Intraoperative intervention decision determination method and device, equipment and storage medium

Abstract

The invention relates to an intraoperative intervention decision determination method and device, equipment and a storage medium. The method comprises the following steps: acquiring an electronic case, a physiological feature information set and a target cue word template of a target object; the physiological feature information set is a set composed of physiological feature information of the target object in the operation; determining a dynamic abstract feature vector based on the electronic case, the physiological feature information set and the target cue word template; determining a physiological feature vector based on the physiological feature information set; and determining an intraoperative intervention decision based on the dynamic abstract feature vector and the physiological feature vector. By adopting the method, the accuracy of an intraoperative intervention decision can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an intraoperative intervention decision determination method and device, equipment and a storage medium. BACKGROUND

[0002] With the development of artificial intelligence technology, a large number of neural network models have emerged. Neural network models can process structured and unstructured information and have strong semantic understanding and context reasoning capabilities, providing new possibilities for intraoperative intervention decision-making.

[0003] In traditional technology, intraoperative drug intervention decisions are mainly determined by anesthesiologists or surgeons. Since complex and rapidly changing multi-dimensional physiological characteristic information (such as heart rate, blood pressure, blood oxygen saturation, and respiratory rate) is generated during surgery, it is difficult for anesthesiologists or surgeons to comprehensively integrate and analyze the above information within a limited time, so they often have to rely on experience to make decisions, resulting in low accuracy of intraoperative intervention decisions. SUMMARY

[0004] Therefore, it is necessary to provide an intraoperative intervention decision determination method, device, equipment and storage medium capable of improving the accuracy of intraoperative intervention decisions to solve the above technical problems.

[0005] In a first aspect, the present application provides an intraoperative intervention decision determination method, comprising:

[0006] obtaining an electronic case, a physiological characteristic information set and a target prompt word template of a target object; the physiological characteristic information set is a set of physiological characteristic information of the target object during surgery;

[0007] determining a dynamic summary feature vector based on the electronic case, the physiological characteristic information set and the target prompt word template;

[0008] determining a physiological characteristic vector based on the physiological characteristic information set;

[0009] determining an intraoperative intervention decision based on the dynamic summary feature vector and the physiological characteristic vector.

[0010] In some exemplary embodiments, determining a dynamic summary feature vector based on the electronic case, the physiological characteristic information set and the target prompt word template comprises:

[0011] determining an abnormal feature based on the physiological characteristic information set and an abnormal threshold;

[0012] determining abnormal medical information corresponding to the abnormal feature based on the abnormal feature and a medical knowledge graph;

[0013] dividing and compressing the physiological characteristic information set to obtain a physiological characteristic information sub-set sequence;

[0014] inputting the electronic case, the abnormal feature, the abnormal medical information, the sub-set sequence of physiological feature information, and the target prompt word template into the generative large model to obtain a dynamic summary of the target object;

[0015] inputting the dynamic summary into a text embedding model to obtain a dynamic summary feature vector of the dynamic summary.

[0016] In some example embodiments, based on the abnormal feature and the medical knowledge graph, the abnormal medical information corresponding to the abnormal feature is determined, including:

[0017] determining an abnormal feature vector corresponding to the abnormal feature;

[0018] obtaining a node feature vector corresponding to each node in the medical knowledge graph, and determining a similarity between the abnormal feature vector and each node feature vector;

[0019] based on the similarity corresponding to each node, determining a disease node, a drug node, and a complication node corresponding to the abnormal feature;

[0020] based on the disease node, the drug node, and the complication node, determining the abnormal medical information corresponding to the abnormal feature.

[0021] In some example embodiments, based on the set of physiological feature information, a physiological feature vector is determined, including:

[0022] dividing the set of physiological feature information based on a preset time step to obtain a sub-set sequence of physiological feature information; the sub-sets of physiological feature information in the sub-set sequence of physiological feature information are arranged in chronological order;

[0023] using a time sequence encoding unit to sequentially encode the sub-sets of physiological feature information in the sub-set sequence of physiological feature information;

[0024] until the last sub-set of physiological feature information, obtaining the physiological feature vector.

[0025] In some example embodiments, based on the dynamic summary feature vector and the physiological feature vector, an intraoperative intervention decision is determined, including:

[0026] performing two-way cross multi-head attention fusion on the dynamic summary feature vector and the physiological feature vector to obtain a fusion feature vector;

[0027] based on the fusion feature vector, determining the intraoperative intervention decision.

[0028] In some example embodiments, based on the fusion feature vector, the intraoperative intervention decision is determined, including:

[0029] input the fusion feature vector into an intervention probability binary classification model to obtain an intervention probability;

[0030] input the fusion feature vector into a multi-label model to obtain an intervention vector; the intervention vector includes probabilities of using various intervention drugs and a probability of not using an intervention drug;

[0031] generate an intraoperative intervention decision based on the intervention probability and the intervention vector.

[0032] In a second aspect, the present application further provides an intraoperative intervention decision determination apparatus, comprising:

[0033] an acquisition module configured to acquire an electronic case of a target object, a physiological feature information set, and a target prompt word template; the physiological feature information set is a set of physiological feature information of the target object during surgery;

[0034] a first vector generation module configured to determine a dynamic summary feature vector based on the electronic case, the physiological feature information set, and the target prompt word template;

[0035] a second vector generation module configured to determine a physiological feature vector based on the physiological feature information set;

[0036] a decision generation module configured to determine an intraoperative intervention decision based on the dynamic summary feature vector and the physiological feature vector.

[0037] In a third aspect, the present application further provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of any method in the first aspect when executing the computer program.

[0038] In a fourth aspect, the present application further provides a computer readable storage medium, which stores a computer program, and the computer program implements the steps of any method in the first aspect when executed by a processor.

[0039] In a fifth aspect, the present application further provides a computer program product, comprising a computer program, and the computer program implements the steps of any method in the first aspect when executed by a processor.

[0040] The method, device, computer device, storage medium and computer program product for determining the intraoperative intervention decision, obtain an electronic medical record, a physiological characteristic information set and a target prompt word template of a target object; the physiological characteristic information set is a set of physiological characteristic information of the target object in surgery; a dynamic summary feature vector is determined based on the electronic medical record, the physiological characteristic information set and the target prompt word template; a physiological characteristic vector is determined based on the physiological characteristic information set; and an intraoperative intervention decision is determined based on the dynamic summary feature vector and the physiological characteristic vector. The dynamic summary feature vector is determined based on the electronic medical record, the physiological characteristic information set and the target prompt word template, that is, the real-time time series data of the physiological characteristics and the semantics and structure of the electronic medical record are expressed as a dynamic summary vector representing clinical understanding, the physiological characteristic vector is determined based on the physiological characteristic information set, that is, the real-time dynamic changes of the multi-dimensional physiological characteristics of the patient in surgery are captured based on the physiological characteristic information set, and the intraoperative intervention decision is determined by combining the dynamic summary feature vector and the physiological characteristic vector, so as to realize comprehensive integration and modeling of the multi-dimensional physiological characteristics and the medical record information of the patient in surgery, accurately capture the changes of the real-time physiological state of the patient, avoid information one-sidedness caused by relying on artificial experience, and significantly improve the accuracy of the intraoperative intervention decision. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the description of the embodiments of the present application or the related art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other related drawings can be obtained by those skilled in the art without creative labor.

