A method and system for optimizing medical decisions based on artificial intelligence

Abstract

The application provides a method and system for optimizing medical decision based on artificial intelligence, which comprises the following steps: acquiring multi-modal medical data; the multi-modal medical data comprises structured data, unstructured data and time series data; determining an initial causal tensor based on the causal relationship among the structured data, the unstructured data and the time series data; the initial causal tensor is a three-order tensor, and the dimensions of the three-order tensor are respectively represented as the number of data modalities, the number of variables and the time step; performing low-rank tensor decomposition on the initial causal tensor to obtain a key causal pattern, and updating the initial causal tensor based on the dynamic change of the patient's condition combined with the key causal pattern to obtain a current causal tensor; and performing decision optimization on the current causal tensor based on reinforcement learning to obtain an optimized medical strategy. The application solves the problem of difficult comprehensive and accurate medical decision, and improves the accuracy of the medical strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to a method and system for optimizing medical decision based on artificial intelligence. BACKGROUND

[0002] With the development of artificial intelligence technology, machine learning methods are introduced into emergency medicine, such as medical image analysis based on deep learning and vital sign prediction based on XGBoost / LSTM.

[0003] However, most of these methods are based on statistical learning, which can only identify the correlation between variables and cannot understand the causal relationship, which may lead to incorrect decisions. For example, it is observed that "hypertensive patients have lower survival rate", but simply reducing blood pressure may not improve survival rate, which may be due to underlying diseases leading to high blood pressure and high mortality risk. At the same time, traditional AI models are usually based on static or short-term time series data, which is difficult to adapt to the rapid changes of patient conditions. Moreover, most of the current AI models are trained for single modality data, which cannot effectively fuse multi-modal information such as vital signs, laboratory examination data, medical images and medical record texts, and thus it is difficult to make comprehensive and accurate medical decisions. SUMMARY

[0004] The present application provides a method and system for optimizing medical decision based on artificial intelligence, which solves the problem of difficult comprehensive and accurate medical decision making and improves the accuracy of medical strategy.

[0005] In a first aspect, the present application provides a method for optimizing medical decision based on artificial intelligence, comprising:

[0006] obtaining multi-modal medical data; the multi-modal medical data includes structured data, unstructured data and time series data;

[0007] determining an initial causal tensor based on the causal relationship between the structured data, the unstructured data and the time series data; the initial causal tensor is a three-order tensor, and the dimensions of the three-order tensor are respectively represented as the number of data modalities, the number of variables and the time step;

[0008] performing low-rank tensor decomposition on the initial causal tensor to obtain a key causal pattern, and updating the initial causal tensor based on the dynamic changes of the patient's condition combined with the key causal pattern to obtain a current causal tensor;

[0009] performing decision optimization on the current causal tensor based on reinforcement learning to obtain an optimized medical strategy.

[0010] In a second aspect, the present invention further provides an artificial intelligence-based medical decision-making optimization system, which is applied to the artificial intelligence-based medical decision-making optimization method described in the first aspect; the artificial intelligence-based medical decision-making optimization system includes:

[0011] Acquisition unit: used for acquiring multimodal medical data; the multimodal medical data includes structured data, unstructured data and time series data;

[0012] A causal tensor construction unit is configured to determine an initial causal tensor based on the causal relationship between the structured data, the unstructured data, and the time series data; the initial causal tensor is a third-order tensor, and the dimensions of the third-order tensor are respectively represented by the number of data modes, the number of variables, and the time step;

[0013] A state updating unit is configured to perform low-rank tensor decomposition on the initial causal tensor to obtain a key causal pattern, and update the initial causal tensor based on the dynamic changes of the patient's condition in combination with the key causal pattern to obtain a current causal tensor;

[0014] Strategy optimization unit: used to perform decision optimization on the current causal tensor based on reinforcement learning to obtain an optimized medical strategy.

[0015] In a third aspect, the present invention also provides an electronic device comprising: a memory for storing a computer software program; and a processor for reading and executing the computer software program, thereby implementing any of the above-described methods for optimizing medical decisions based on artificial intelligence.

[0016] In a fourth aspect, the present invention also provides a non-transitory computer-readable storage medium, in which a computer software program is stored. When the computer software program is executed by a processor, it implements any of the above-mentioned methods for optimizing medical decisions based on artificial intelligence.

[0017] In a fifth aspect, the present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-described methods for optimizing medical decisions based on artificial intelligence.

[0018] The method for optimizing medical decisions based on artificial intelligence provided by an embodiment of the present invention constructs an initial causal tensor by using structured data, unstructured data and time series data in multimodal medical data. It can not only effectively integrate the causal relationship between multimodal data, clearly display the complex causal relationship between data, and avoid the problem of incorrect decision-making caused by the inability to understand the causal relationship, but also provide a more accurate basis for the initial causal tensor in the subsequent optimization of medical decisions through the fusion of multimodal data. Afterwards, by performing low-rank tensor decomposition on the initial causal tensor and combining it with the dynamic changes of the patient's condition, a current causal tensor that can reflect the condition in real time is obtained, and the dynamic update of the initial causal tensor is realized to adapt to the dynamic changes of the patient's condition, and further provide a more accurate basis for subsequent medical decisions. Finally, the dynamically updated current causal tensor is combined with reinforcement learning to obtain an optimized medical strategy, which solves the problem of difficulty in making comprehensive and accurate medical decisions and improves the accuracy of the medical strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 1 is a flow chart of a method for optimizing medical decision-making based on artificial intelligence provided by an embodiment of the present invention;

[0020] Figure 2 This is a schematic diagram of the structure of a medical decision-making system based on artificial intelligence optimization provided by an embodiment of the present invention;

