risk score calculation formula is , where is a continuous risk variable, is a categorical risk indicator, and β and γ are the corresponding regression coefficients. The model training uses historical surgery data, with intraoperative complication rates, surgery time-out rates, and patient adverse event rates as target variables. Through cross-validation and grid search optimization, the model parameters are optimized, and the final model's AUC value reaches 0.87, with an accuracy rate of over 82%. The risk score results are divided into low risk (0-0.3), medium risk (0.3-0.7), and high risk (0.7-1.0) according to the probability distribution. The model also provides risk factor contribution analysis to help identify the main sources of risk and provide guidance for the development of risk control measures. This quantitative risk assessment method significantly improves the accuracy and objectivity of surgical risk prediction.

[0084] Standardization processing converts the physician state score and surgical risk score into standardized values with the same dimension and distribution range, ensuring the effectiveness and accuracy of subsequent model fusion. The implementation principle uses a combination strategy of multiple standardization techniques. First, Z-score standardization is used to eliminate the influence of different scoring dimensions. Then, Min-Max normalization is used to map the data to the [0, 1] interval, with the normalization formula ensuring that the outputs of different models have the same numerical range. To handle the skewness of data distribution, Box-Cox transformation is also applied to optimize the data distribution characteristics, with the transformation formula where λ is a transformation parameter, and the optimal value is determined by maximum likelihood estimation. The standardization process also includes outlier detection and processing, using the interquartile range method to identify outliers, and truncation or smoothing of values outside the normal range. The final standardized physician state score and surgical risk score both conform to the standard normal distribution, with a mean of 0 and a standard deviation of 1, and data quality assessment shows that the standardization effect is good, with a coefficient of variation controlled within 15%. This standardization ensures the comparability and compatibility of data from different sources, providing high-quality input data for subsequent risk prediction models.

[0085] In one embodiment, a risk prediction model is constructed based on a machine learning model, and the predicted risk level of the target physician in performing the target surgical task is obtained by the following steps:

[0086] The risk prediction model is constructed based on a machine learning model, and the input features of the risk prediction model at least include the multi-dimensional capability portrait model, the physician state model and the surgical risk model;

[0087] The risk prediction model is trained and verified using historical surgical case data;

[0088] The performance of the trained risk prediction model is evaluated and calibrated;

[0089] The risk prediction model is applied to output the predicted risk level of the target physician in performing the target surgical task.

[0090] In this embodiment, the risk prediction model integrates various machine learning algorithms to fuse multi-dimensional information such as physician capability, state and surgical risk into a unified risk assessment framework, achieving accurate prediction of surgical risk. The implementation principle adopts a multi-level ensemble learning architecture, combining the advantages of gradient boosting decision tree (GBDT), random forest (RF) and deep neural network (DNN) algorithms to construct a heterogeneous ensemble model. The input features of the model include the technical proficiency, clinical judgment and risk control capability scores output by the multi-dimensional capability portrait model, the fatigue index and workload score provided by the physician state model, and the surgical complexity and patient risk factors calculated by the surgical risk model, totaling more than 150 feature dimensions. Principal component analysis (PCA) and recursive feature elimination (RFE) methods are used for feature selection in the feature engineering stage, and the dimension reduction formula is where X is the original feature matrix, W is the principal component weight matrix, and Y is the feature representation after dimension reduction. The construction of the basic learner adopts a hierarchical training strategy, with the GBDT model processing numerical features, the random forest processing categorical features, and the deep neural network capturing the nonlinear interaction between features. The ensemble strategy adopts a weighted average method, and the weights are determined through cross-validation. The ensemble formula is where the sum of each weight is 1. The model also integrates an uncertainty quantification mechanism, which estimates the prediction confidence through Bayesian methods, triggering a manual review process when the confidence is below a threshold. To handle the class imbalance problem, the SMOTE oversampling technique and cost-sensitive learning method are used to ensure the model's sensitivity to high-risk cases. The final risk prediction model can output a continuous risk probability value between 0 and 1, and is divided into low, medium, and high risk levels according to the preset threshold, providing quantitative risk assessment support for clinical decision-making.