[0042] Figure 1 An application environment diagram of the intraoperative intervention decision determination method in an embodiment;

[0043] Figure 2 A flowchart of the intraoperative intervention decision determination method in an embodiment;

[0044] Figure 3 A schematic diagram of a target prompt word template in an embodiment;

[0045] Figure 4 A schematic diagram of physiological monitoring data in an embodiment;

[0046] Figure 5 A training schematic diagram of a multi-modal fusion model in an embodiment;

[0047] Figure 6 A flowchart of the dynamic summary feature vector determination step in an embodiment;

[0048] Figure 7A flowchart for determining a physiological feature information sub-set sequence in an embodiment;

[0049] Figure 8 A flowchart for determining abnormal medical information in an embodiment;

[0050] Figure 9 A schematic diagram of the overall architecture of an intraoperative intervention decision system in an embodiment;

[0051] Figure 10 A flowchart of an intraoperative intervention decision method in an embodiment;

[0052] Figure 11 A block diagram of an intraoperative intervention decision determination apparatus in an embodiment;

[0053] Figure 12 An internal structure diagram of a computer device in an embodiment. DETAILED DESCRIPTION

[0054] In order to make the purposes, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0055] The intraoperative intervention decision determination method provided by the embodiments of the present application can be applied in an application environment as shown in the figure. Figure 1 The terminal 102 communicates with the server 104 through a network. The data storage system can store data required to be processed by the server 104. The data storage system can be integrated on the server 104, or placed on a cloud or other network server. The terminal and the server can be used alone to execute the intraoperative intervention decision determination method provided in the embodiments of the present application. The terminal and the server can also be used cooperatively to execute the intraoperative intervention decision determination method provided in the embodiments of the present application. For example, the terminal 102 acquires an electronic case of a target object, a physiological feature information set and a target prompt word template; the physiological feature information set is a set composed of physiological feature information of the target object in surgery; determines a dynamic summary feature vector based on the electronic case, the physiological feature information set and the target prompt word template; determines a physiological feature vector based on the physiological feature information set; and determines an intraoperative intervention decision based on the dynamic summary feature vector and the physiological feature vector. The terminal 102 can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The Internet of Things devices can be smart speakers, smart televisions, smart air conditioners, smart vehicle devices, etc. The portable wearable devices can be smart watches, smart bracelets, head-mounted devices, etc. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers.

[0056] In some example embodiments, as shown in Figure 2 A method for intraoperative intervention decision determination is provided, and the embodiment takes the method applied to a computer device as an example for illustration, which includes steps 202 to 208.

[0057] In step 202, an electronic medical record (EMR) of a target object, a physiological characteristic information set, and a target prompt word template are obtained. The physiological characteristic information set is a set of physiological characteristic information of the target object during surgery.

[0058] The electronic medical record (EMR) refers to the electronic medical record data of a patient, including demographic information, medical history, surgical name and type, anesthesia method, preoperative examination results, and other structured and unstructured information, which can provide important background and potential risk factors. The physiological characteristic information set refers to a set of information determined according to the collected multi-dimensional physiological monitoring data during the surgery of the patient. The physiological monitoring data includes heart rate, blood pressure, blood oxygen saturation, respiratory rate, body temperature, end-tidal carbon dioxide, cardiac output, etc. The physiological monitoring data reflects the key and immediate physiological state of the patient. It can be understood that the physiological characteristic information set is composed of physiological characteristic information corresponding to multiple physiological characteristics. The physiological characteristic information refers to a data sequence obtained by sampling, missing value processing, and standardization processing of the physiological monitoring data.

[0059] The target prompt word template refers to a prompt word template for generating a dynamic summary. The target prompt word template includes but is not limited to a patient information layer, an abnormal sign layer, a knowledge graph analysis layer, and a generation constraint layer. The patient information layer includes surgical information (such as surgical name and surgical type), demographics (such as patient age, patient gender, and patient body mass index (BMI)), and anesthesia method, etc. The abnormal sign layer includes an abnormal characteristic set. The knowledge graph analysis layer includes abnormal medical information. The generation constraint layer includes constraint conditions for generating a dynamic summary (such as dynamic summary length ≤ 512 word units, containing abnormal signs, possible causes, and intervention suggestion categories). For example, as shown in Figure 3 The target prompt word template includes patient information (i.e., the patient information layer), abnormal signs (i.e., the abnormal sign layer), knowledge graph analysis (i.e., the knowledge graph analysis layer), and generation constraints (i.e., summary content requirements). The target object refers to an object for collecting the physiological characteristic information set. The target object can be a patient during surgery or a patient who needs surgery.

[0060] Exemplarily, the computer device acquires physiological monitoring data corresponding to each physiological feature in a current time period, determines, for each physiological feature, physiological feature information corresponding to the physiological feature based on the physiological monitoring data corresponding to the physiological feature, determines a physiological feature information set of the current time period based on the physiological feature information corresponding to each physiological feature, and acquires an electronic case of the target object and a target prompt word template. The current time period refers to a current time period, which can be between the start of the operation and the first use of the intervention drug, and the duration of the current time period is equal to the duration of the collection window. It can be understood that the current time period refers to a rolling time window divided by the duration of the collection window between the start of the operation and the first use of the intervention drug. For example, if the duration of the collection window is 5 minutes, the duration of the current time period is 5 minutes. The physiological monitoring data refers to data obtained by collecting the vital signs of the patient during the operation through the medical monitoring device. The physiological monitoring data can be numerical data or waveform data. For example, as shown in the schematic diagram of the physiological monitoring data Figure 4

[0061] In some exemplary embodiments, determining, for each physiological feature, physiological feature information corresponding to the physiological feature based on the physiological monitoring data corresponding to the physiological feature, and determining a physiological feature information set of the current time period based on the physiological feature information corresponding to each physiological feature, includes: for each physiological feature, sampling the physiological monitoring data corresponding to the physiological feature to obtain a sampling value sequence corresponding to the physiological feature; in the case that the sampling value sequence has missing values, performing interpolation processing on the sampling value sequence to obtain an interpolation sequence; performing standardization processing on the interpolation sequence to obtain the physiological feature information corresponding to the physiological feature (i.e., a standard value sequence); and obtaining the physiological feature information set of the current time period based on the physiological feature information corresponding to each physiological feature.

[0062] In step 204, a dynamic summary feature vector is determined based on the electronic case, the physiological feature information set, and the target prompt word template.

[0063] The dynamic summary feature vector refers to a vectorized representation of the intraoperative state of the patient obtained based on the electronic case, the physiological feature information set, and the target prompt word template via a generative large model and a text embedding model.

[0064] ​Exemplarily, the computer device determines a dynamic summary feature vector based on the electronic medical record, the set of physiological feature information, and a target prompt word template, through a generative large model and a text embedding model. The generative large model refers to a deep learning model trained on a large amount of data, which has natural language understanding and generation capabilities, and can automatically generate text results that meet semantic logic according to input prompts or context. The generative large model can be a prompt word template driven model, such as DeepSeek-V3-Chat and LLM (Large Language Model). The text embedding model refers to a model that maps input text information to vector representation. The text embedding model can be a Clinical-Longformer model, which is a pre-trained and fine-tuned model on clinical medical corpus (such as electronic medical record EMR) and is specifically used for natural language processing tasks in medical scenarios (such as clinical text understanding, medical question answering, medical record analysis).