[0021] Figure 3 An embodiment diagram of an electronic device provided by an embodiment of the present invention;

[0022] Figure 4 An embodiment diagram of a computer-readable storage medium provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0024] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the specified features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0025] In the description of the present application, the term "for example" is used to mean "serving as an example, instance, or illustration." Any embodiment described as "for example" in the present application is not necessarily to be construed as preferred or advantageous over other embodiments. The following description is presented to enable any person skilled in the art to make and use the application. In the following description, for the purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present application. It will be apparent, however, to one skilled in the art that the present application can be practiced without using these specific details. In other instances, well-known structures and processes are not elaborated upon in order to avoid obscuring the subject matter of the present application. Thus, the present application is not intended to be limited by the embodiments shown, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

[0026] Referring to Figure 1 , Figure 1 is a flowchart of the method for optimizing medical decision based on artificial intelligence provided by the present application. The execution subject of the method for optimizing medical decision based on artificial intelligence in the embodiments of the present application is a medical decision system. Therefore, the method for optimizing medical decision based on artificial intelligence comprises the following steps:

[0027] Step 10, acquiring multi-modal medical data; the multi-modal medical data comprises structured data, unstructured data and time series data.

[0028] Optionally, the medical decision system collects various medical data of the patient through various channels and technical means, and meanwhile, the medical data can also be directly input into the medical decision system by the corresponding medical staff. Among them, for the structured data, the medical decision system can extract the basic information (such as name, age, gender, etc.), diagnosis results (such as disease name, diagnosis time, etc.), vital sign data (such as heart rate, blood pressure, blood oxygen saturation, etc.) and inspection results (such as blood sugar, blood lipid, blood routine, etc.) and other data with definite format and meaning of the patient from the hospital information system, and when acquiring the structured data, the medical decision system can also access the electronic medical record system, the hospital information system (HIS), the laboratory information management system (LIS) and other databases storing a large amount of structured medical data. Data exchange and sharing are carried out through data interface.

[0029] Furthermore, for unstructured data, the medical decision-making system first accesses the hospital's file servers or databases to obtain data without a predefined structure, such as medical imaging data (such as CT, MRI, and X-rays) and medical records. Simultaneously, when acquiring medical imaging data or case text data, it uses image recognition and processing technologies to analyze and extract key information. When acquiring medical records, it uses natural language processing to parse doctors' handwritten and electronic medical records to extract key information. For time series data, the medical decision-making system can connect to medical devices (such as monitors, electrocardiograms, and blood glucose meters) to collect patients' vital signs at regular or periodic intervals. The medical decision-making system can also access the hospital's information system to obtain vital sign data stored in time series format. This enables comprehensive collection of multimodal medical data that covers all aspects of the patient's condition, providing a rich and accurate information foundation for subsequent causal analysis and decision-making.

[0030] In one embodiment, real-time data of a sepsis patient is obtained as an example. The multimodal medical data includes: vital signs (numeric data): such as blood pressure, heart rate, and blood oxygen saturation; laboratory examination data (numeric data): such as lactate and inflammatory markers (CRP); imaging data (CT / MRI): such as lung infection status (score 1-5); medical record text data: patient complaints and medical order records (processed using NLP). These data constitute the patient's current real-time status.

[0031] Step 20: Determine an initial causal tensor based on the causal relationship between the structured data, the unstructured data, and the time series data; the initial causal tensor is a third-order tensor, and the dimensions of the third-order tensor are respectively represented by the number of data modes, the number of variables, and the time step.

[0032] Optionally, after acquiring the structured data, the unstructured data, and the time series data, the medical decision-making system first preprocesses the data so that the various data can be subsequently processed in a standardized form. A causal relationship learning algorithm is then used to analyze the causal relationships between the different modal data and effectively integrate the causal relationships between the different modal data to obtain an initial causal tensor. The specific process is described in steps 201 to 205. Representing the causal relationships of multimodal data in tensor form can intuitively demonstrate the complex causal relationships between different modalities, variables, and time steps, providing a clear structural framework for subsequent analysis and decision-making.

[0033] Furthermore, the three dimensions of the third-order tensor represent the number of data modalities (such as vital signs, imaging, and laboratory test data, i.e., structured data, unstructured data, and time series data), the number of variables (such as blood pressure, heart rate, and blood sugar), and the time step (used to track disease progression). This third-order structure integrates relevant factors from multiple data modalities, rather than making erroneous judgments based solely on a single data point. Simultaneously, the fusion of multimodal data provides richer information for the initial causal tensor, allowing data from different modalities to describe the patient's condition from different perspectives. These data, such as laboratory indicators in structured data, textual descriptions of medical records in unstructured data, and changes in vital signs in time series data, can be integrated into the initial causal tensor, providing a more accurate basis for subsequent optimized medical decision-making. Because this more comprehensive patient information is included, various factors can be more precisely considered when formulating medical strategies, improving decision accuracy.

[0034] Step 30, performing low-rank tensor decomposition on the initial causal tensor to obtain a key causal pattern, and updating the initial causal tensor based on the dynamic changes of the patient's condition in combination with the key causal pattern to obtain a current causal tensor.

[0035] Optionally, after obtaining the initial causal tensor, the medical decision-making system extracts the key causal model by performing low-rank tensor decomposition on the initial causal tensor to simplify the data structure, thereby reducing computational complexity. This process is described in steps 3011 through 3014.