[0091] The model training and validation process optimizes model parameters and verifies prediction performance through deep mining of large-scale historical surgery data, ensuring the accuracy and generalization ability of the model. The implementation principle is based on a supervised learning framework, collecting surgery case data from the past 5 years as the training set, including more than 50,000 surgery records covering different departments and different complexity of surgery types. In the data preprocessing stage, multiple imputation methods are used to handle missing values, and features with a missing rate of more than 30% are excluded. Outliers are identified and smoothed through boxplot methods. The label construction uses a composite index system, combining intraoperative complication occurrence, surgery time overrun, patient adverse events, postoperative infection, and other risk events into a binary classification label, with a risk event occurrence rate of about 15%. The training strategy uses stratified sampling 5-fold cross-validation to ensure the consistency of the positive and negative sample proportions in the training and validation sets. The early stopping mechanism is used in model training to prevent overfitting, and the AUC index on the validation set is monitored, and training is stopped when there is no improvement for 10 consecutive epochs. The Bayesian optimization method is used for hyperparameter optimization, and the search space includes learning rate (0.01-0.1), tree depth (3-10), regularization parameter (0.01-1.0), and other key parameters. Learning rate decay strategy is used during training, with an initial learning rate of 0.05, and a decay of 0.9 times every 50 epochs. To improve the robustness of the model, data augmentation techniques are used to generate perturbed samples, and training data is expanded by adding Gaussian noise and feature replacement.

[0092] Model performance evaluation and calibration ensure the accuracy and reliability of the prediction results through a multi-dimensional index system and calibration techniques, improving the model's performance in actual applications. The implementation principle uses a comprehensive evaluation framework, combining classification performance indicators, calibration performance indicators, and clinical practicality indicators for comprehensive evaluation. Classification performance evaluation uses ROC curve, precision-recall curve, and confusion matrix, etc. to calculate sensitivity, specificity, positive predictive value, negative predictive value, and other key indicators. Calibration performance evaluation uses Hosmer-Lemeshow test and calibration chart analysis to evaluate the consistency of predicted probability and actual incidence, and the calibration statistic calculation formula is where g is the number of groups, and are the observed and expected values ​​of group i, respectively. The model calibration uses two methods: Platt scaling and isothermal regression. Platt scaling adjusts the predicted probability through the sigmoid function. The formula is: , where A and B are calibration parameters and f(x) is the original model output. Isothermal regression uses the rank-preserving regression method to establish a monotonic mapping relationship to ensure that the calibrated probability has a good ranking. The clinical practicality evaluation uses the decision curve analysis (DCA) method to calculate the net benefit under different risk thresholds. The formula is , where TP and FP are the number of true positives and false positives respectively, The evaluation results showed that the calibrated model's predicted probabilities across all risk intervals were highly consistent with the actual incidence rates, with calibration slopes close to 1, intercepts close to 0, and a Brier score of 0.12, indicating excellent calibration performance. Decision curve analysis revealed that the model had a positive net benefit across risk thresholds of 0.1-0.8, demonstrating its practical value in clinical decision-making.

[0093] The practical application of the risk prediction model provides clinicians with personalized surgical risk assessments and authorization recommendations through real-time data input and intelligent decision support. This implementation is based on a real-time inference engine. When a physician applies for surgical authorization, the model automatically retrieves the latest feature data from various data sources, including the physician's current competency profile score, real-time work status parameters, and the risk characteristics of the target surgery. The data preprocessing module transforms the input features according to the standardized training process to ensure that the data format and distribution are consistent with the training data. Model inference utilizes a combination of batch and stream processing. For routine authorization requests, batch processing is used, achieving response times within 5 seconds; for emergency surgery requests, stream processing is enabled, achieving millisecond-level responses. Risk classification utilizes a dynamic threshold strategy, adjusting the classification threshold based on the hospital's risk tolerance and quality control requirements, with a low-risk threshold set at 0.2 and a high-risk threshold set at 0.6. Output includes risk probability values, risk level classifications, analysis of the contribution of major risk factors, and personalized recommendations. For example, a predicted risk probability of 0.35 is classified as medium risk. Key risk factors may include elevated physician fatigue (30% contribution), greater surgical complexity (25% contribution), and a high number of patient comorbidities (20% contribution). Personalized recommendations may include increasing preoperative preparation time, arranging for assistance from experienced physicians, and strengthening intraoperative monitoring. The model also provides uncertainty assessment, prompting for manual review when the prediction confidence falls below 80%. Practical application results show that the model's prediction accuracy exceeds 85%, with a false positive rate below 10%. This significantly improves the scientific nature and safety of surgical authorization decisions and effectively reduces the incidence of medical risk events.