[0065] In step 206, a physiological feature vector is determined based on the set of physiological feature information.

[0066] The physiological feature vector refers to a vectorized representation of the patient's real-time physiological state determined based on the set of physiological feature information.

[0067] Exemplarily, the computer device determines the physiological feature vector based on the set of physiological feature information through a time sequence coding unit. The time sequence coding unit (Gated Recurrent Unit, GRU for short) is a variant of recurrent neural network, which is used to process and model time series data.

[0068] In step 208, an intraoperative intervention decision is determined based on the dynamic summary feature vector and the physiological feature vector.

[0069] The intraoperative intervention decision refers to a decision result of whether drug intervention is needed and the specific drug intervention category based on patient state information during surgery. The intraoperative intervention decision includes but is not limited to intervention probability and intervention vector.

[0070] Exemplarily, the computer device determines the intraoperative intervention decision based on the dynamic summary feature vector and the physiological feature vector through a multi-modal fusion model. The multi-modal fusion model refers to an artificial intelligence model capable of unified representation and joint modeling of data of different modalities (such as text, image, voice, time series, etc.), and is used to align, correlate and represent and calculate data features of different modalities in a unified semantic space, so as to fully utilize the complementarity between various types of information and improve the expression ability and prediction performance of the model.

[0071] In some exemplary embodiments, a training schematic diagram of the multi-modal fusion model is as shown in Figure 5 The method comprises the following steps: inputting a training sample into a to-be-trained multi-modal fusion model, determining a binary classification task performance index and a multi-label classification task performance index; determining a comprehensive performance index based on the binary classification task performance index and the multi-label classification task performance index; comparing the comprehensive performance index with a current optimal performance index; if the comprehensive performance index is greater than the current optimal performance index, updating the comprehensive performance index as the current optimal performance index, determining zero as the updated training number, calculating a current error loss, updating the to-be-trained multi-modal fusion model based on the current error loss to obtain an updated to-be-trained multi-modal fusion model, and returning to execute the step of inputting the training sample into the to-be-trained multi-modal fusion model to determine the binary classification task performance index and the multi-label classification task performance index; if the comprehensive performance index is less than or equal to the current optimal performance index, adding 1 to the training number to obtain an updated training number, comparing the updated training number with a patience value; if the updated training number is less than the patience value, returning to execute the step of inputting the training sample into the to-be-trained multi-modal fusion model to determine the binary classification task performance index and the multi-label classification task performance index; and if the updated training number is equal to the patience value, determining the to-be-trained multi-modal fusion model as a trained multi-modal fusion model.

[0072] The training sample refers to a data set used to train the multi-modal fusion model to be trained. The binary classification task performance indicator refers to an indicator for measuring the performance of the binary classification prediction task. The binary classification task performance indicator can be F1-score (harmonic mean of precision and recall), AUROC (Area Under Receiver Operating Characteristic Curve), AUPRC (Area Under Precision-Recall Curve), accuracy, and precision, etc. The multi-label classification task performance indicator refers to an indicator for measuring the performance of the multi-label prediction task. The multi-label classification task performance indicator can be F1-score, AUROC, AUPRC, accuracy, and precision, etc. The multi-label classification task performance indicator can be the same as the binary classification task performance indicator, or different from the binary classification task performance indicator. The comprehensive performance indicator refers to a performance indicator obtained by fusing the binary classification task performance indicator and the multi-label classification task performance indicator, used to measure the overall performance of the multi-modal fusion model. The fusion method can include weighted average (i.e. , is the binary classification task performance indicator F1-score, is the multi-label classification task performance indicator F1-score, or , , is the binary classification task performance indicator, is the multi-label classification task performance indicator), taking the minimum value (i.e. ), wherein is the binary classification task performance indicator F1-AUROC, is the multi-label classification task performance indicator AUROC), or other combination functions to ensure balanced optimization of the model on the two tasks. The current optimal performance indicator refers to the historical best comprehensive performance indicator obtained in the previous training process during the training process. The patience value is a parameter in the early stopping strategy, used to represent the number of consecutive training cycles without performance improvement before stopping training. For example, if the patience value is equal to 10, it means that the training is terminated early when there is no performance improvement for 10 consecutive cycles. The error loss refers to the difference between the prediction result of the multi-modal fusion model and the true label. The error loss can be a dynamic balanced multi-task loss function (i.e. , wherein L is the error loss; is the binary classification task loss; is the multi-label task loss; and is the training parameter).

[0073] In the above method for determining the intraoperative intervention decision, the dynamic summary feature vector is determined based on the electronic medical record, the physiological characteristic information set and the target prompt word template, that is, the real-time time series data of the physiological characteristics and the semantic and structured expression of the electronic medical record are expressed as a dynamic summary vector representing clinical understanding, the physiological characteristic vector is determined based on the physiological characteristic information set, that is, the real-time dynamic change of the multi-dimensional physiological characteristics of the patient during the operation is captured based on the physiological characteristic information set, and the intraoperative intervention decision is determined by combining the dynamic summary feature vector and the physiological characteristic vector, so as to realize comprehensive integration and modeling of the multi-dimensional physiological characteristics and the medical record information of the patient during the operation, accurately capture the change of the real-time physiological state of the patient, avoid information one-sidedness caused by relying on artificial experience, and significantly improve the accuracy of the intraoperative intervention decision.

[0074] In some exemplary embodiments, as shown in Figure 6 Based on the electronic medical record, the physiological characteristic information set and the target prompt word template, the dynamic summary feature vector is determined, including:

[0075] In step 602, the abnormal feature is determined based on the physiological characteristic information set and the abnormal threshold.

[0076] The abnormal threshold refers to a reference value for judging whether the physiological characteristic is abnormal. When at least one value in the physiological characteristic information exceeds the abnormal threshold, the physiological characteristic is determined to be an abnormal feature. The abnormal feature refers to an abnormal physiological characteristic.

[0077] For example, the computer device compares the physiological characteristic information of the physiological characteristic with the abnormal threshold for each physiological characteristic. If at least one value in the physiological characteristic information is greater than the abnormal threshold, the physiological characteristic is determined to be an abnormal feature.

[0078] In one embodiment, the abnormal threshold τ = 2, and the abnormal feature set is as follows:

[0079] Formula (1)

[0080] A is the abnormal feature set; is a physiological characteristic, such as blood pressure, heart rate, blood oxygen saturation, etc. is a standardized value of the physiological characteristic .

[0081] In step 604, the abnormal medical information corresponding to the abnormal feature is determined based on the abnormal feature and the medical knowledge graph.

[0082] The medical knowledge graph refers to a graph structure data constructed based on medical field knowledge, including diseases, symptoms, drugs and their relationships, and is used to support the association analysis of abnormal features and clinical medical knowledge. The medical knowledge graph can be PrimeKG (Prime Knowledge Graph, medical comprehensive knowledge graph). The abnormal medical information refers to the medical explanation, potential cause or possible intervention measure obtained by associating the abnormal features with the medical knowledge graph.

