Urinary operation risk AI intelligent assessment method and system

Abstract

The invention relates to the technical field of intelligent medical information, in particular to a urinary surgery risk AI intelligent assessment method and system. The method comprises the steps that intraoperative data in the urinary surgery operation process are collected in real time; inputting the intraoperative data into a risk feature fusion AI model, performing context association analysis of different stages of the operation, and fusing each intraoperative multi-source modal feature to generate a risk assessment comprehensive feature set; analyzing the risk assessment comprehensive feature set by using a risk assessment prediction AI model, dynamically adjusting the priority and prediction weight of each risk index in combination with a doctor judgment priority weight adjustment mechanism, and predicting to obtain an intra-operative real-time comprehensive risk score; and visually displaying the intraoperative real-time comprehensive risk score and the risk prompt information in a terminal in real time. According to the AI intelligent assessment method, the causal relationship between the risks is constructed by associating the data of each stage of the operation with doctor judgment, and the AI intelligent assessment method capable of dynamically predicting the key risks in the operation process is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent medical information technology, in particular to a urinary surgery risk AI intelligent evaluation method and system. BACKGROUND

[0002] Under the background of the gradual popularization of minimally invasive surgery, complex open reconstruction surgery and other procedures in urology, intraoperative risk control has become a core factor affecting the incidence of postoperative complications, functional recovery level and medical safety quality. In recent years, with the improvement of the informatization and intelligentization level of the operating room, the acquisition and utilization ability of intraoperative multi-source data has been significantly enhanced, including intraoperative image sequences, physiological parameters, intraoperative voice recordings, etc. However, most of the current surgical risk assessment methods still remain at the postoperative summary or preoperative static modeling level, lacking dynamic assessment capability for real-time intraoperative data. Especially when unforeseen risk events occur during surgery, the traditional method relying on experience and intuition lacks systematic support, which restricts the initiative and accuracy of risk prevention and control.

[0003] In specific applications, the risk factors involved in different stages of surgery have significant differences in form, timing characteristics and data structure. However, the existing technology still has the following three key problems: first, during the operation process in different stages of surgery, the patient's physical indicators are constantly changing as the surgery progresses, and the surgical risk cannot be guaranteed, so preoperative risk assessment cannot replace intraoperative surgical risk assessment; second, most existing intraoperative risk assessment systems use intra-stage static feature modeling methods, which cannot identify the evolution of local high-risk features in the previous stage into major risk events in the subsequent stage, and the risk transmission chain is cut off between stages, which may lead to insufficient warning of high-risk stages; finally, the existing system usually uses parallel modeling strategy for multiple risk indicators, but there is a close relationship between intraoperative complication risk and functional risk, for example, in a bladder tumor resection surgery, if the unexpected bladder wall bleeding during the electrocision process is not controlled in time, it is easy to cause postoperative bladder volume reduction and damage to the detrusor muscle function, but the existing system does not consider bleeding events as a leading factor when assessing functional risk, which leads to the failure to accurately identify the risk linkage in the early stage. SUMMARY

[0004] The present application aims to provide a urinary surgery risk AI intelligent evaluation method and system to solve the three key problems mentioned in the background.

[0005] To achieve the above-mentioned purpose, the technical solution of the present application is: a urinary surgery risk AI intelligent evaluation method, comprising:

[0006] S1, collecting intraoperative data of a urological surgery process in real time, and performing time synchronization processing and preprocessing;

[0007] The intraoperative data includes intraoperative objective data and intraoperative subjective data.

[0008] S2, inputting the intraoperative data into a risk feature fusion AI model to extract intraoperative multi-source modal features, and performing context correlation analysis of different stages of the surgery, in the context correlation analysis process, performing subjective-objective consistency comparison analysis on the intraoperative multi-source modal features, obtaining a consistency comparison result, and fusing each intraoperative multi-source modal feature to generate a risk assessment comprehensive feature set;

[0009] S3, using a risk assessment prediction AI model to analyze the risk assessment comprehensive feature set, calculating and analyzing the risk causal influence relationship between the complication risk index and the functional risk index based on a multi-risk causal interaction graph and a causal graph embedding modeling mechanism, dynamically adjusting the priority and prediction weight of each risk index combined with a doctor judgment priority weight adjustment mechanism, and predicting to obtain an intraoperative real-time comprehensive risk score;

[0010] S4, real-time visualizing the intraoperative real-time comprehensive risk score and risk prompt information in the terminal.

[0011] Preferably, in S1, the intraoperative objective data includes intraoperative video data, physiological parameter data, intraoperative image data, intraoperative pathological detection data, operating room environment data, and surgical instrument state data; the intraoperative subjective data is real-time language judgment information of the leading doctor acquired by a voice acquisition device.

[0012] The time synchronization processing and preprocessing is an operation on the intraoperative objective data and the intraoperative subjective data, specifically as follows:

[0013] Based on a unified time reference, the intraoperative objective data and the intraoperative subjective data are time-stamped aligned, format standardized, noise removed, and missing value filled according to the data type, and the intraoperative subjective data is converted into structured labels representing the stage of surgery, the state of the patient, and the type of risk.