[0036] Furthermore, after the medical decision-making system obtains the key causal pattern through low-rank tensor decomposition, since the patient's condition changes at any time, in order to improve the timeliness and accuracy of the current causal tensor, the key causal pattern is combined with the dynamic changes of the patient's condition. Then, when a certain vital sign of the patient changes, the initial causal tensor can be updated in time. The specific process is as described in steps 3021 to 3025, so that the updated current causal tensor can reflect the changes in the patient's condition in real time, providing accurate data basis for subsequent optimization of medical decisions.

[0037] Step 40: Optimize the decision of the current causal tensor based on reinforcement learning to obtain an optimized medical strategy.

[0038] Optionally, when making medical decisions, the medical decision-making system employs reinforcement learning, as traditional decision-making methods are difficult to optimize in complex medical environments and cannot effectively handle the causal relationships of multimodal, high-dimensional data. This allows the system to identify optimal strategies through trial and error in complex, dynamic environments. Furthermore, the system can learn the complex causal relationships of multimodal, high-dimensional data. Combined with the current design of the causal tensor and reward function, this system can make decisions more aligned with actual medical needs, improving the quality and effectiveness of medical decisions. This is described in detail in steps 401-404. Furthermore, reinforcement learning uses a Markov decision process to redefine state transition probabilities, reward functions, and discount factors. This makes the decision-making process more systematic and organized, and enhances the interpretability of decisions.

[0039] Furthermore, after the medical decision-making system has optimized the medical strategy, doctors can evaluate and adjust the strategy based on their own professional knowledge and clinical experience. If the decision involves drug treatment, the drug's applicability and side effects will be considered; if it is surgical treatment, the surgical risks and feasibility will be assessed. Furthermore, during the implementation of the optimized medical strategy, it is necessary to continuously monitor the patient's condition changes, physiological indicators, etc. The effectiveness of the strategy is determined based on the monitoring results. If it is ineffective or ineffective, the current causal tensor is re-evaluated, and the parameters of the decision optimization process are adjusted or the decision optimization is re-performed to ensure the treatment effect. The implementation process and results of the optimized medical decision are recorded and analyzed to provide a reference for subsequent decisions on similar cases, thereby continuously improving the medical decision-making model and improving the overall level of medical decision-making.

[0040] The embodiment of the present invention constructs an initial causal tensor through structured data, unstructured data and time series data in multimodal medical data, which can not only effectively integrate the causal relationship between multimodal data, clearly display the complex causal relationship between data, and avoid the problem of incorrect decision-making caused by the inability to understand the causal relationship, but also provide a more accurate basis for the initial causal tensor in the subsequent optimization of medical decisions through the fusion of multimodal data. Afterwards, by performing low-rank tensor decomposition on the initial causal tensor and combining it with the dynamic changes of the patient's condition, a current causal tensor that can reflect the condition in real time is obtained, and the dynamic update of the initial causal tensor is realized to adapt to the dynamic changes of the patient's condition, and further provide a more accurate basis for subsequent medical decisions. Finally, the dynamically updated current causal tensor is combined with reinforcement learning to obtain an optimized medical strategy, which solves the problem of difficulty in making comprehensive and accurate medical decisions and improves the accuracy of the medical strategy.

[0041] In one embodiment, steps 201 to 205 are described as follows:

[0042] Step 201 : Preprocessing and feature extraction are performed on the structured data, the unstructured data, and the time series data respectively to obtain structured features, unstructured features, and time series features.

[0043] Optionally, after the medical decision-making system obtains the structured data, the unstructured data and the time series data, the original multimodal data cannot be directly used for causal analysis due to problems such as inconsistent format, high noise, and unclear features. Therefore, through preprocessing and feature extraction, the potential value of the data can be explored and the accuracy and efficiency of causal analysis can be improved.

[0044] Furthermore, the medical decision-making system preprocesses and extracts features from structured data. Data cleaning is first performed to remove erroneous and missing values, followed by normalization (e.g., Z-score standardization) to map numerical values ​​to specific intervals. Categorical data is then encoded using techniques such as one-hot encoding to extract features such as patient information and test results. For unstructured data, if the data is text, word segmentation and part-of-speech tagging are performed first. Word embedding models (e.g., Word2Vec, BERT) are then used to convert the text into numerical vectors and extract semantic features. If the data is medical imaging, image preprocessing techniques such as grayscale transformation and filtering are applied. Convolutional neural networks (CNNs) are then used to extract image feature vectors, such as edge and texture features. For time series data, non-stationary sequences are converted to stationary sequences using differencing methods. Trend and cyclical features are then extracted. For example, for electrocardiogram (ECG) time series data, stationarization is achieved by calculating the difference between adjacent moments, and cyclical features are then extracted using Fourier transforms.

[0045] In one embodiment, taking a sepsis patient as an example, the vital signs and laboratory test data in the structured data are as follows:

[0046] Variable Value Explanation Blood pressure (BP) 85 mmHg Low blood pressure Heart rate (HR) 120 bpm Tachycardia Blood oxygen saturation (Sp02) 90％ Low blood oxygen Lactate 5.5 mmol / L Tissue hypoxia Inflammatory marker (CRP) 100 mg / L Severe inflammation

[0047] ; The image data (CT) of unstructured data is as follows:

[0048] Variable Explanation Pulmonary infection Significant infection Pulmonary edema Mild pulmonary edema

[0049] ; Using neural network f cv Extract image features and convert CT data into numerical vectors:

[0050] X CT =[0.8,0.4,0.2]

[0051] ; Among them: 0.8 represents a higher lung infection score; 0.4 represents the degree of pulmonary edema; 0.2 represents the uncertainty weight of the imaging modality.

[0052] For time series data (medical record text data), the original text description is: "A 60-year-old patient with a history of diabetes developed a fever of 39°C, decreased blood pressure, and increased lactate. Sepsis is suspected, and immediate antibiotics and pressor medication are recommended."