[0094] In one embodiment, the multi-level early warning mechanism is triggered according to the predicted risk level and the daily behavior score data in the multi-modal historical data, and the dynamic adjustment strategy of the target physician's operation permission is automatically generated according to the preset intervention strategy rule base, which includes the following steps:

[0095] Setting a threshold based on the predicted probability or risk level of adverse events and the triggering condition of the daily behavior score data in the multi-modal historical data to activate different levels of risk early warning;

[0096] Establishing a multi-level early warning mechanism including at least information prompt level, suggestion guidance level and warning review level;

[0097] According to the preset intervention strategy rule base and the risk warning in the early warning level in the multi-level early warning mechanism, the dynamic adjustment strategy for the target physician's operation permission is matched from the preset intervention measures.

[0098] In this embodiment, the risk early warning trigger mechanism realizes accurate identification and graded early warning of physician operation risk by setting a multi-dimensional threshold system, ensuring the forward-looking and effectiveness of medical safety management. The implementation principle is based on a multi-factor decision tree model, which considers the adverse event probability output by the risk prediction model, the daily behavior score data of the physician and the dynamic changes of environmental factors. The risk level threshold is set by statistical methods combined with clinical expert experience, and the optimal segmentation point is determined by analyzing the distribution characteristics of risk events in the historical data. In specific implementation, the risk probability is divided into three intervals: the low risk interval (0-0.3) corresponds to green early warning, the medium risk interval (0.3-0.7) corresponds to yellow early warning, and the high risk interval (0.7-1.0) corresponds to red early warning. The daily behavior score data includes the physician's attendance rate, training participation, colleague evaluation, patient satisfaction, medical error record and other dimensions, and the comprehensive behavior score is calculated by weighted average, the formula is wherein is the score of the i-th behavior dimension, is the corresponding weight. The triggering condition uses a logical combination rule, and the yellow early warning is triggered when the risk probability exceeds 0.3 or the behavior score is less than 70, and the red early warning is triggered when the risk probability exceeds 0.7 or the behavior score is less than 50. To avoid frequent false alarms, a time window mechanism is set, and only when the triggering condition is met for 3 consecutive days, the early warning is activated. In addition, a dynamic threshold adjustment mechanism is established to regularly calibrate the threshold parameters according to seasonal changes, department characteristics and overall risk level of the hospital. For example, the early warning threshold is appropriately lowered during the operation peak period, and the tolerance is appropriately increased after the holiday.

[0099] The multi-level warning mechanism realizes differentiated response and precise management for different risk levels through a hierarchical and progressive warning system, ensuring the systematicness and pertinence of medical safety management. The implementation principle is based on risk management theory and human factors engineering principles, and a three-level warning response system is constructed, with each level corresponding to different risk levels and intervention intensity. The information prompt level warning corresponds to low risk situations, mainly through information means to send reminder information to physicians and management departments, and the warning trigger condition is that the risk probability is between 0.2-0.3 or the behavior score is between 70-80 points. In specific implementation, through the hospital information system, emails, short messages or APP push are automatically sent, and the reminder content includes the current risk state, main risk factors and basic precautions. The suggestion guidance level warning corresponds to medium risk situations, in addition to information reminders, it also needs active guidance intervention, and the trigger condition is that the risk probability is between 0.3-0.7 or the behavior score is between 50-70 points. This level of warning will automatically arrange the department director or senior physician for face-to-face guidance, provide individualized improvement suggestions, and require the physician to participate in relevant training or skill enhancement activities. The warning audit level warning corresponds to high risk situations, which needs immediate audit and mandatory intervention measures, and the trigger condition is that the risk probability is more than 0.7 or the behavior score is less than 50 points. This level of warning will immediately notify the medical department and the department head, and start the emergency assessment procedure, which may include measures such as suspension of surgery permission, mandatory rest, psychological assessment, etc. Each warning level has a clear response time limit, the information prompt level requires response within 24 hours, the suggestion guidance level requires response within 8 hours, and the warning audit level requires response within 2 hours. The warning mechanism also integrates escalation and de-escalation mechanisms, which automatically escalate the warning level when the risk continues to worsen, and gradually de-escalate when the risk is effectively controlled.