[0014] Preferably, in S2, the risk feature fusion AI model is constructed based on a combination of a multi-source data feature encoding network, a surgery stage context modeling network, and a consistency comparison feature fusion network, for feature extraction, context correlation analysis, and consistency comparison feature fusion of the intraoperative objective data and the intraoperative subjective data, to generate a risk assessment comprehensive feature set.

[0015] The construction method of the risk feature fusion AI model is as follows:

[0016] The feature encoding subnetwork is constructed to encode the intraoperative objective data and the intraoperative subjective data respectively to obtain the intraoperative multi-source modal features; the surgery stage context modeling subnetwork is constructed to model the time sequence correlation of the intraoperative multi-source modal feature sequences of different surgery stages, and to perform cross-stage mode migration during modeling to map the intraoperative multi-source modal features of the previous stage to the feature space of the current stage; the consistency comparison feature fusion subnetwork is constructed to introduce a constraint loss function based on the consistency comparison results of the subjective and objective during model training to adjust the fusion weight of the intraoperative multi-source modal features; the risk feature fusion AI model is trained based on the multi-source training samples, and the training is terminated when the prediction accuracy of the intraoperative real-time comprehensive risk score of the fusion features reaches a preset threshold in the validation set.

[0017] Preferably, in S2, the context correlation analysis of different stages of surgery refers to establishing a time sequence feature sequence according to the order of surgery stages for the intraoperative multi-source modal features in different stages of surgery, and using a cross-stage mode migration mechanism to map the intraoperative multi-source modal features of the previous stage to the feature space of the current stage.

[0018] The cross-stage mode migration mechanism is used to establish risk feature transmission between different stages of surgery during the surgery process, specifically as follows: a time sequence correlation modeling network is used to model the stage feature sequence to calculate the feature dependency strength between stages; for stage pairs with feature dependency strength higher than a preset threshold, the intraoperative multi-source modal features of the previous stage are projected to the feature space of the current stage through a feature mapping function, and context feature completion processing is performed on the intraoperative multi-source modal features corresponding to the current stage.

[0019] Preferably, in S2, the subjective and objective consistency comparison analysis is a similarity calculation of the intraoperative subjective modal features and the intraoperative objective modal features, which is used to identify the risk cognitive bias between subjective cognition and objective indicators, and serves as an influencing factor for feature fusion weight adjustment in the generated risk assessment comprehensive feature set; the specific method of subjective and objective consistency comparison analysis is as follows:

[0020] The intraoperative subjective modal features and the intraoperative objective modal features are obtained, and a similarity calculation function is used to match the similarity of the intraoperative subjective modal features and the intraoperative objective modal features of the same stage to obtain the subjective and objective consistency comparison results; the subjective and objective consistency comparison results are introduced as a constraint factor into a subjective and objective consistency weight function to dynamically adjust the fusion weight of the intraoperative subjective modal features and the intraoperative objective modal features, and a risk assessment comprehensive feature set is generated by fusion.

[0021] Preferably, in S3, the risk assessment prediction AI model is constructed based on a multi-risk causal interaction graph, a causal graph embedding modeling mechanism, and a doctor judgment priority weight adjustment mechanism, and is used for intraoperative real-time risk score prediction of the risk assessment comprehensive feature set.

[0022] The construction method of the risk assessment prediction AI model is as follows: inputting the risk assessment comprehensive feature set into a causal structure modeling network, modeling the statistical dependence between the complication risk indicators and the functional risk indicators by a causal graph, generating a stable risk causal interaction representation; based on a multi-task prediction structure, training the scoring output network of each risk indicator, establishing a corresponding nonlinear prediction path for each risk indicator; introducing a doctor judgment priority weight adjustment mechanism to construct a risk weight adjustment function to dynamically adjust the prediction weight of each risk indicator; jointly optimizing the causal interaction modeling and the risk score prediction tasks, and when the scoring error in the verification set converges and reaches a preset precision threshold, the training is terminated and the model parameters are fixed.

[0023] Preferably, the multi-risk causal interaction graph is composed of a plurality of nodes representing intraoperative risk events and directed edges therebetween, for representing the causal path relationship between the plurality of risk events and serving as the structured input of the causal graph embedding modeling mechanism; wherein the nodes are set based on the intraoperative risk event types recorded in the historical sample data; and the directed edges are constructed based on the time sequence relationship between different risk events in the sample data and the causal correlation defined in the clinical knowledge.

[0024] The causal graph embedding modeling mechanism uses a graph neural network model to extract causal dependence embedding features of each intraoperative risk event based on the multi-risk causal interaction graph, and generates node embedding feature vectors of the causal correlation relationship between the complication risk indicators and the functional risk indicators.