[0053] Using the neural network f nlp Extract text features and convert text data into numerical vectors:

[0054] X Text =[1,0,1,0,0.8]

[0055] ; where: 1 represents diabetes (extracted from medical records); 0 represents no history of hypertension; 1 represents description related to sepsis; 0 represents no mention of heart disease; 0.8 represents the importance of antibiotic intervention recommendation.

[0056] Step 202 : constructing an initial Bayesian network by taking the structured features, the unstructured features, and the time series features as nodes and the independent relationships between the nodes as edges.

[0057] Optionally, the medical decision-making system uses the extracted structured features, unstructured features, and time series features as nodes of the Bayesian network. Based on domain knowledge and preliminary analysis of the data, it determines the potential independent relationships between the nodes and uses this relationship as an edge to connect the corresponding nodes. For example, in a medical scenario, it is known that blood pressure (structured feature) is associated with cardiac output (structured feature). An edge from the blood pressure node to the cardiac output node can be constructed in the Bayesian network. During the construction process, the direction of the edge represents the possible direction of the causal relationship, and medical knowledge and prior research results can be used as reference. Finally, the initial Bayesian network is constructed, which can intuitively represent the causal relationship structure between multimodal data features and provide a clear framework for subsequent causal analysis.

[0058] Step 203 : performing information entropy processing on each node in the initial Bayesian network based on a preset information entropy function to obtain node information entropy.

[0059] Optionally, the information entropy function stored in the medical decision-making system is:

[0060]

[0061] ; Where X0 represents the random variable represented by the node; x i Expressed as the value of X0, p(x i ) represents the value x i The probability of n is the total number of values ​​of the random variable X0. The information entropy function is used to process the information entropy of each node in the initial Bayesian network. That is, for each node X0, the probability of its different values ​​p(xi ), then substitute the information entropy formula into the node information entropy. For example, for a node indicating whether a patient is infected (yes / no), the ratio of infected to uninfected samples to the total number of samples is used as the probability to calculate the information entropy of the node. Information entropy reflects the uncertainty of a node; higher entropy values ​​indicate greater uncertainty. This allows us to quantify the uncertainty of each node and understand the information content of each feature in the causal relationship. This helps in subsequent screening of important nodes and analysis of the strength of causal relationships, highlighting nodes that contribute significantly to the causal relationship and improving the specificity of causal analysis.

[0062] Step 204 : Analyze and process the nodes at both ends of each edge in the initial Bayesian network based on a preset mutual information function to obtain node mutual information between the nodes at both ends of each edge.

[0063] Optionally, the mutual information function stored in the medical decision-making system is:

[0064]

[0065] ; where X Δ and Y Δ Represents the nodes at both ends of each edge, P(x,y) is represented by X Δ and Y Δ The joint probability of X is P(x) and P(y) is Δ With Y Δ Their respective marginal probability distributions, when X Δ and Y Δ When they are independent of each other, P(x,y)=P(x)P(y), then I(X Δ ; Y Δ )=0; when X Δ and Y Δ When the two nodes are completely correlated, the mutual information reaches its maximum value. Therefore, in the medical decision-making system, for each edge connecting two nodes X Δ and Y Δ , first calculate their joint probability P(x,y) and their respective probabilities P(x) and P(y), and then substitute them into the mutual information function to calculate the node mutual information. For example, for the blood pressure node X Δ and heart rate node Y Δ The joint probability is calculated by counting the proportion of samples with different blood pressure and heart rate value combinations to the total number of samples. The probabilities of different blood pressure and heart rate values ​​are then calculated separately, and the mutual information between them is calculated. Mutual information measures the degree of dependence between two nodes. A greater mutual information indicates a stronger connection between the two nodes. This provides an important basis for determining edge weights.

[0066] Step 205 : Adjust the weight of each edge in the initial Bayesian network based on the node information entropy and the node mutual information, determine the current Bayesian network, and perform tensor transformation on the current Bayesian network to obtain an initial causal tensor.

[0067] Optionally, the medical decision-making system presets a weight adjustment function to adjust the weights of the edges in the Bayesian network based on the node information entropy obtained in step 203 and the node mutual information obtained in step 204, wherein the weight adjustment function is:

[0068]

[0069] ; where ω ij Represented as a node and The weight of the edge between Indicates a point and The mutual information between nodes, and Represents the information entropy of each node. The current Bayesian network is determined by calculating the weight of each edge in the initial Bayesian network, and then converted into an initial causal tensor. According to the structure and node characteristics of the Bayesian network, the information of nodes and edges is mapped into a third-order tensor according to the dimensional requirements of the number of data modalities, the number of variables, and the time step. Assuming that the data modalities are structured data, unstructured data, and time series data (M0=3), the number of variables is N0, and the time step is P0, a tensor of M0*N0*P0 is constructed, and the causal relationship information in the Bayesian network is filled in the corresponding position. Ultimately, by adjusting the weight of the edge through the comprehensive node information entropy and mutual information, the Bayesian network can more accurately reflect the strength of the causal relationship between multimodal data. Then, the Bayesian network is converted into a causal tensor, which can use the mathematical properties of the tensor for more efficient calculation and analysis, providing convenience for subsequent causal reasoning and decision optimization.

[0070] In one embodiment, taking the real-time data of sepsis patients as an example, the initial causal tensor obtained is:

[0071]

[0072] Where M0 represents the number of modalities, including numerical values, images, and text. In this example, M0 = 3. N0 represents the number of variables, including the number of clinical variables that doctors are concerned about, such as blood pressure, heart rate, and lung infection score. Here, N0 = 5. P0 represents the time step, which records the causal relationship over time. r Represented as the set of real numbers.