[0083] The computer device determines an abnormal feature vector of the abnormal feature, obtains a node feature vector corresponding to each node in the medical knowledge graph, and determines abnormal medical information corresponding to the abnormal feature based on the abnormal feature vector and the node feature vector corresponding to each node in the medical knowledge graph.

[0084] In step 606, the physiological feature information set is divided and compressed to obtain a physiological feature information sub-set sequence.

[0085] The physiological feature information sub-set sequence refers to a sub-sequence obtained by dividing and compressing the physiological feature information set according to time periods or feature dimensions, which is used to reduce the computational complexity and retain key dynamic information.

[0086] The computer device divides the physiological feature information set based on a preset time step to obtain a plurality of physiological feature information initial sets, and the physiological feature information initial set includes physiological feature sub-information of each physiological feature. For each physiological feature in the physiological feature information initial set, the mean value, the latest value and the trend of the physiological feature sub-information corresponding to the physiological feature are determined, and the physiological feature information sub-set is determined based on the mean value, the latest value and the trend of each physiological feature. The physiological feature information sub-set sequence is determined based on a plurality of physiological feature information sub-sets. The preset time step refers to a pre-set division time length, for example, the preset time step is 1 minute. The latest value refers to the last value in the physiological feature sub-information. The trend refers to the slope of a plurality of values in the physiological feature sub-information, which can be calculated by a linear regression slope.

[0087] In some exemplary embodiments, a flowchart of the physiological feature information sub-set sequence determination is shown in Figure 7 as follows:

[0088] (1) Obtain physiological monitoring data of each physiological feature in the sliding window. The sliding window length of the sliding window is 5 minutes, and the physiological monitoring data is 5 minutes long.

[0089] (2) Signal synchronous sampling. Each physiological feature is sampled at an interval of 2 seconds to form a sampling value sequence set, which can be represented by a matrix as an original signal matrix Wherein, N is the dimension of the physiological characteristics, T is the number of collection, for example, N=66, T=160.

[0090] (3) Missing value processing. Linear interpolation is used for missing values, and the interpolation formula is:

[0091] Formula (2)

[0092] Wherein, is the interpolation value at time t; is the sampling value at time t-1; t is the time corresponding to the missing value; is the starting time of the i-th sliding window; is the ending time of the i-th sliding window; is the sampling value at time t; is the sampling value at time t. is the sampling value at time t.

[0093] (4) Z-score standardization. Each physiological characteristic is standardized to obtain a physiological characteristic information set, and the standardization formula is as follows:

[0094] Formula (3)

[0095] Formula (4)

[0096] Formula (5)

[0097] Wherein, is the standardized value of the n-th physiological characteristic at time t; is the sampling value of the n-th physiological characteristic at time t; is the mean value of the n-th physiological characteristic; is the standard deviation of the n-th physiological characteristic; is the total number of sampling values of the n-th physiological characteristic; D is the set of sampling values of the n-th physiological characteristic; d is the identification of the sampling value of the n-th physiological characteristic.

[0098] (5) Compression division. The 5-minute physiological characteristic information set is divided into 5 one-minute physiological characteristic information initial sets, each physiological characteristic in the physiological characteristic information initial set corresponds to 30 standardized values, the mean value, the latest value and the trend of the 30 standardized values are determined, based on the mean value, the latest value and the trend of each physiological characteristic, a physiological characteristic information sub-set is determined; 5 physiological characteristic information sub-sets generate a physiological characteristic information sub-set sequence in chronological order.

[0099] ​At step 608, the electronic case, the abnormal feature, the abnormal medical information, the physiological feature information sub-set sequence, and the target prompt word template are input into the generative large model to obtain a dynamic summary of the target object.

[0100] The dynamic summary refers to a text description generated by the generative large model based on the electronic case, the abnormal feature, the abnormal medical information, the physiological feature information sub-set sequence, and the target prompt word template, reflecting the patient's intraoperative state and possible intervention needs.

[0101] For example, the computer device inputs the electronic case, the abnormal feature, the abnormal medical information, the physiological feature information sub-set sequence, and the target prompt word template into the generative large model to obtain the dynamic summary of the target object.

[0102] At step 610, the dynamic summary is input into a text embedding model to obtain a dynamic summary feature vector of the dynamic summary.

[0103] For example, the computer device inputs the dynamic summary into the text embedding model to obtain the dynamic summary feature vector of the dynamic summary.

[0104] In some exemplary embodiments, the DeepSeek-V3-Chat is used to generate the dynamic summary, and the Clinical-Longformer model is used to encode the dynamic summary into a 768-dimensional dynamic summary feature vector.

[0105] Equation (6)

[0106] Where s is the dynamic summary feature vector; Longformer is the Clinical-Longformer model that converts the dynamic summary into the dynamic summary feature vector; summary is the dynamic summary generated by the DeepSeek-V3-Chat; The dimension of the dynamic summary feature vector is 768.

[0107] In this embodiment, by introducing an abnormal threshold and a medical knowledge graph, the physiological feature is recognized and the medical significance is explained, the physiological feature information is divided and compressed into a physiological feature information sub-set sequence, the key feature information is effectively extracted, then the generative large model is used to generate the dynamic summary combined with the target prompt word template, and the text embedding model is used to convert the dynamic summary into the dynamic summary feature vector, the patient's intraoperative abnormal condition is semantically and structurally expressed, and the abnormal feature is deeply associated with the medical knowledge, so that the dynamic summary feature vector more accurately represents the patient's clinical state, providing more abundant and accurate information support for subsequent determination of intraoperative intervention decisions based on the physiological feature vector, which helps to improve the accuracy and reliability of the intraoperative intervention decisions.

[0108] In some example embodiments, based on the abnormal feature and the medical knowledge graph, the abnormal medical information corresponding to the abnormal feature is determined, including:

[0109] The abnormal feature vector corresponding to the abnormal feature is determined, the node feature vector corresponding to each node in the medical knowledge graph is obtained, the similarity between the abnormal feature vector and each node feature vector is determined, the disease node, the drug node and the complication node corresponding to the abnormal feature are determined based on the similarity corresponding to each node, and the abnormal medical information corresponding to the abnormal feature is determined based on the disease node, the drug node and the complication node.

[0110] The abnormal feature vector refers to a vector representing the abnormal feature. The node feature vector refers to a vector representing the node feature. The similarity refers to a correlation measure between the abnormal feature vector and the node feature vector. Common calculation methods include cosine similarity, Euclidean distance or dot product. The higher the similarity, the stronger the semantic or clinical relevance of the abnormal feature to the node. The disease node refers to a node in the medical knowledge graph corresponding to the abnormal feature, representing a disease. The disease node reflects the possible disease background of the abnormal feature. The drug node refers to a node in the medical knowledge graph corresponding to the abnormal feature, representing a drug. The drug node reflects the possible drug intervention scheme of the abnormal feature. The complication node refers to a node in the medical knowledge graph corresponding to the abnormal feature, representing a complication node. The complication node is used to prompt the potential risk of the abnormal feature.

[0111] For example, the computer device determines the abnormal feature vector corresponding to the abnormal feature, obtains the node feature vector corresponding to each node in the medical knowledge graph, determines the similarity between the abnormal feature vector and each node feature vector, determines the disease node, the drug node and the complication node corresponding to the abnormal feature based on the similarity corresponding to each node, and determines the abnormal medical information corresponding to the abnormal feature based on the disease node, the drug node and the complication node.