[0025] The risk causal influence relationship between the complication risk indicators and the functional risk indicators is calculated and analyzed based on the multi-risk causal interaction graph and the causal graph embedding modeling mechanism, and the specific method is as follows:

[0026] The node embedding feature vectors are generated by the causal graph embedding modeling mechanism based on the multi-risk causal interaction graph, the causal influence strength between the nodes is calculated, and the risk causal relationship embedding matrix between the complication risk indicators and the functional risk indicators is constructed, i.e., the risk causal influence relationship.

[0027] Preferably, the doctor judgment priority weight adjustment mechanism is based on intraoperative subjective data to extract a subjective risk weight factor, and combines the intraoperative real-time comprehensive risk score of the previous surgery stage to dynamically adjust the priority and prediction weight of each risk indicator to predict the intraoperative real-time comprehensive risk score of the current surgery stage.

[0028] Preferably, in S4, the intraoperative real-time comprehensive risk score is matched with the corresponding risk prompt information and displayed in a graphical manner in real time on the doctor terminal interface to assist in intraoperative risk intervention decision-making.

[0029] In another aspect, the present application provides a urinary surgery risk AI intelligent evaluation system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the urinary surgery risk AI intelligent evaluation method described above.

[0030] Compared with the prior art, the above technical scheme of the present application has the following beneficial technical effects:

[0031] 1. In the present application, by correlating the data at different stages of the surgery process and comparing the differences between the doctor's judgment and the actual data, the possible risks at each stage can be more comprehensively sorted out;

[0032] 2. In the present application, by constructing a causal relationship graph between risks and dynamically adjusting the prediction weight of key risks in combination with the real-time judgment of the doctor, it can more accurately identify which risks are more worthy of attention, especially in the case of concurrent complications and functional problems, which helps the doctor make more reasonable decisions during the operation. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 A flowchart of an embodiment of the present application is shown. DETAILED DESCRIPTION

[0034] Embodiment one, as shown in the figure, the present application provides a urinary surgery risk AI intelligent evaluation method, the specific implementation steps are as follows: Figure 1

[0035] S1, real-time acquisition of intraoperative data of urological surgery process, and time synchronization processing and pretreatment;

[0036] Among them, the intraoperative data includes intraoperative objective data and intraoperative subjective data;

[0037] S2, input the intraoperative data into the risk feature fusion AI model to extract the intraoperative multi-source modal features, and perform context correlation analysis of different stages of the surgery, and perform consistency comparison analysis on the intraoperative multi-source modal features in the context correlation analysis process, obtain the consistency comparison result, and fuse each intraoperative multi-source modal feature to generate a risk evaluation comprehensive feature set;

[0038] S3, using the risk evaluation prediction AI model to analyze the risk evaluation comprehensive feature set, calculating and analyzing the risk causal influence relationship between the complication risk index and the functional risk index based on the multi-risk causal interaction graph and the causal graph embedding modeling mechanism, dynamically adjusting the priority and prediction weight of each risk index in combination with the doctor's judgment priority weight adjustment mechanism, and predicting to obtain the real-time comprehensive risk score in the operation;

[0039] ​S4, real-time visualization of the intraoperative real-time comprehensive risk score and risk prompt information in the terminal.

[0040] In the embodiment S1, the intraoperative objective data includes intraoperative video data, physiological parameter data, intraoperative image data, intraoperative pathological detection data, operating room environment data and surgical instrument state data; the intraoperative subjective data is real-time language judgment information of the leading surgeon acquired through a voice acquisition device;

[0041] The time synchronization processing and preprocessing are operations on the intraoperative objective data and the intraoperative subjective data, and are specifically as follows:

[0042] Based on a unified time reference, the intraoperative objective data and the intraoperative subjective data are time stamped, format standardized, noise removed and missing value filled according to the data type, and the intraoperative subjective data is converted into structured labels representing the surgical stage, the patient state and the risk type.

[0043] In the embodiment S1, the acquisition of the intraoperative objective data includes the following specific manners: the intraoperative video data is acquired by a high frame rate monitoring camera device in the operating room, and the frame rate is not less than 30 fps; the physiological parameter data is acquired by a multi-parameter monitoring device connected to the patient in real time, and the parameter types include but are not limited to heart rate, blood pressure, blood oxygen saturation, body temperature and respiratory rate, etc., and the data frequency needs to be higher than 0.5 Hz; the intraoperative image data includes image sequences generated by intraoperative ultrasound, CT or intraoperative navigation system, and must have DICOM format standard and time stamp information; the intraoperative pathological detection data is mainly derived from the combination of intraoperative rapid frozen section analysis results and AI image recognition results, and is input synchronously through a double-channel path; the operating room environment data can be acquired by an Internet of Things sensing network to obtain indoor temperature and humidity, illumination intensity, air pressure and cleanliness index, etc., and the sensing frequency is recommended to be set at a 5-minute sampling period; the surgical instrument state data is read through an interface connected to the instrument control system or the intelligent operating bed, and the instrument usage state, opening time, operation mode and historical operation record, etc. are acquired, and the application period and object in the surgical process are marked synchronously.