[0100] The intervention strategy matching mechanism automatically generates individualized surgery permission adjustment strategies according to the warning level and specific risk characteristics through an intelligent rule engine, achieving precise intervention and dynamic management. The implementation principle is based on expert systems and case-based reasoning technology, and a strategy knowledge base containing hundreds of intervention rules is established, each rule clearly defines the trigger condition, intervention measures and expected effect. The construction of the rule base adopts a hierarchical structure, with the top layer being the warning level classification, the middle layer being the risk factor category, and the bottom layer being the specific intervention measures. For example, for the information prompt level warning, the rule base contains "skill reminder", "training recommendation", "workload suggestion" and other mild intervention measures; for the suggestion guidance level warning, it contains "tutor guidance", "skill assessment", "work adjustment" and other moderate intervention measures; for the warning audit level warning, it contains "permission suspension", "mandatory training", "psychological assessment" and other intensive intervention measures. The strategy matching adopts multi-dimensional similarity calculation, and the matching formula is where is the jth feature of the current situation, is the corresponding feature in the rule, where w is the feature weight, and sim is the similarity function. The strategy generation process also considers the physician's personal characteristics, historical intervention effects, and current working environment, and continuously optimizes the matching accuracy through machine learning algorithms. In specific implementation, when a yellow warning is triggered and the main risk factor is fatigue overload, the possible matching strategies include: adjusting the working shift, arranging mandatory rest, providing fatigue management training, etc. When a red warning is triggered and the risk factor is skill deficiency, the possible strategies include: suspending the permission of complex surgery, arranging skill enhancement training, assigning experienced physicians for guidance, etc. The strategy execution adopts a closed-loop management mode, including four links of strategy formulation, execution supervision, effect evaluation and feedback adjustment. The execution effect is evaluated through quantitative indicators such as the decline amplitude of risk level, the improvement degree of behavior score, the incidence of subsequent events, etc.

[0101] In one embodiment, the method further comprises the steps of:

[0102] establishing an automated data collection process for clinical outcome data and user feedback data;

[0103] continuously monitoring the performance of the multi-dimensional capability portrait model and the risk prediction model on real-world data using the data collected by the data collection process;

[0104] when the model performance of the multi-dimensional capability portrait model or the risk prediction model is below a preset threshold, adjusting the multi-dimensional capability portrait model and the risk prediction model using the accumulated clinical outcome data and user feedback data.