[0112] In some example embodiments, the similarity between the abnormal feature vector and each node feature vector is determined using the following formula:

[0113] Formula (7)

[0114] wherein, is the similarity between the abnormal feature vector and the node feature vector of the i-th node; q is the abnormal feature vector; is the node feature vector of the i-th node in the medical knowledge graph.

[0115] In some example embodiments, the determination process of the abnormal medical information is as shown in Figure 8As shown, comprising: determining an abnormal feature; obtaining a medical knowledge graph; performing entity matching between an abnormal feature vector corresponding to the abnormal feature and a node feature vector corresponding to each node in the medical knowledge graph, to obtain a disease node, a drug node and a complication node corresponding to the abnormal feature; and determining abnormal medical information corresponding to the abnormal feature based on the disease node, the drug node and the complication node, and an association path between the disease node, the drug node and the complication node.

[0116] In this embodiment, by calculating the similarity between the abnormal feature vector and the node feature vector in the medical knowledge graph, the automatic matching of the abnormal feature and the medical semantic nodes such as diseases, drugs and complications is realized, so as to convert the low-level physiological data anomaly into clinically understandable abnormal medical information. This method not only establishes a semantic and structured association between the abnormal feature and the potential disease cause, the applicable drug intervention and the possible complications, but also realizes automatic reasoning from physiological anomaly to medical knowledge, provides accurate and semantically complete basic information for subsequent dynamic summary generation, and effectively improves the medical rationality and interpretability of the dynamic summary.

[0117] In some exemplary embodiments, based on the set of physiological feature information, a physiological feature vector is determined, comprising:

[0118] Based on a preset time step, the set of physiological feature information is divided and compressed to obtain a sequence of sets of physiological feature information; the sets of physiological feature information in the sequence of sets of physiological feature information are arranged in chronological order; a time sequence coding unit is used to sequentially code the sets of physiological feature information in the sequence of sets of physiological feature information; and until the last set of physiological feature information, a physiological feature vector is obtained.

[0119] The division refers to dividing the continuous set of physiological feature information into a plurality of data sets corresponding to time segments according to the preset time step. The compression refers to a process of data dimensionality reduction or information condensation on the divided data sets. The compression result can be a small number of key indicators such as the mean value, the latest value and the trend of each physiological feature. The set of physiological feature information refers to a data set obtained by dividing and compressing the set of physiological feature information according to the preset time step. The set of physiological feature information includes the mean value, the latest value and the trend of each physiological feature. The sequence of sets of physiological feature information refers to a sequence arranged in chronological order by a plurality of sets of physiological feature information.

[0120] Exemplarily, the computer device divides the set of physiological feature information based on a preset time step to obtain a plurality of initial sets of physiological feature information, each initial set of physiological feature information including physiological feature sub-information of each physiological feature; for each physiological feature in the initial set of physiological feature information, determining a mean value, a latest value and a trend of the physiological feature sub-information corresponding to the physiological feature, determining a sub-set of physiological feature information based on the mean value, the latest value and the trend of each physiological feature; arranging the plurality of sub-sets of physiological feature information in chronological order to obtain a sequence of sub-sets of physiological feature information; and using a time sequence coding unit to sequentially encode each sub-set of physiological feature information in the sequence of sub-sets of physiological feature information until the last sub-set of physiological feature information to obtain a physiological feature vector.

[0121] In some exemplary embodiments, the formula for encoding each sub-set of physiological feature information in the sequence of sub-sets of physiological feature information using the time sequence coding unit is as follows:

[0122] Formula (8)

[0123] Formula (9)

[0124] wherein, is a hidden state vector output by the time sequence coding unit after encoding the tthsub-set of physiological feature information; is the tthsub-set of physiological feature information; is a hidden state vector output by the time sequence coding unit after encoding the t-1thsub-set of physiological feature information, used to record context information of the sequence of sub-sets of physiological feature information; is a physiological feature vector, i.e., a hidden state vector output by the time sequence coding unit after encoding the last sub-set of physiological feature information in the sequence of sub-sets of physiological feature information; N is the total number of sub-sets of physiological feature information in the sequence of sub-sets of physiological feature information, for example, N = 5 indicates that there are 5 sub-sets of physiological feature information in the sequence of sub-sets of physiological feature information; is a physiological feature vector with a dimension of 256.

[0125] In this embodiment, by dividing and compressing the physiological feature information set into an ordered physiological feature information sub-set sequence according to a preset time step, and using a time sequence coding unit to gradually code each physiological feature information sub-set until a physiological feature vector representing the overall dynamic process is obtained, the time sequence dependence and dynamic change law of the multi-dimensional physiological characteristics of the patient during the operation can be fully captured. This method not only avoids the problems of computational complexity and information redundancy caused by directly processing long time sequences, but also extracts key state change features across time periods, so that the generated physiological feature vector is significantly improved in terms of representation accuracy and real-time performance, providing more accurate and comprehensive basis information for subsequent determination of intraoperative intervention decisions.

[0126] In some exemplary embodiments, based on the dynamic summary feature vector and the physiological feature vector, an intraoperative intervention decision is determined, including:

[0127] The dynamic summary feature vector and the physiological feature vector are subjected to two-way cross multi-head attention fusion to obtain a fusion feature vector; and based on the fusion feature vector, an intraoperative intervention decision is determined.

[0128] In some exemplary embodiments, the two-way cross multi-head attention fusion refers to establishing a correlation between the dynamic summary feature vector and the physiological feature vector in the fusion process by using a two-way interaction mode, and calculating the mutual dependence between the two types of vectors from different “attention angles” in parallel through a multi-head attention mechanism, so as to obtain a fusion feature vector that can comprehensively reflect the static medical record information and the dynamic physiological state of the patient. It can be understood that, in the two-way cross multi-head attention fusion, the dynamic summary feature vector is taken as Query (query), and the physiological feature vector is taken as Key / Value (key / value), and the physiological feature vector is taken as Query, and the dynamic summary feature vector is taken as Key / Value. After the outputs of the two attention are spliced along the feature dimension, a linear projection layer is used to fuse them into a unified fusion feature vector, realizing deep bidirectional interaction between the physiological state and the medical record semantics. The fusion feature vector refers to a feature vector obtained through the two-way cross multi-head attention fusion mechanism, which contains the clinical background and semantic information contained in the dynamic summary feature vector, as well as the real-time physiological change features contained in the physiological feature vector, and is a deep fusion result of the two types of information.

[0129] Exemplarily, the computer device performs two-way cross multi-head attention fusion on the dynamic summary feature vector and the physiological feature vector to obtain a fusion feature vector, and determines an intraoperative intervention decision based on the fusion feature vector.

[0130] In some exemplary embodiments, the dynamic summary feature vector and the physiological feature vector are subjected to two-way cross multi-head attention fusion, and the Query and Key / Value mapping relationship of one-way cross multi-head attention fusion is as follows:

[0131] Equation (10)

[0132] Equation (11)

[0133] Equation (12)

[0134] wherein Q is a Query in the attention mechanism; s is a dynamic summary feature vector; is a linear transformation weight matrix for generating the Query Q; K is a Key in the attention mechanism; is a physiological feature vector; is a linear transformation weight matrix for generating the Key K; V is a Value in the attention mechanism; is a linear transformation weight matrix for generating the Value V.