[0073] The embodiment of the present invention integrates multimodal medical data, from data preprocessing and feature extraction to the construction of Bayesian networks, and then to the calculation of information entropy and mutual information to adjust the weights of edges, and finally converts it into an initial causal tensor, forming a complete causal relationship modeling process. It can fully explore the complex causal relationships between multimodal data and improve the accuracy and reliability of causal relationship representation. Converting the Bayesian network into a causal tensor provides an efficient data structure for subsequent causal reasoning, decision optimization, etc., which helps to achieve more accurate medical decisions and improve the quality of medical services.

[0074] In one embodiment, steps 3011 to 3014 are described as follows:

[0075] Step 3011: perform feature division on the initial causal tensor to obtain multiple local features, and perform feature analysis on each of the local features to obtain a local feature analysis result.

[0076] Optionally, the medical decision system targets the initial causal tensor. Since the initial causal tensor is a third-order tensor, its dimensions represent the number of data modalities, the number of variables, and the time step. In order to analyze the information in the tensor more carefully, a partitioning strategy based on window sliding is adopted. Assume that the initial causal tensor is Where m0 represents the data modality, n0 represents the number of variables, and p0 represents the time step. Let the window size be s c *p c *q c (s c ≤m0,p c ≤n0,q c ≤p0), starting from the starting position of the tensor, the sliding window operation is performed with a step size of 1. The area covered by each window is a local feature. For example, at the i-th sliding, the local feature obtained is Its elements are represented as where a t ∈{1,…,s c}, b t ∈{1,…,p c}, c t ∈{1,…,q c}, and satisfy the corresponding relationship between i and the starting position of the window, such as i=(r c -1)*p c *q c +(c t -1)*p c +b t ,(r cThe initial causal tensor is divided into multiple overlapping or non-overlapping local features. This allows large-scale tensor data to be divided into multiple smaller local features, reducing the complexity of data processing. Local features focus more on information in specific areas of the tensor, which is conducive to discovering local causal patterns and improving the precision of analysis.

[0077] Furthermore, after the medical decision system determines multiple local features, for each local feature T ti , the principal component analysis (PCA) method is used for feature analysis. First, the covariance matrix C of the local feature tensor is calculated i .for Expands into matrix form (i.e., the modal dimension is the row and the other two dimensions are expanded into columns), the covariance matrix Then, the covariance matrix C i Perform eigenvalue decomposition, that is Among them U i Expressed as the eigenvector matrix, Λ i is represented as an eigenvalue diagonal matrix with eigenvalue λ ij (j=1,…,s c ) Arrange from largest to smallest. Select the first k c principal components (k c c ), the eigenvectors and eigenvalues ​​corresponding to the principal components constitute the local feature analysis results. Through principal component analysis, we can extract the most important change direction and degree in the local features. This can remove redundant information in the local features and retain the main feature information.

[0078] Step 3012: Adjust the kernel function parameters based on the local feature analysis result to obtain an adjusted kernel function.

[0079] Optionally, the kernel function in the medical decision system is a Gaussian kernel function Where x and y are both input vectors, σ is the bandwidth parameter of the kernel function, and the formula for adjusting the parameter σ is designed based on the principal component eigenvalues ​​obtained from local feature analysis. Let λ min and λ max are the maximum and minimum eigenvalues ​​obtained by principal component analysis of local features, respectively. The adjustment formula is: where σ old Expressed as the bandwidth parameter of the initial kernel function, σ new ​The adjusted bandwidth parameter is then expressed as the adjusted bandwidth parameter. The kernel function bandwidth is then adjusted based on the adjustment formula and the degree of change in local features to better adapt the kernel function to the distribution of local data. Adjusting the kernel function parameters based on the characteristics of local features can enhance the kernel function's ability to fit local data and more accurately measure similarity between data.

[0080] Step 3013: Perform low-rank decomposition on the initial causal tensor based on the adjusted kernel function to obtain a low-rank tensor factor.

[0081] Optionally, the medical decision system uses the adjusted kernel function to perform low-rank decomposition on the initial causal tensor. A low-rank decomposition method based on kernel function is adopted, such as the idea of ​​kernel principal component analysis (KPCA). The elements in the causal tensor are regarded as sample points, and the data are mapped to a high-dimensional space through the kernel function, and low-rank decomposition is performed in the high-dimensional space. Assume that the initial causal tensor is T t , after kernel function mapping, find the orthogonal basis u in high-dimensional space k 、v k and w k (k=1,2,…,K - , K - is expressed as the rank of the low-rank decomposition), so that T t It can be approximately expressed as where λ k Expressed as the corresponding weight value, u k ,v k ,w k are respectively expressed as low-rank tensor factors, Expressed as an outer product operation. Through iterative optimization algorithms, such as alternating least squares (ALS): The goal is to minimize the original causal tensor T t The error of the causal model after decomposition, variables U, V, W are the causal relationship matrices to be optimized, and u is continuously updated k ,λ k 、v k and w k , until the convergence condition is met, and the low-rank decomposition result, i.e., the low-tensor tensor factor, is obtained. The low-rank decomposition based on the kernel function can better capture the nonlinear relationship of data in high-dimensional space. Compared with traditional low-rank decomposition methods, it can more accurately extract the key causal patterns in complex causal tensors.

[0082] Step 3014: Error verification is performed on the low-rank tensor factor and the initial causal tensor, and analysis and judgment are performed based on the error verification results to obtain the key causal pattern.