[0105] In this embodiment, the automated data collection process ensures that high-quality feedback data required for continuous learning of the model can be obtained and processed in a timely and accurate manner through the construction of a real-time collection and integration mechanism for multi-source heterogeneous data. The implementation principle is based on event-driven architecture and streaming data processing technology, and a data monitoring network covering the entire surgical process is established. The clinical outcome data collection uses a distributed sensor network, with data collection devices deployed at key nodes such as the operating room, ward, and laboratory to capture objective data such as surgery duration, complication occurrence, patient vital sign changes, and postoperative recovery indicators in real time. The data collection interface uses the HL7 FHIR standard to ensure interoperability between different medical devices and information systems, and the data transmission uses a message queue mechanism to ensure reliable data transmission and real-time performance. In specific implementation, structured data such as surgical records, anesthesia records, and nursing records are automatically extracted from the electronic medical record system through the API interface, preoperative and postoperative contrast images are obtained from the medical imaging system, and laboratory test data such as blood tests and biochemical indicators are collected from the laboratory information system. User feedback data collection is achieved through multi-channel interactive interfaces, including the physician APP, patient satisfaction survey system, peer review platform, etc. Physician feedback uses a combination of structured questionnaires and open-ended evaluations, with evaluation dimensions including perceived surgical difficulty, team collaboration evaluation, equipment satisfaction, and time arrangement rationality, etc., and the score uses a Likert 5-point scale. Patient feedback is automatically sent through the postoperative follow-up system, covering pain level, recovery progress, and medical service satisfaction. Data quality control uses a multi-layer verification mechanism, including data integrity check, logical consistency verification, and outlier identification, etc., and the quality score formula is where each C value represents the integrity, consistency, and accuracy score, respectively. Data storage uses a hybrid architecture of time-series database and relational database to support high-concurrency writing and complex query requirements.

[0106] The model performance monitoring mechanism continuously tracks the performance of multi-dimensional capability portrait models and risk prediction models in real clinical environments, and timely detects performance degradation and concept drift problems by establishing a comprehensive evaluation index system and real-time monitoring platform. The implementation principle is based on statistical process control theory and machine learning model monitoring technology, and a multi-level performance evaluation framework is constructed. The monitoring index system includes three levels of prediction accuracy indicators, model stability indicators, and business impact indicators. The prediction accuracy monitoring uses a sliding window method to calculate the AUC, precision, recall, F1 score, and other key indicators of the model within the recent time window, and compares them with historical benchmarks. The accuracy change detection formula is Performance warning is triggered when the variation exceeds 5%. Model stability monitoring identifies concept drift by analyzing the changes in the predictive distribution, using statistical methods such as the Kolmogorov-Smirnov test and the Population Stability Index (PSI), whose formula is When the PSI value exceeds 0.25, it indicates that the model stability has significantly decreased. Business impact monitoring focuses on the impact of model predictions on actual clinical decision-making, including changes in surgical success rates, medical malpractice rates, and physician satisfaction. The monitoring platform uses real-time dashboards to display dynamic changes in various indicators, supporting multi-dimensional drilling analysis and abnormal alarm functions. Performance degradation detection uses control chart methods, setting upper and lower control limits, and triggering investigations when consecutive 7 data points exceed the control limits or show a clear trend. The monitoring frequency is set according to the data volume and business importance, with key indicators monitored daily and general indicators monitored weekly. Monitoring results are sent to relevant technical teams and clinical experts through an automated reporting system on a regular basis to ensure that problems can be discovered and addressed in a timely manner.

[0107] The model adjustment and optimization mechanism uses intelligent retraining and incremental learning strategies to fine-tune the model using continuously accumulated clinical data and user feedback, ensuring that the model performance remains at its best. The implementation principle is based on online learning and transfer learning theories, combined with model distillation and knowledge preservation techniques, to build an adaptive model update framework. The model adjustment trigger mechanism uses multi-condition judgment, automatically starting the adjustment process when the monitoring indicators show that the model performance has decreased by more than the pre-set threshold (such as a 5% decrease in AUC, a PSI exceeding 0.25, or user satisfaction below 80%) or the amount of new data accumulated reaches a certain size. In the data preparation stage, the newly collected clinical outcome data and user feedback data are cleaned and labeled, and semi-supervised learning methods are used to handle some unlabeled data, generating pseudo-labels using model prediction results and expert knowledge. The model adjustment strategy uses different methods according to the reasons for performance degradation: for performance degradation caused by data distribution drift, domain adaptation techniques are used for feature space alignment; for prediction errors caused by new patterns, incremental learning methods are used to expand model capacity; for performance fluctuations caused by label noise, robust learning methods are used to improve model anti-interference ability.