[0135] Attention weight calculation:

[0136] Equation (13)

[0137] wherein, is an attention weight of the Query in the i-th position to the Key in the j-th position, indicating a correlation weight of the i-th Query and the j-th Key; i is a position index of the Query, such as the dynamic summary feature vector s after linear transformation to obtain Q, Q may have multiple vectors Q1, Q2…, i is the position index of the vector in Q; j is a position index of the Key; k is a Key index in the normalization denominator; is a Query in the i-th position index; is a Key in the j-th position index.

[0138] The fusion feature vector is:

[0139] Equation (14)

[0140] Equation (15)

[0141] Equation (16)

[0142] wherein Z is a fusion feature vector; is to splice the two attention results together to form the final fusion feature vector; Attn is a standard attention calculation; is an attention calculated according to the Query generated according to the dynamic summary feature vector, the Key generated according to the physiological feature vector, and the Value generated according to the physiological feature vector; attention calculated for the query generated according to the physiological feature vector, the key generated according to the dynamic summary feature vector, and the value generated according to the dynamic summary feature vector; attention weight for the jth Key to the ith Query; jth Value vector; attention weight for the jth Query to the ith Key; jth Value vector.

[0143] In this embodiment, by performing two-way cross multi-head attention fusion on the dynamic summary feature vector and the physiological feature vector, deep association modeling of patient background information and real-time physiological dynamics is realized, so that the fusion feature vector not only retains the medical knowledge in the electronic case and the clinical semantic summary, but also can capture the key physiological changes of the patient in real time during the operation, so as to comprehensively consider the individual differences and dynamic state changes of the patient when generating the intraoperative intervention decision, which effectively overcomes the one-sidedness caused by single modal information, makes the intraoperative intervention decision more comprehensive, accurate and interpretable, and improves the scientificity and reliability of the intraoperative intervention suggestion.

[0144] In some exemplary embodiments, based on the fusion feature vector, an intraoperative intervention decision is determined, including:

[0145] The fusion feature vector is input into an intervention probability binary classification model to obtain an intervention probability, and the fusion feature vector is input into a multi-label model to obtain an intervention vector; the intervention vector includes the use probability of various intervention drugs and the probability of not using intervention drugs; based on the intervention probability and the intervention vector, an intraoperative intervention decision is generated.

[0146] The intervention probability binary classification model refers to a binary classification prediction model for determining whether the patient needs drug intervention, and the output result is an intervention probability, the numerical range of which is usually between 0 and 1, which is used to measure the possibility of "need intervention" and "no need intervention". The multi-label model refers to a machine learning model that can simultaneously predict multiple class labels, which is used to predict the probability of using intervention drugs and the probability of not using any drugs, and the output result is an intervention vector, which contains the use probability of intervention drugs and the probability of not using any drugs. The intervention vector refers to a multi-dimensional vector output by the multi-label model, each dimension of which corresponds to the use probability of an intervention drug or the probability of "not using intervention drugs", for example, the output intervention vector is a 7-dimensional vector, which respectively represents the use probability of 6 kinds of intervention drugs and the probability of not using intervention drugs, for example, the numerical value of each dimension in the 7-dimensional vector represents the use probability of phenylephrine, the use probability of dopamine, the use probability of prostaglandin E1, the use probability of nitroglycerin, the use probability of infusion, and the probability of not using intervention drugs.

[0147] Exemplarily, the computer device inputs the fusion feature vector into the intervention probability binary classification model to obtain an intervention probability, inputs the fusion feature vector into the multi-label model to obtain an intervention vector, and generates an intraoperative intervention decision according to the intervention probability and the intervention vector.

[0148] In an exemplary embodiment, the calculation formula of the intervention probability is as follows:

[0149] Formula (17)

[0150] wherein, p is the intervention probability; is a Sigmoid function, which maps the result to [0, 1]; is a binary classification task weight matrix; is a binary classification task bias vector.

[0151] In an exemplary embodiment, the calculation formula of the value of each dimension in the intervention vector is as follows:

[0152] Formula (18)

[0153] wherein, pdrugs is the use probability of the intervention drug, or the probability of not using the intervention drug; is a Sigmoid function, which maps the result to [0, 1]; is a multi-label classification weight matrix; is a multi-label classification bias vector.

[0154] In this embodiment, by combining the intervention probability binary classification model and the multi-label model, the overall judgment of “whether intervention is needed” and the subdivided prediction of “the use probability of the intervention drug and the probability of not using the intervention drug” are decoupled and linked, which not only can accurately judge the necessity of intervention, but also can probabilistically push for different drug categories, realizes the whole-link prediction from macro decision to micro selection, makes the intraoperative intervention decision not only have global rationality, but also have detailed personalized and differentiated recommendation, can effectively improve the refinement and clinical operability of the intervention scheme, thereby significantly enhancing the accuracy and clinical practical value of the intervention decision.

[0155] In an exemplary embodiment, the overall architecture diagram of the intraoperative intervention decision system is as shown in Figure 9 The intraoperative intervention decision system is composed of two core modules of “feature engineering module” and “multi-modal fusion and double-task prediction module”, and each module is further subdivided into multiple sub-modules. The feature engineering module is responsible for uniformly converting the heterogeneous data generated in operation into embedding vectors for model processing, including the following sub-modules: ​

[0156] Data loading unit:

[0157] ①Pull real-time vital sign stream data (i.e. physiological monitoring data) through intraoperative monitoring devices or obtain data from the intraoperative patient physiological dataset in training mode, and use the sliding window Z-score standardization algorithm to mark the abnormality of the multi-dimensional physiological features of the physiological information in the window.

[0158] ②Fill the missing values of the timing features, and obtain the multi-dimensional drug intervention label vector of the dataset in training mode.

[0159] ③Read the static information in the patient's electronic medical record.

[0160] Dynamic summary generation unit:

[0161] ①Knowledge graph retrieval subunit: map abnormal features to a medical knowledge graph to retrieve associated diseases, complications, and drug contraindications.

[0162] ②Dynamic summary generation subunit: generate a dynamic summary based on a hierarchical prompt word template using a large model.

[0163] ③Text embedding subunit: encode the dynamic summary into a 768-dimensional dynamic summary vector using a text embedding model.

[0164] The multi-modal fusion and dual-task prediction module includes the following sub-modules:

[0165] Time series encoding unit:

[0166] ①Use a GRU model to process the X×N-dimensional standardized feature matrix, outputting a 256-dimensional time series intermediate representation vector.

[0167] Dual-path cross-multi-head attention fusion unit:

[0168] ①One path cross-attention takes the dynamic summary vector as Query (query) and the physiological feature vector as Key / Value (key / value), and the other path takes the physiological feature vector as Query and the dynamic summary vector as Key / Value. Calculate two-way multi-head cross-attention and concatenate, output a 256-dimensional fusion feature vector, and realize the deep interaction between physiological sequences and clinical semantics.

[0169] Multi-task prediction head unit:

[0170] ①Binary classification prediction head: fully connected layer + Sigmoid, output intervention probability p∈[0,1].

[0171] ②Multi-label prediction head: fully connected layer + Sigmoid, output intervention vector.