[0044] In the embodiment S1, in terms of intraoperative subjective data acquisition, a voice acquisition device is used to acquire the language judgment information of the leading surgeon in real time, the acquisition system should have a noise reduction function and a sound source directional recognition mechanism to filter background environmental interference and distinguish the main speaker role, the voice data needs to be converted into structured text labels through a voice recognition system after acquisition, and a semantic analysis model is called to perform intent recognition and keyword extraction, and finally a structured label sequence representing the surgical stage, the patient state and the risk type is constructed.

[0045] In the embodiment S1, the time synchronization and preprocessing of the intraoperative objective data and the intraoperative subjective data are as follows: a unified time reference construction mechanism is used to connect all acquisition devices to a central clock synchronization system, a timestamp mapping algorithm is used to correct the time axis of each modality data, and interpolation or clipping processing is performed according to the sampling frequency to construct equal time interval data frames. When performing format standardization, the image data is uniformly converted into a gray image tensor with a resolution of 512*512, the time series numerical data is converted into a normalized floating point array sequence, and the semantic label is uniformly converted into a triple format, indicating (category, content, time). At the same time, a noise detection mechanism is used to identify abnormal jump points of physiological parameters, image blur frames and non-speaking sections of voice, and a missing value filling strategy is used for repair.

[0046] In the embodiment S2, the risk feature fusion AI model is constructed based on a multi-source data feature encoding network, a surgical stage context modeling network and a consistency comparison feature fusion network, and is used for feature extraction, context association analysis and consistency comparison feature fusion of the intraoperative objective data and the intraoperative subjective data to generate a risk assessment comprehensive feature set.

[0047] The construction method of the risk feature fusion AI model is as follows:

[0048] A feature encoding subnetwork is constructed to encode the intraoperative objective data and the intraoperative subjective data respectively to obtain intraoperative multi-source modality features. A surgical stage context modeling subnetwork is constructed to model the time series correlation of the intraoperative multi-source modality feature sequence in different surgical stages, and to perform cross-stage mode migration during modeling to map the intraoperative multi-source modality features of the previous stage to the feature space of the current stage. A consistency comparison feature fusion subnetwork is constructed to introduce a constraint loss function based on the consistency comparison result of the subjective and objective data in the model training to adjust the fusion weight of the intraoperative multi-source modality features. The risk feature fusion AI model is jointly trained based on multi-source training samples, and the training is terminated when the prediction accuracy of the intraoperative real-time comprehensive risk score of the fusion features reaches a preset threshold in the validation set.

[0049] In the embodiment S2, the intraoperative multi-source modality features refer to the modality feature representations obtained by processing the intraoperative objective data and the intraoperative subjective data through the feature encoding subnetwork, and the intraoperative multi-source modality features include intraoperative subjective modality features and intraoperative objective modality features. The intraoperative objective modality features include intraoperative video space-time features, physiological parameter time series features, intraoperative image texture features, intraoperative pathological index features, operating room environment state features and surgical instrument operation features, and the intraoperative subjective modality features are feature representations of real-time language judgment information of the leading surgeon after semantic encoding.

[0050] In the embodiment, the multi-source data feature encoding network is composed of two independent neural networks, corresponding to the feature extraction process of intraoperative objective data and intraoperative subjective data respectively; for different data types in the intraoperative objective data, a branch structure is used to construct a convolutional encoding network for processing image data (intraoperative video frames, intraoperative image images), a time series embedding network for processing physiological parameter and environmental data sequences, and a graph structure embedding network for processing device interdependence information of surgical instrument state data; after the intraoperative subjective data is converted by speech and structured labels, a semantic feature encoder based on the Transformer structure is used for embedding mapping to output a high-dimensional feature representation representing the content of the doctor's subjective judgment; the surgery stage context modeling network models the intraoperative multi-source modal feature sequence of each surgery stage based on the Bi-LSTM structure, uses the stage timestamp as the stage division basis, and inputs the feature sequence divided by stage into the context modeling network for time series modeling processing. The stage indication variable embedded in the model is used to maintain the consistency of the stage semantics; the cross-stage mode migration mechanism realizes the mapping of the intraoperative multi-source modal features of the previous stage to the feature space of the current stage through the construction of a stage migration mapping matrix, and after being concatenated with the original features of the current stage, it is normalized as the feature input of the current stage; the consistency comparison feature fusion network uses a parallel fusion architecture, introduces a subjective and objective consistency loss function in the training stage, and the loss function is propagated in the reverse direction according to the similarity difference between the intraoperative subjective modal features and the intraoperative objective modal features, adjusts the fusion weight parameters in the feature fusion layer, and finally generates a fused risk assessment comprehensive feature set; in the joint training process, the cross-entropy loss function is used as the main loss function, and the intraoperative risk score prediction result under the fused feature input is used as the supervision target for training. The termination condition of model training is that the prediction accuracy of the intraoperative real-time comprehensive risk score on the validation set reaches the set threshold and is stable without falling within multiple epochs.