[0108] In specific implementation, the adjustment of the multi-dimensional ability portrait model uses the knowledge distillation method to maintain the consistency of the original ability evaluation while incorporating new evaluation dimensions. The distillation loss function is , is the task loss, and For knowledge distillation loss, alpha is a balance parameter. The adjustment of the risk prediction model adopts an ensemble learning method, and the newly trained sub-model is combined with the original model by weighting. The weight is determined by the performance of the verification set. The model update adopts an A / B test framework. The new model is first tested in a small-scale user group, and then the improvement effect is verified by comparison experiments before being deployed comprehensively. In order to prevent catastrophic forgetting, the elastic weight consolidation (EWC) technology is used to protect important parameters from being modified too much. The regularization term is wherein is the Fisher information matrix diagonal element, is the important parameter value. The adjusted model needs to pass through strict verification tests, including offline evaluation, online A / B test and expert review. Only when all indicators meet the requirements can it be formally put online.

[0109] The application further discloses a doctor surgery dynamic authorization system based on AI intelligent analysis, which comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor.

[0110] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), ready programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The application does not limit this.

[0111] The memory can be an internal storage unit of the computer device, for example, a hard disk or a memory of the computer device, or an external storage device of the computer device, for example, a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD) or a flash memory card (FC) equipped on the computer device. The memory can also be a combination of the internal storage unit and the external storage device of the computer device. The memory is used to store the computer program and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or will be output. The application does not limit this.

[0112] The application further discloses a computer readable storage medium, which stores instructions. When the instructions are executed by a processor, the processor is configured to execute the doctor surgery dynamic authorization method based on AI intelligent analysis described in any one of the above embodiments.

[0113] The computer program can be stored in a machine readable medium, and the computer program includes computer program code. The computer program code can be in a form of source code, object code, executable file, or some intermediate form. The machine readable medium includes any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the machine readable medium includes but is not limited to the above-mentioned elements.

[0114] The computer program code is stored in the computer readable storage medium, and is loaded and executed on the processor, so as to facilitate the storage and application of the method.

[0115] It should be understood by those skilled in the art that the above discussion of any embodiment is only exemplary, and is not intended to limit the protection scope of the present application to these examples; under the idea of the present application, the above embodiments or technical features in different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of one or more embodiments of the present application as described above. In order to be brief, they are not provided in details.

[0116] One or more embodiments of the present application are intended to cover all such alternatives, modifications and variations falling within the broad scope of the present application. Therefore, any omission, modification, equivalent replacement, improvement, etc. made within the spirit and principles of one or more embodiments of the present application should be included in the protection scope of the present application.

Claims

1. A method for dynamic authorization of physician surgery based on AI intelligent analysis, characterized in that: The steps include: Acquire multimodal historical data related to the target physician; Based on multimodal historical data and through multimodal feature fusion and deep learning representation methods, a multidimensional capability profile model of the target physician is constructed, and surgical authorization is granted to the target physician based on the multidimensional capability profile model; Recent workload data of the target physician is obtained from the hospital information system, and the current work status of the target physician is quantitatively modeled into a physician status model based on the recent workload data. The inherent risk of the target surgical task planned to be performed by the target physician is quantitatively modeled into a surgical risk model. Integrate the multi-dimensional capability profile model, physician status model, and surgical risk model, and use a machine learning model to build a risk prediction model to predict the target physician's risk level when performing the target surgical task. A multi-level early warning mechanism is triggered based on the predicted risk level and daily behavior scoring data in multimodal historical data to automatically generate a dynamic adjustment strategy for the target physician's surgical authority based on the preset intervention strategy rule library.

2. The method for dynamic authorization of physician surgery based on AI intelligent analysis according to claim 1 is characterized in that: The step of obtaining multimodal historical data related to the target physician includes the following steps: Obtain target physicians' structured archival data through standardized data interfaces; Collecting the target physician's surgical simulation history data from the surgical simulator, the surgical simulation history data including the target physician's historical surgical simulation operation data, physiological index data, and eye movement pattern data; The structured archival data and surgical simulation historical data are cleaned and integrated to form a multimodal historical data set of the target physician.