[0172] Early stopping and monitoring unit:

[0173] ①Introduce learnable weights and , automatically balance the dynamic balance loss of the two tasks.

[0174] ②Calculate the comprehensive index after each round of verification, if there is no improvement for N consecutive rounds, terminate the training, and automatically roll back to the best model parameters.

[0175] Based on the above intraoperative intervention decision system, the flowchart of the intraoperative intervention decision method is as shown in Figure 10 , which includes:

[0176] Obtain the physiological monitoring data corresponding to each physiological feature in the current time period, for each physiological feature, sample the physiological monitoring data corresponding to the physiological feature to obtain a sequence of sampling values corresponding to the physiological feature; in the case of missing values in the sequence of sampling values, perform interpolation processing on the sequence of sampling values to obtain an interpolated sequence; perform standardization processing on the interpolated sequence to obtain physiological feature information corresponding to the physiological feature (i.e. a sequence of standard values); based on the physiological feature information corresponding to each physiological feature, obtain a set of physiological feature information for the current time period.

[0177] Divide the set of physiological feature information based on a preset time step to obtain a plurality of initial sets of physiological feature information, each initial set of physiological feature information including physiological feature sub-information of each physiological feature; for each physiological feature in the initial set of physiological feature information, determine the mean, latest value and trend of the physiological feature sub-information corresponding to the physiological feature, determine a sub-set of physiological feature information based on the mean, latest value and trend of each physiological feature; determine a sequence of sub-sets of physiological feature information based on the plurality of sub-sets of physiological feature information.

[0178] For each physiological feature, compare the physiological feature information of the physiological feature with an abnormal threshold, if at least one value in the physiological feature information is greater than the abnormal threshold, determine the physiological feature as an abnormal feature, and determine an abnormal feature vector of the abnormal feature.

[0179] Obtain the node feature vector corresponding to each node in the medical knowledge graph, determine the similarity between the abnormal feature vector and each node feature vector, determine the disease node, drug node and complication node corresponding to the abnormal feature based on the similarity corresponding to each node, and determine the abnormal medical information corresponding to the abnormal feature based on the disease node, drug node and complication node.

[0180] Input the electronic medical record, abnormal feature, abnormal medical information, sequence of sub-sets of physiological feature information and target prompt word template into the generative large model to obtain a dynamic summary of the target object. Input the dynamic summary into a text embedding model to obtain a dynamic summary feature vector of the dynamic summary.

[0181] The physiological feature information subsets in the sequence of physiological feature information subsets are sequentially encoded by the time sequence encoding unit until the last physiological feature information subset, to obtain a physiological feature vector.

[0182] The dynamic summary feature vector and the physiological feature vector are fused by a two-way cross multi-head attention to obtain a fusion feature vector.

[0183] The fusion feature vector is input into an intervention probability binary classification model to obtain an intervention probability, and the fusion feature vector is input into a multi-label model to obtain an intervention vector, and an intraoperative intervention decision is generated according to the intervention probability and the intervention vector.

[0184] Based on the above intraoperative intervention decision system and intraoperative intervention decision method, a simulation scenario is constructed using real surgery data, simulating a 5-minute window from the start of the surgery to the first unplanned drug intervention. The system outputs a drug intervention prediction vector based on real-time physiological data and patient electronic medical record information, including:

[0185] Intervention necessity (i.e., intervention probability): output a probability value between 0 and 1, indicating whether a drug intervention is needed at present;

[0186] Intervention category (i.e., intervention vector): the system outputs a K-dimensional probability vector (K=7, corresponding to the probability of using 6 kinds of intervention drugs and the probability of not using intervention drugs), indicating the recommended priority of each type of drug.

[0187] For example, for a 65-year-old male patient undergoing laparoscopic cholecystectomy, in terms of time series data, the mean arterial pressure has been continuously below 65 mmHg (millimeters of mercury) in the current 5 minutes, and the heart rate has risen to 110 bpm (beats per minute), then the system generates a dynamic summary as follows:

[0188] A 65-year-old male patient underwent laparoscopic cholecystectomy under general anesthesia. The current mean arterial pressure is <65 mmHg (hypotension), accompanied by compensatory tachycardia (heart rate 110 beats per minute), indicating possible inadequate perfusion or volume depletion. The increased heart rate may further increase myocardial oxygen consumption, thereby increasing the risk of heart. It is necessary to evaluate whether there is bleeding, anesthesia too deep or vasodilation, etc. While optimizing the hemodynamic state, drugs contraindicated for patients with cardiovascular disease should be avoided. Close monitoring of end-organ hypoperfusion signs is required. After processing these two data by the fusion model,

[0189] The system will output the following intraoperative intervention decision:

[0190] Intervention necessity: 0.87

[0191] Intervention category probability:

[0192] - Vasopressors (e.g. norepinephrine): 0.72

[0193] - Volume adjustment drugs (e.g. crystalloid): 0.65

[0194] - Continuous monitoring without intervention: 0.13

[0195] - Other categories: all below 0.10

[0196] The above simulation process is run on a notebook computer equipped with Intel Core i9-13900H CPU (14 cores), 32GB DDR4 memory, and Windows 11 operating system.

[0197] The simulation process verifies that the system can generate intraoperative real-time prediction and intervention recommendation through local computing environment without connecting to real operating room terminal devices, and has the feasibility of deployment in actual clinical terminals.

[0198] The verification results on VitalDB (Vital Signs Database) dataset (6388 surgeries):

[0199] Binary classification task: accuracy 97.76%, F1 0.9357, AUROC 0.9781;

[0200] Multi-label task: Micro-F1 0.964, Macro-F1 0.750;

[0201] Real-time: end-to-end delay 21 seconds, core model inference 4.2 milliseconds, meeting the intraoperative real-time decision-making requirements.

[0202] In the above intraoperative intervention decision-making method, the dynamic summary feature vector is determined by the electronic case, the physiological characteristic information set and the target prompt word template, that is, the real-time time series data of physiological characteristics and the semanticization and structurization of electronic case are expressed as a dynamic summary vector representing clinical understanding, the physiological characteristic vector is determined by the physiological characteristic information set, that is, the real-time dynamic change of multi-dimensional physiological characteristics of the patient during the operation is captured by the physiological characteristic information set, and the intraoperative intervention decision is determined by combining the dynamic summary feature vector and the physiological characteristic vector, thereby realizing the comprehensive integration and modeling of multi-dimensional physiological characteristics and patient medical record information during the operation, accurately capturing the changes of the real-time physiological state of the patient, avoiding the information partiality caused by relying on artificial experience only, and significantly improving the accuracy of intraoperative intervention decision-making.

[0203] It should be understood that although the steps in the flowcharts involved in the embodiments described above are shown in sequence according to the arrows, the steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of the steps is not strictly limited in sequence, and the steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the embodiments described above can include multiple 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 steps or stages is not necessarily sequential, but can be alternately or alternately executed with at least part of other steps or steps or stages in other steps.

[0204] Based on the same inventive concept, the embodiments of the present application also provide an intraoperative intervention decision determination device for implementing the above-mentioned intraoperative intervention decision determination method. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more intraoperative intervention decision determination device embodiments provided below can refer to the limitations of the intraoperative intervention decision determination method in the above, which will not be repeated here.