[0051] In the embodiment S2, the context association analysis of different stages of surgery means that the intraoperative multi-source modal features are established in time sequence according to the order of surgery stages, and the cross-stage mode migration mechanism is used to map the intraoperative multi-source modal features of the previous stage to the feature space of the current stage.

[0052] The cross-stage mode migration mechanism is used to establish the risk feature transmission between different surgery stages during the surgery process, specifically as follows: a time sequence association modeling network is used to model the stage feature sequence, and the feature dependence strength between stages is calculated; for the stage pair with a feature dependence strength higher than a preset threshold, the intraoperative multi-source modal features of the previous stage are projected to the feature space of the current stage through a feature mapping function, and context feature completion processing is performed on the intraoperative multi-source modal features corresponding to the current stage.

[0053] In the embodiment S2, the context correlation analysis of different stages of the operation is based on the stage division of the intraoperative time sequence, each stage is controlled by the pre-labeled operation stage label, and the sequence structure of the intraoperative multi-source modality features at the stage level is constructed; for the stage sequence, a bidirectional long short-term memory network is used for modeling to generate the context feature state vector of each stage; the cross-stage mode migration mechanism calculates the feature dependence strength of the modeled stage context vector, and the dependence strength is jointly determined by the attention score between the previous stage and the current stage and the gradient normalized similarity; when the dependence strength is higher than the set threshold, the intraoperative multi-source modality features of the previous stage are converted to the feature space of the current stage through a feature mapping function, the mapping features and the original intraoperative multi-source modality features of the current stage are spliced at the vector level, and the completed context fusion features are generated as the feature input of the current stage through a normalization fusion module.

[0054] In the embodiment S2, the subjective-objective consistency comparison analysis is similarity calculation of the intraoperative subjective modality features and the intraoperative objective modality features, which is used to identify the risk cognitive bias between the subjective cognition and the objective indicators, and is used as an influencing factor for adjusting the feature fusion weight of the risk assessment comprehensive feature set; the specific method of the subjective-objective consistency comparison analysis is as follows:

[0055] The intraoperative subjective modality features and the intraoperative objective modality features are obtained, and a similarity calculation function is used to match the similarity of the intraoperative subjective modality features and the intraoperative objective modality features in the same stage to obtain the subjective-objective consistency comparison result; the subjective-objective consistency comparison result is used as a constraint factor, introduced into a subjective-objective consistency weight function, and the fusion weight of the intraoperative subjective modality features and the intraoperative objective modality features is dynamically adjusted to generate a risk assessment comprehensive feature set.

[0056] In the embodiment S2, the subjective-objective consistency comparison analysis includes constructing a feature pair set after modality normalization processing of the intraoperative subjective modality features and the intraoperative objective modality features, respectively, the intraoperative subjective modality features are structured feature vectors obtained by semantic analysis, emotion recognition and context identification of the intraoperative language information of the leading surgeon, and the intraoperative objective modality features include a combined vector composed of physiological parameter features, image recognition features, instrument operation features and environmental change features, each pair of subjective and objective feature vectors in the feature pair set is matched for similarity using a weighted cosine similarity function, the obtained subjective-objective consistency comparison result is converted into a subjective-objective consistency weight factor through a preset mapping function, and the factor is input into a feature fusion network as a constraint parameter to adjust the fusion proportion of each modality feature, control the weight deviation of the subjective features and the objective features in the fusion result while maintaining the semantic integrity, and finally generate a fused risk assessment comprehensive feature set.

[0057] In the embodiment S3, the risk assessment prediction AI model is constructed based on a multi-risk causal interaction graph, a causal graph embedding modeling mechanism, and a doctor judgment priority weight adjustment mechanism, and is used for predicting intraoperative risk scores of the risk assessment comprehensive feature set in real time;

[0058] The construction method of the risk assessment prediction AI model is as follows: the risk assessment comprehensive feature set is input into a causal structure modeling network, a causal graph modeling is performed on statistical dependency relationships between the complication risk indicators and the functional risk indicators, a stable risk causal interaction representation is generated; based on a multi-task prediction structure, score output networks of the risk indicators are jointly trained, corresponding nonlinear prediction paths of the risk indicators are established; a doctor judgment priority weight adjustment mechanism is introduced, a risk weight adjustment function is constructed to dynamically adjust prediction weights of the risk indicators; the two tasks of causal interaction modeling and risk score prediction are jointly optimized, when score errors in a verification set converge and reach a preset precision threshold, the training is terminated and model parameters are fixed.