[0083] Optionally, the medical decision system calculates the error between the tensor reconstructed by the low-tensor tensor factor and the initial causal tensor, and uses the Frobenius norm to measure the error, which is: Where ||·||F represents the Frobenius norm. An error threshold ∈ is set. If the calculated error E0 is less than ∈, the low-rank decomposition result is considered to have well preserved the key information of the initial causal tensor, and the causal pattern corresponding to the current low-rank causal tensor factor is the key causal pattern. If E0 is greater than ∈, the kernel function parameters need to be readjusted, and the low-rank decomposition and error verification process is repeated until the error requirement is met. Error verification can quantitatively evaluate the accuracy of the low-rank decomposition results, ensuring that the extracted key causal patterns effectively represent the important information in the initial causal tensor.

[0084] The embodiment of the present invention systematically extracts key causal patterns through a series of steps, including feature partitioning of the initial causal tensor, kernel function parameter adjustment, low-rank decomposition, and error verification. Each step is closely linked. Feature partitioning and local analysis provide a basis for kernel function parameter adjustment. The adjusted kernel function improves the effect of low-rank decomposition, and error verification ensures the accuracy of key causal patterns. The overall method can effectively process complex causal tensors and accurately extract key causal patterns, providing strong support for subsequent medical decision optimization, causal relationship analysis, etc., and improving the reliability and effectiveness of analysis and decision-making based on causal tensors.

[0085] In one embodiment, steps 3021 to 3025 are described as follows:

[0086] Step 3021: Analyze the causal flow between the variables in the key causal model based on the dynamic change state of the disease to obtain a causal flow analysis result.

[0087] Optionally, the medical decision-making system continuously collects multimodal medical data at continuous time points. Multimodal medical data includes structured data (such as vital sign values), unstructured data (such as doctor's medical record text), and time series data (such as the change of specific indicators over time). Multimodal medical data is associated with key causal patterns and analyzed, such as using a causal analysis algorithm (an extension of the Granger causality test). Assuming variables X0 and Y0, in the time series t0, considering the impact of the past value of X0 on the current value of Y0, the formula is: in represents the value of variable Y0 at time t0, represents the value of variable X0 at time t0-j, α i and β j are expressed as coefficients, γ and δ are expressed as lag orders, Expressed as an error term. By testing β jWhether X0 is significantly different from zero determines whether there is a causal influence on Y0. This analysis is performed for each pair of variables in the key causal model to determine the causal flow and generate the causal flow analysis results. Clarifying the causal flow between variables helps understand the underlying mechanisms of disease progression and provides a clear logical path for subsequent medical decision-making. It can uncover causal relationships hidden in the data and prevent erroneous decisions based solely on superficial correlations.

[0088] Step 3022: Determine a time window for causal flow analysis based on the speed and trend of the dynamic changes in the disease condition.

[0089] Optionally, the medical decision-making system quantitatively analyzes the speed and trend of the dynamic changes of the disease, mainly by calculating the rate of change of key indicators related to the disease (such as body temperature, blood pressure, etc.) over time to determine the speed of disease change. Let the key indicator sequence be Z(t μ ), the change rate formula is Where Δt represents the time interval. By fitting the time series data, such as using a polynomial fit Z(t μ )=a n t μ m +…+a1t μ +a0, judge the change trend based on the slope and curvature of the fitting curve. If the slope of the curve increases and the curvature is positive, it indicates that the condition is deteriorating rapidly; if the slope decreases and the curvature is negative, the condition is improving significantly. According to the change speed and trend, the time window size T is adaptively determined. c For example, when the condition changes dramatically, the time window can be shortened (e.g. T c0 The time window is represented as the initial setting) to more accurately capture short-term causal relationships; when the disease changes slowly, the time window is appropriately increased. Implementing adaptive adjustment of the time window can more accurately capture the causal relationship during the disease change process.

[0090] Step 3023: Perform strength calculation processing on the causal flow analysis result based on the transfer entropy function to obtain the causal flow strength.

[0091] Optionally, the medical decision-making system presets a transfer entropy function to measure the degree of information transfer from one variable to another variable, that is, to reflect the strength of the causal relationship. Therefore, for the two variables in the key causal model and The transfer formula is: where x t-1 and y t-1 Represented as a variable and The value at time t-1, y t Represented as a variable The value at time t, P(xt-1 ,y t-1 ,y t ) is expressed as a joint probability distribution, P(y t |x t-1 ,y t-1 ) is expressed as a conditional probability distribution, that is, x t-1 and y t-1 Known time y t The probability of occurrence, P(y t |y t-1 ) is represented by y t-1 When y is known t The probability of occurrence is calculated. Based on the causal flow direction determined in the causal flow analysis results, the transfer entropy between each pair of causal variables is calculated to obtain the causal flow strength. This makes the description of causal relationships more scientific and precise. It also distinguishes the importance of different causal relationships, providing more valuable information for medical decision-making.

[0092] Step 3024: Update the initial causal tensor based on the causal flow intensity and the time window to obtain an updated causal tensor.

[0093] Optionally, the medical decision system updates the initial causal tensor according to the calculated causal flow intensity and the determined time window. t , whose element T t (M0, N0, P0), where M0 represents the modal number; N0 represents the number of variables; P0 represents the time step. and time window T c , the initial causal tensor is updated. For tensor elements corresponding to causal relationships with greater causal flow intensity and within the time window, a greater weight is given. The update formula is: where α t Expressed as a weight coefficient (the value range is between 0-1, adjusted according to the stability and importance of the disease changes), Expressed as the maximum value among all causal flow intensities. Using the update formula, the causal flow intensities are incorporated into the initial causal tensor, resulting in an updated causal tensor. This update method effectively incorporates information about the dynamic changes in the disease state into the causal tensor, enabling it to more accurately reflect the causal relationships between variables under the current disease state. By adjusting the weight coefficients, the proportion of historical causal information and new causal flow strength information can be flexibly controlled to adapt to varying disease state changes.