[0205] In some exemplary embodiments, as shown in Figure 11 An intraoperative intervention decision determination device is provided, comprising: an acquisition module 1102, a first vector generation module 1104, a second vector generation module 1106, and a decision generation module 1108, wherein:

[0206] The acquisition module 1102 is configured to acquire an electronic case of a target object, a set of physiological characteristic information, and a target prompt word template; the set of physiological characteristic information is a set composed of physiological characteristic information of the target object in surgery;

[0207] The first vector generation module 1104 is configured to determine a dynamic summary feature vector based on the electronic case, the set of physiological characteristic information, and the target prompt word template;

[0208] The second vector generation module 1106 is configured to determine a physiological characteristic vector based on the set of physiological characteristic information;

[0209] The decision generation module 1108 is configured to determine an intraoperative intervention decision based on the dynamic summary feature vector and the physiological characteristic vector.

[0210] In some example embodiments, the first vector generation module 1104 is further configured to: determine an abnormal feature based on the set of physiological feature information and the abnormal threshold; determine abnormal medical information corresponding to the abnormal feature based on the abnormal feature and the medical knowledge graph; divide and compress the set of physiological feature information to obtain a sequence of sets of physiological feature information; input the electronic medical record, the abnormal feature, the abnormal medical information, the sequence of sets of physiological feature information, and the target prompt word template into the generative large model to obtain a dynamic summary of the target object; and input the dynamic summary into the text embedding model to obtain a dynamic summary feature vector of the dynamic summary.

[0211] In some example embodiments, the first vector generation module 1104 is further configured to: determine an abnormal feature vector corresponding to the abnormal feature; obtain a node feature vector corresponding to each node in the medical knowledge graph, and determine a similarity between the abnormal feature vector and each node feature vector; determine a disease node, a drug node, and a complication node corresponding to the abnormal feature based on the similarity of each node; and determine the abnormal medical information corresponding to the abnormal feature based on the disease node, the drug node, and the complication node.

[0212] In some example embodiments, the second vector generation module 1106 is further configured to: divide and compress the set of physiological feature information based on a preset time step to obtain a sequence of sets of physiological feature information; arrange the sets of physiological feature information in the sequence of sets of physiological feature information in chronological order; and encode the sets of physiological feature information in the sequence of sets of physiological feature information in sequence using a time sequence encoding unit until the last set of physiological feature information to obtain a physiological feature vector.

[0213] In some example embodiments, the decision generation module 1108 is further configured to: perform two-way cross multi-head attention fusion on the dynamic summary feature vector and the physiological feature vector to obtain a fused feature vector; and determine the intraoperative intervention decision based on the fused feature vector.

[0214] In some example embodiments, the decision generation module 1108 is further configured to: input the fused feature vector into an intervention probability binary classification model to obtain an intervention probability; input the fused feature vector into a multi-label model to obtain an intervention vector; the intervention vector includes a use probability of each intervention drug and a probability of not using the intervention drug; and generate the intraoperative intervention decision based on the intervention probability and the intervention vector.

[0215] The above-described modules in the intraoperative intervention decision determination apparatus can be implemented in whole or in part by software, hardware, and combinations thereof. The above-described modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a memory in a computer device in software form to be called and executed by a processor.

[0216] In some exemplary embodiments, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 12 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for determining intraoperative intervention decisions. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0217] Those skilled in the art will understand that Figure 12 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0218] In some exemplary embodiments, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0219] In some exemplary embodiments, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above-described method embodiments.

[0220] In some exemplary embodiments, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0221] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties.

[0222] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to memory, database or other medium used in the embodiments provided by the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided by the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0223] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present disclosure.

[0224] The above-described embodiments are merely illustrative of several embodiments of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for determining intraoperative intervention decisions, characterized in that, The method includes: The electronic medical record, physiological characteristic information set, and target prompt word template of the target object are obtained; the physiological characteristic information set is a set of physiological characteristic information of the target object during surgery. Based on the electronic medical record, the set of physiological feature information, and the target prompt word template, a dynamic summary feature vector is determined; Based on the set of physiological feature information, determine the physiological feature vector; Based on the dynamic summary feature vector and the physiological feature vector, intraoperative intervention decisions are determined.

2. The method according to claim 1, characterized in that, The determination of dynamic summary feature vectors based on the electronic medical record, the set of physiological feature information, and the target cue word template includes: Based on the set of physiological feature information and the abnormal threshold, the abnormal features are determined; Based on the aforementioned abnormal features and the medical knowledge graph, the abnormal medical information corresponding to the abnormal features is determined; The set of physiological feature information is divided and compressed to obtain a sequence of physiological feature information subsets; The electronic medical record, the abnormal features, the abnormal medical information, the sequence of the subset of physiological feature information, and the target prompt word template are input into the generative large model to obtain a dynamic summary of the target object; The dynamic summary is input into the text embedding model to obtain the dynamic summary feature vector.

3. The method according to claim 2, characterized in that, The step of determining the abnormal medical information corresponding to the abnormal features based on the abnormal features and the medical knowledge graph includes: Determine the abnormal feature vector corresponding to the abnormal feature; Obtain the node feature vector corresponding to each node in the medical knowledge graph, and determine the similarity between the abnormal feature vector and each node feature vector; Based on the similarity of each node, disease nodes, drug nodes, and complication nodes corresponding to the abnormal features are determined. Based on the disease node, the drug node, and the complication node, the abnormal medical information corresponding to the abnormal feature is determined.

4. The method according to claim 1, characterized in that, The step of determining the physiological feature vector based on the set of physiological feature information includes: Based on a preset time step, the set of physiological feature information is divided and compressed to obtain a sequence of physiological feature information subsets; the physiological feature information subsets in the sequence are arranged in chronological order according to time periods. The physiological feature information subsets in the sequence of physiological feature information subsets are encoded sequentially using a time-series coding unit; The physiological feature vector is obtained by continuing until the last subset of physiological feature information is reached.

5. The method according to claim 1, characterized in that, The determination of intraoperative intervention decisions based on the dynamic summary feature vector and the physiological feature vector includes: The dynamic summary feature vector and the physiological feature vector are fused by a dual-path cross-multi-head attention to obtain a fused feature vector; Based on the fused feature vector, intraoperative intervention decisions are determined.

6. The method according to claim 5, characterized in that, The determination of intraoperative intervention decisions based on the fused feature vector includes: The fused feature vector is input into the intervention probability binary classification model to obtain the intervention probability; The fused feature vector is input into a multi-label model to obtain an intervention vector; the intervention vector includes the probability of using various intervention drugs and the probability of not using intervention drugs. Based on the intervention probability and the intervention vector, an intraoperative intervention decision is generated.

7. An intraoperative intervention decision-making device, characterized in that, The device includes: The acquisition module is used to acquire the target object's electronic medical record, physiological characteristic information set, and target prompt word template; the physiological characteristic information set is a collection of the target object's physiological characteristic information during surgery; The first vector generation module is used to determine the dynamic summary feature vector based on the electronic medical record, the set of physiological feature information, and the target prompt word template; The second vector generation module is used to determine the physiological feature vector based on the set of physiological feature information. The decision generation module is used to determine intraoperative intervention decisions based on the dynamic summary feature vector and the physiological feature vector.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images