[0059] In the embodiment S3, in the risk assessment prediction AI model, the causal structure modeling network is used for graph structure modeling based on a multi-risk causal interaction graph, mapping feature nodes related to intraoperative risk events in the risk assessment comprehensive feature set to risk nodes corresponding to the multi-risk causal interaction graph, and constructing a directed causal edge structure based on time sequences, joint occurrence probabilities, and clinical knowledge between nodes, establishing a stable causal graph representation structure for causal paths between the complication risk indicators and the functional risk indicators, the graph structure performs node information propagation and aggregation through a graph neural network, and outputs a risk causal interaction feature matrix; the multi-task prediction structure includes a plurality of risk score sub-networks, each sub-network is used for predicting a single risk indicator, each sub-network receives the shared risk causal interaction feature matrix as input, and performs nonlinear feature mapping and score output through an independent parameter path to form a plurality of risk score output channels; score results predicted by the sub-networks are jointly used for loss function calculation to form a joint training target, the loss function is composed of a causal modeling error and a risk score error, a multi-task loss balance parameter is set in the target function of joint optimization, and the causal modeling task and the score prediction task are converged at the same time.

[0060] In the embodiment S3, the doctor judges the priority weight adjustment mechanism to construct a weight adjustment function in the model training process, which dynamically adjusts the loss function weight coefficients of each risk indicator based on the subjective risk weight factors extracted from the intraoperative subjective data and the comprehensive score output of the historical stage, and participates in the error back propagation in the model training stage; the total score error is evaluated on the validation set after each iteration in the training process, and the error reduction trend is recorded; when the error reduction of consecutive iterations is less than the preset change threshold, and the current score error is less than the preset precision threshold, the training process is terminated and all model parameters are fixed, which are used for real-time scoring prediction of intraoperative risk in the inference stage.

[0061] In the embodiment S3, the multi-risk causal interaction graph is composed of multiple nodes representing intraoperative risk events and directed edges between them, used to represent the causal path relationship between multiple risk events, and as the structured input of the causal graph embedding modeling mechanism; wherein the nodes are set based on the intraoperative risk event types recorded in the historical sample data; the directed edges are constructed based on the time sequence relationship between different risk events in the sample data and the causal association defined in the clinical knowledge;

[0062] The causal graph embedding modeling mechanism is based on the multi-risk causal interaction graph, uses a graph neural network model to extract causal dependence embedding features of each intraoperative risk event, and generates node embedding feature vectors of the causal association relationship between the complication risk indicators and the functional risk indicators;

[0063] The causal graph embedding modeling mechanism is based on the multi-risk causal interaction graph, uses a graph neural network model to extract causal dependence embedding features of each intraoperative risk event, and generates node embedding feature vectors of the causal association relationship between the complication risk indicators and the functional risk indicators;

[0064] The causal graph embedding modeling mechanism generates node embedding feature vectors based on the multi-risk causal interaction graph, calculates the causal influence strength between nodes, and constructs the risk causal relationship embedding matrix between the complication risk indicators and the functional risk indicators, which is the risk causal influence relationship.

[0065] In the embodiment S3, the multi-risk causal interaction graph is a directed graph for representing the causal path structure between various intraoperative risk events, and the construction process includes two steps of node setting and edge construction; in the node setting, the complete set of intraoperative risk events is extracted based on the types of intraoperative risk events recorded in the historical surgery sample data, each type of risk event is defined as a node in the graph, and each node in the node set has a unique identifier and is bound to the corresponding event attribute; in the edge construction, all risk events are sorted in sequence using the event timestamps annotated in the sample data, and the directed edges are established between the node pairs that have a time sequence relationship and have a medical causal logic based on the judgment rules about the causal mechanism between events in the existing medical clinical knowledge base, the directed edges take the starting node as the cause event and the target node as the result event, and the causal weight parameter is added to record the association strength between events.

[0066] In the embodiment S3, in the causal graph embedding modeling mechanism, the multi-risk causal interaction graph is used as the structured input of the graph neural network, the initial risk event features of each node are embedded and propagated and aggregated by multiple layers using the graph convolution network, the causal dependence embedding feature vector of each node is generated, representing the strength mode of the node being influenced and influencing others in the causal path; subsequently, the node embedding features between the complication risk indicator nodes and the functional risk indicator nodes are calculated two by two, the causal influence strength is calculated by the vector similarity calculation function and the path dependence function, the risk causal relationship embedding matrix is constructed, the embedding matrix is the quantitative expression result of the causal action between the complication risk and the functional risk, and is used as the risk weight adjustment and prediction input in the subsequent steps.

[0067] In the embodiment S3, the doctor judgment priority weight adjustment mechanism is based on the subjective risk weight factor extracted from the intraoperative subjective data, and combines the intraoperative real-time comprehensive risk score of the previous surgery stage to dynamically adjust the priority and prediction weight of each risk indicator, so as to predict the intraoperative real-time comprehensive risk score of the current surgery stage.

[0068] In the embodiment S3, the doctor judgment priority weight adjustment mechanism is used to dynamically adjust the prediction weights of the risk indicators in the operation according to the real-time judgment tendency of the leading doctor, and the specific process is as follows: structured label information containing the doctor's risk judgment intention is extracted from the intraoperative subjective data, the label information contains the doctor's identification type, subjective weight level and judgment time of the current stage risk event, the system converts the label information into a subjective risk weight factor set and performs stage classification and type mapping; before entering the risk assessment of the current operation stage, the real-time comprehensive risk score result generated in the previous operation stage is called as the starting point of the current assessment, and the subjective risk weight factor of the current stage is combined to perform the risk indicator priority adjustment operation, that is, the risk indicators with higher matching degree in the risk indicator set are given higher priority order, and the weight factor coefficient of the non-linear prediction path is dynamically increased; before performing the risk prediction calculation in the current stage, the indicator weight coefficient after the priority weight adjustment is used to update the weight distribution configuration of the risk assessment prediction AI model, so that it is more suitable for the real-time judgment structure of the doctor and outputs the intraoperative real-time comprehensive risk score of the current stage.