[0094] Step 3025: Perform low-rank reconstruction on the updated causal tensor to obtain the current causal tensor.

[0095] Optionally, when the medical decision-making system receives the updated causal tensor, since the updated causal tensor may still have high dimensionality and computational complexity, low-rank reconstruction can be performed on it to reduce the dimensionality of the causal tensor and reduce computational complexity while preserving key causal information. This is described in detail in steps 30251 to 30254.

[0096] The embodiments of the present invention closely integrate the dynamic changes of a patient's condition, forming a complete causal tensor update system from analyzing causal flow directions, determining time windows, calculating causal flow strength, to updating causal tensors and performing low-rank reconstruction. This system can accurately reflect the causal relationships between variables under changing conditions in real time, providing a reliable basis for medical decision-making. Furthermore, through quantitative analysis and reasonable updating and reconstruction methods, it improves the accuracy and efficiency of causal relationship analysis, helping doctors formulate treatment plans more scientifically and improving the quality of medical services.

[0097] In one embodiment, steps 30251 to 30254 are described as follows:

[0098] Step 30251: Perform singular value decomposition on the updated causal tensor to obtain a first singular value matrix.

[0099] Optionally, after the medical decision system receives the updated causal tensor, since the updated causal tensor incorporates the dynamic change information of the patient's condition and has a high dimension and complex structure, it performs singular value decomposition (SVD) on it and assumes that the updated causal tensor is T tnew , dimension is M g *N g *P g (M g Expressed as the number of data modes, N g Expressed as the number of variables, P g In order to facilitate the singular value decomposition, it is first expanded into a matrix form according to certain rules. Assume that the expanded matrix is ​​Z i (Size is M g N g *P g ). According to the singular value decomposition principle, there exists an orthogonal matrix A diagonal matrix containing singular values ​​Σ(M g N g *P g ) and orthogonal matrices Make The diagonal elements of Σ These are the singular values. Arrange the singular values ​​in descending order to form the first singular value matrix Σ1. Singular value decomposition can decompose complex causal tensors into several simple matrix product forms, making the tensor structure clearer.

[0100] Step 30252: Update the singular values ​​in the first singular value matrix based on a preset singular value threshold to obtain a second singular value matrix.

[0101] Optionally, a singular value threshold is preset in the medical decision system Singular value threshold The setting of takes into account many factors, such as the noise level of the data, the importance of causal relationships, etc., to achieve the purpose of removing small singular values ​​by setting a threshold, thus achieving noise reduction and de-redundancy processing of the data. The retained larger singular values ​​and their corresponding singular vectors can more centrally represent the key information in the updated causal tensor, thereby improving the quality and effectiveness of the data. Therefore, for the first singular value matrix Σ1, the second singular value matrix Σ2 is constructed. When When, Σ2(i,i)=Σ1(i,i); when When Σ2(i,i)=0, we obtain the second singular value matrix Σ2. The purpose is to remove singular values ​​that have a small contribution to causal relationships. Small singular values ​​correspond to noise or unimportant information in the data. By removing small singular values, we can further highlight key causal patterns.

[0102] Step 30253: Construct geometric constraints based on the spatial geometric characteristics of the multimodal medical data.

[0103] Optionally, the medical decision-making system receives multimodal medical data including structured data, unstructured data, and time series data, each of which has unique spatial geometric characteristics. Structured data, such as a patient's physiological indicators, exhibit certain spatial correlations. For example, blood pressure and heart rate are physiologically related, and this correlation can be measured by spatial distance or correlation coefficient. Unstructured imaging data, such as medical CT images, has a distinct spatial structure, and the adjacency and spatial distribution characteristics of pixels contain important information. Time series data, on the other hand, exhibits continuity and trend in the temporal dimension. These characteristics are comprehensively considered to construct geometric constraints.

[0104] Furthermore, suppose d feature vectors l1, l2, ..., l are extracted from multimodal medical data. d , build a covariance matrix C d ∈R d*d , where C d (i,j)=Cov(l i ,l j ), represents the feature vector l i and l j The covariance between. For the covariance matrix C d Perform eigenvalue decomposition and get where Qd represents an orthogonal matrix, A d It is represented as a diagonal matrix with diagonal elements ζ i is the eigenvalue. According to the eigenvalue and eigenvector, the geometric constraints are constructed. For example, for the low-rank reconstructed matrix X W (i.e., the matrix to be reconstructed later is related to the second singular value matrix), which is required to satisfy Among them α1 and β q It is represented as parameters set according to data characteristics and analysis requirements. The constraints reflect the spatial geometric structure of multimodal medical data, restricting the shape of the reconstructed matrix in the data space, thereby making it more consistent with the geometric characteristics of the original data.

[0105] Step 30254: perform low-rank reconstruction on the second singular value matrix based on the geometric constraint to obtain the current causal tensor.

[0106] Optionally, the medical decision system uses the constructed geometric constraints to perform low-rank reconstruction on the second singular value matrix Σ2 to obtain the current causal tensor. An optimization algorithm, such as the alternating direction multiplier method (ADMM), is used to introduce geometric constraints during the reconstruction process. Suppose the objective function of the reconstruction is: where ||·|| F represents the Frobenius norm, which is used to measure the difference between matrices; Indicates that when performing the above minimization operation, the variable X W Conditions that need to be met; U e Expressed as the left singular vector matrix corresponding to Σ2; V e By solving this optimization problem, we can obtain the low-rank reconstruction matrix T that satisfies the geometric constraints. dan , convert it back to tensor form, and we get the current causal tensor T r Low-rank reconstruction is performed under geometric constraints, ensuring that the reconstruction result retains key information (through singular value threshold screening) while conforming to the spatial geometric structure of multimodal medical data. This enables the resulting current causal tensor to more accurately reflect the causal relationship of the patient's condition, providing more reliable data support for subsequent medical decision-making.