3. The method for dynamic authorization of physician surgery based on AI intelligent analysis according to claim 2 is characterized in that: The method of constructing a multi-dimensional capability portrait model of a target physician based on multimodal historical data and through multimodal feature fusion and deep learning representation method includes the following steps: Perform feature extraction and independent encoding on different modal data in multimodal historical data; The preset fusion strategy is used to fuse the feature vectors obtained after independent encoding of each modality data to generate a unified multimodal fusion feature; The multimodal fusion features are input into a deep neural network, and trained through supervised learning to obtain an evaluation model that outputs quantitative scores of the target physician on predefined ability dimensions, forming a multidimensional ability profile model. Apply model interpretability techniques to analyze key input features that affect the evaluation results of the multi-dimensional capability profiling model; Monitor the dynamic changes of key input features to obtain a dynamically updated multi-dimensional capability portrait model.

4. The method for dynamic authorization of physician surgery based on AI intelligent analysis according to claim 3 is characterized in that: The predefined capability dimensions include at least technical skills dimension, cognitive skills dimension, non-technical skills dimension and learning adaptability dimension.

5. The method for dynamic authorization of physician surgery based on AI intelligent analysis according to claim 1 is characterized in that: The method of obtaining recent workload data of a target physician from a hospital information system, quantitatively modeling the current work status of the target physician into a physician status model based on the recent workload data, and quantitatively modeling the inherent risk of a target surgical task planned to be performed by the target physician into a surgical risk model includes the following steps: Obtain the target physician's recent workload data from the hospital information system, and estimate the target physician's fatigue index based on the recent workload data to form a physician status model that quantifies the target physician's current work status; Assess the inherent complexity of the surgery and patient-specific risk factors based on the type of surgery the target physician plans to perform, the estimated duration, and the patient's basic conditions; Combining the inherent complexity of surgery with patient-specific risk factors, and using a risk scoring model to construct a surgical risk model for the target physician to perform the target surgical task; The quantified physician status model and surgical risk model were standardized.

6. The method for dynamic authorization of physician surgery based on AI intelligent analysis according to claim 5 is characterized in that: The method of constructing a risk prediction model based on a machine learning model to predict the risk level of a target physician in currently performing a target surgical task comprises the following steps: Constructing a risk prediction model based on a machine learning model, wherein the input features of the risk prediction model include at least a multidimensional capability profile model, a physician status model, and a surgical risk model; Use historical surgical case data to train and validate the risk prediction model; Conduct performance evaluation and calibration of the trained risk prediction model; The risk prediction model is applied to output the predicted risk level of the target physician's current target surgical task.

7. The method for dynamic authorization of physician surgery based on AI intelligent analysis according to claim 1 is characterized in that: The multi-level warning mechanism is triggered based on the predicted risk level and the daily behavior score data in the multimodal historical data to automatically generate a dynamic adjustment strategy for the target physician's surgical authority based on a preset intervention strategy rule library, including the following steps: Setting thresholds based on the predicted probability or risk level of adverse events and trigger conditions based on daily behavior score data in multimodal historical data to activate risk warnings at different levels; Establish a multi-level early warning mechanism that includes at least information prompt level, suggestion guidance level, and warning review level; According to the preset intervention strategy rule base and the warning level of risk warning in the multi-level warning mechanism, a dynamic adjustment strategy for the surgical authority of the target physician is matched from the preset intervention measures.

8. The method for dynamic authorization of physician surgery based on AI intelligent analysis according to claim 1 is characterized in that: The method further comprises the steps of: Establish an automated data collection process for clinical outcome data and user feedback data; Utilize the data collected during the data collection process to continuously monitor the performance of the multi-dimensional capability profiling model and risk prediction model on real-world data; When the model performance of the multidimensional capability profiling model or the risk prediction model is lower than the preset threshold, the accumulated clinical outcome data and user feedback data are used to adjust the multidimensional capability profiling model and the risk prediction model.

9. A dynamic authorization system for physician surgery based on AI intelligent analysis, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for dynamic authorization of physician surgery based on AI intelligent analysis as described in any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instruction is executed by a processor, the processor is configured to execute the method for dynamic authorization of physician surgery based on AI intelligent analysis according to any one of claims 1 to 8.

Images