[0069] In the embodiment S3, the intraoperative real-time comprehensive risk score and the corresponding risk prompt information are matched and displayed in the embodiment S4, and are displayed in a graphical manner in the doctor terminal interface in real time, which is used to assist the intraoperative risk intervention decision.

[0070] In the embodiment S3, the risk prompt information includes prompt number, risk type label, warning level, risk indicators that may be involved, and recommended intervention opportunity and operation action; the graphical manner includes hierarchical color identification of risk score, icon label of risk prompt, index heat bar, and phase-by-phase score trend curve, and the like visual elements, and the display data unit is updated at a fixed refresh period to ensure continuous display of intraoperative information.

[0071] In the embodiment S3, the risk prompt information includes prompt number, risk type label, warning level, risk indicators that may be involved, and recommended intervention opportunity and operation action; the graphical manner includes hierarchical color identification of risk score, icon label of risk prompt, index heat bar, and phase-by-phase score trend curve, and the like visual elements, and the display data unit is updated at a fixed refresh period to ensure continuous display of intraoperative information.

[0072] The embodiments of the application are described in detail above with reference to the drawings, but the application is not limited thereto, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. A urinary surgery risk AI intelligent evaluation method, characterized in that, The method comprises the following steps: S1, collecting intraoperative data of a urological surgery process in real time, and performing time synchronization processing and preprocessing; The intraoperative data comprises objective intraoperative data and subjective intraoperative data; S2, inputting the intraoperative data into a risk feature fusion AI model to extract intraoperative multi-source modal features, and performing context correlation analysis of different stages of the surgery, in the context correlation analysis process, performing subjective-objective consistency comparison analysis on the intraoperative multi-source modal features, obtaining a consistency comparison result, and fusing each intraoperative multi-source modal feature to generate a risk assessment comprehensive feature set; S3, using a risk assessment prediction AI model to analyze the risk assessment comprehensive feature set, calculating and analyzing the risk causal influence relationship between the complication risk index and the functional risk index based on a multi-risk causal interaction graph and a causal graph embedding modeling mechanism, dynamically adjusting the priority and prediction weight of each risk index combined with a doctor judgment priority weight adjustment mechanism, and predicting to obtain an intraoperative real-time comprehensive risk score; S4, displaying the intraoperative real-time comprehensive risk score and risk prompt information in real time on a terminal.

2. The urinary surgery risk AI intelligent evaluation method according to claim 1, characterized in that: In S1, the objective intraoperative data comprises intraoperative video data, physiological parameter data, intraoperative image data, intraoperative pathological detection data, operating room environment data and surgical instrument state data; the subjective intraoperative data is real-time language judgment information of the leading doctor acquired through a voice acquisition device; The time synchronization processing and preprocessing is an operation on the objective intraoperative data and the subjective intraoperative data, and specifically as follows: Based on a unified time reference, the objective intraoperative data and the subjective intraoperative data are time stamped, standardized in format, noise removed, and missing value filled, and the subjective intraoperative data is converted into structured labels representing the stage of surgery, the state of the patient and the type of risk.

3. The urinary surgery risk AI intelligent evaluation method according to claim 2, characterized in that: In S2, the risk feature fusion AI model is constructed based on a multi-source data feature coding network, a surgery stage context modeling network and a consistency comparison feature fusion network, and is used for feature extraction, context correlation analysis and consistency comparison feature fusion of the objective intraoperative data and the subjective intraoperative data to generate a risk assessment comprehensive feature set; The construction method of the risk feature fusion AI model is as follows: A feature coding subnetwork is constructed to code the modal features of the objective intraoperative data and the subjective intraoperative data to obtain intraoperative multi-source modal features; A surgery stage context modeling subnetwork is constructed to model the time sequence correlation of the intraoperative multi-source modal feature sequence of different surgery stages, and to perform cross-stage mode migration in the modeling process to map the intraoperative multi-source modal features of the previous stage to the feature space of the current stage; a consistency comparison feature fusion subnetwork is constructed to introduce a constraint loss function based on the subjective-objective consistency comparison result in model training to adjust the fusion weight of the intraoperative multi-source modal features; the risk feature fusion AI model is jointly trained based on multi-source training samples, and the training is terminated when the prediction accuracy rate of the intraoperative real-time comprehensive risk score of the fusion features reaches a preset threshold in the verification set.

4. The urinary surgery risk AI intelligent evaluation method according to claim 3, characterized in that: In S2, the context correlation analysis of different stages of the operation refers to establishing a time sequence feature sequence in the order of the operation stages according to the intraoperative multi-source modal characteristics in different stages of the operation, and using a cross-stage mode migration mechanism to map the intraoperative multi-source modal characteristics of the previous stage to the feature space of the current stage; The cross-stage mode migration mechanism is used to establish the risk feature transmission between different stages of the operation, and the specific method is as follows: a time sequence correlation modeling network is used to model the stage feature sequence, and the feature dependency strength between stages is calculated; for the stage pair with a feature dependency strength higher than a preset threshold, the intraoperative multi-source modal characteristics of the previous stage are projected to the feature space of the current stage through a feature mapping function, and the context feature completion processing is performed on the intraoperative multi-source modal characteristics corresponding to the current stage.

5. The urinary surgery risk AI intelligent evaluation method according to claim 4, characterized in that: In S2, the subjective-objective consistency comparison analysis is a similarity calculation of the intraoperative subjective modal characteristics and the intraoperative objective modal characteristics, which is used to identify the risk cognitive bias between the subjective cognition and the objective indicators, and serves as an influencing factor for adjusting the feature fusion weight of the risk assessment comprehensive feature set; the specific method of the subjective-objective consistency comparison analysis is as follows: The intraoperative subjective modal characteristics and the intraoperative objective modal characteristics are obtained, and a similarity calculation function is used to perform similarity matching on the intraoperative subjective modal characteristics and the intraoperative objective modal characteristics in the same stage to obtain a subjective-objective consistency comparison result; the subjective-objective consistency comparison result is introduced into a subjective-objective consistency weight function as a constraint factor to dynamically adjust the fusion weight of the intraoperative subjective modal characteristics and the intraoperative objective modal characteristics, and a risk assessment comprehensive feature set is generated by fusion.

6. The urinary surgery risk AI intelligent evaluation method according to claim 5, characterized in that: In S3, the risk assessment prediction AI model is constructed based on a multi-risk causal interaction graph, a causal graph embedding modeling mechanism and a doctor judgment priority weight adjustment mechanism, and is used for real-time risk score prediction of the risk assessment comprehensive feature set during the operation; The construction method of the risk assessment prediction AI model is as follows: the risk assessment comprehensive feature set is input into a causal structure modeling network, a causal graph modeling is performed on the statistical dependency relationship between the complication risk indicators and the functional risk indicators, and a stable risk causal interaction representation is generated; based on a multi-task prediction structure, a scoring output network of each risk indicator is trained to establish a corresponding nonlinear prediction path for each risk indicator; A doctor judgment priority weight adjustment mechanism is introduced to construct a risk weight adjustment function to dynamically adjust the prediction weight of each risk indicator; The causal interaction modeling and risk score prediction tasks are jointly optimized, and when the scoring error in the verification set converges and reaches a preset precision threshold, the training is terminated and the model parameters are fixed.

7. The urinary surgery risk AI intelligent evaluation method according to claim 6, characterized in that: The multi-risk causal interaction graph is composed of multiple nodes representing intraoperative risk events and directed edges between them, which is used to represent the causal path relationship between multiple risk events and serves as the structured input of the causal graph embedding modeling mechanism; wherein the nodes are set based on the intraoperative risk event types recorded in the historical sample data; the directed edges are constructed based on the time sequence relationship between different risk events in the sample data and the causal correlation defined in the clinical knowledge; The causal graph embedding modeling mechanism is based on a multi-risk causal interaction graph, uses a graph neural network model to extract causal dependence embedding features of each intraoperative risk event, and generates a node embedding feature vector of the causal association relationship between the complication risk indicators and the functional risk indicators; The risk causal influence relationship between the complication risk indicators and the functional risk indicators is calculated and analyzed based on the multi-risk causal interaction graph and the causal graph embedding modeling mechanism, and the specific method is as follows: The node embedding feature vector is generated based on the multi-risk causal interaction graph by the causal graph embedding modeling mechanism, the causal influence strength between nodes is calculated, and a risk causal relationship embedding matrix between the complication risk indicators and the functional risk indicators is constructed, that is, the risk causal influence relationship.

8. The urinary surgery risk AI intelligent evaluation method according to claim 7, characterized in that: The doctor judgment priority weight adjustment mechanism is based on the extraction of a subjective risk weight factor from intraoperative subjective data, and combines the intraoperative real-time comprehensive risk score of the previous surgical stage to dynamically adjust the priority and prediction weight of each risk indicator to predict the intraoperative real-time comprehensive risk score of the current surgical stage. 9.The urinary surgery risk AI intelligent evaluation method according to claim 8, characterized in that: In S4, the intraoperative real-time comprehensive risk score is matched with the corresponding risk prompt information and displayed in a graphical manner in real time on the doctor terminal interface to assist in intraoperative risk intervention decision-making.

10. A urinary surgery risk AI intelligent evaluation system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the urinary surgery risk AI intelligent evaluation method of any one of claims 1-9.

Images