[0107] The embodiment of the present invention removes noise and redundant information from the updated causal tensor through singular value decomposition and threshold processing, thereby improving data quality and computational efficiency. Constraints are constructed based on the spatial geometric characteristics of multimodal medical data and low-rank reconstruction is performed, making full use of the intrinsic structural information of the data so that the current causal tensor more accurately reflects the causal relationship between multimodal data. The overall steps can effectively improve the quality of the causal tensor, provide a solid data foundation for subsequent medical decision-making optimization based on the causal tensor, and help improve the accuracy of medical decisions.

[0108] In one embodiment, steps 401 to 404 are described as follows:

[0109] Step 401 : determining a state transition probability based on the correlation between the current causal tensor and a preset clinical intervention action set.

[0110] Optionally, in the process of the medical decision system receiving the current causal tensor and determining the state transition probability, due to the current causal tensor T cur It reflects the causal relationship of the patient's multimodal medical data. And the preset clinical intervention action set B={b1,b2,…,b m} contains various possible medical intervention measures. In order to determine the state transition probability, we first define a similarity measurement function to measure the degree of change of the current causal tensor under different intervention actions. For example, if we use an expansion function based on cosine similarity, for the current causal tensor T r and intervention action b i , calculate the state transition probability P(ψ * |ψ,b i ) is:

[0111]

[0112] ; where ψ represents the current state (given by the current causal tensor T cur OK); * Indicates taking intervention action b i The next state after M0, N0, and T0 represent the size of the causal tensor in terms of the number of data modalities, the number of variables, and the time step dimension, respectively; Represented as and intervention action b i The relevant weight function is used to adjust the importance of different data elements when calculating similarity.

[0113] Furthermore, the weight function The calculation of can be based on domain knowledge and historical data. For example, for intervention action b iThe variables and time steps that are directly related are given higher weights; the parts that are less related are given lower weights. Assume that the intervention action b i has a greater impact on the variable n0 at time step t0, then the value of n0 at time step t0 is greater. The value of the corresponding position is greater.

[0114] Step 402, taking the current causal tensor as a state space, and dividing the state space based on the physiological index range and disease severity of the patient to obtain a plurality of subspaces.

[0115] Optionally, the medical decision system divides the current causal tensor T cur into a plurality of subspaces. r The dimension of the current causal tensor T s is M0*N0*T0. Then, a set of physiological indexes P m of the patient (such as heart rate, blood pressure, etc.) and a disease severity index S zz (represented as a comprehensive score) are obtained. The state space ψ is divided according to the range of these indexes. For a physiological index p j , assuming that its normal range is it is divided into f j intervals For the disease severity index S zz , assuming that it is divided into h levels By combining these intervals and levels, the state space ψ of the current causal tensor is divided. For example, a subspace S sub can be defined as: when the physiological index p and the disease verification degree S , the corresponding part of the current causal tensor constitutes a subspace. Mathematically, it is expressed as:

[0116] Step 403, analyzing the real-time state and disease trend of the patient to determine the causal allocation weight in each subspace.

[0117] Optionally, after the medical decision system determines each subspace S sub , it first analyzes the real-time state of the patient in the subspace. Assuming that by statistically analyzing the data of the current causal tensor in the subspace, the mean and standard deviation of each variable are obtained. Then, according to the disease trend, the causal allocation weight is determined. The disease trend can be measured by calculating the difference between the causal tensor of the current time step and the previous time steps. Let (Δt represents the time interval), which represents the variable The change in time step t^. Then the causal distribution weight in Representing variables The magnitude of the change at time step t^; Expressed as the number of variables in the causal tensor, For normalization, Considering variables The smaller the fluctuation, the greater the weight, because stable variables may be more indicative of the development of the disease.

[0118] Step 404: construct a Markov decision process based on the causal distribution weights, the state transition probabilities, the multiple subspaces, the preset clinical intervention action set, and the preset comprehensive objective reward function, and analyze and process the Markov decision process to obtain an optimized medical strategy.

[0119] Optionally, after the medical decision system determines the causal distribution weights, state transition probabilities, multiple subspaces, a preset clinical intervention action set, and a preset target reward function, reinforcement learning is constructed based on the Markov decision process, that is, in the Markov decision process (MDP), the state refers to the subspace S after division. sub , action refers to the operation b in the preset clinical intervention action set B i , state transition probability P(ψ * |ψ,b i ) refers to the possibility of transferring from one subspace to another by performing a certain action, the comprehensive objective reward function R(ψ, b i ) is used to evaluate the quality of each state-action combination and the discount factor γ∈[0,1]. Construct the above Markov decision process and analyze the MDP through algorithms such as policy iteration or value iteration to find the optimal strategy. In policy iteration, a strategy π is first initialized, and then the state value function V under the strategy is continuously evaluated. π (S sub ) and the action-value function Q π (S sub ,b i ), by comparing the values ​​of different actions, the policy π is updated until the policy converges and the optimized medical policy is obtained.

[0120] Furthermore, in policy iteration, the state value function V π (S sub ) is updated as follows:

[0121]

[0122] ; Where B represents the action set; π(b i |S sub ) is represented as in state S sub Next take action b i The probability of (determined by the strategy); P(S sub * |S sub ,b i ) represents the state S sub Collection action b i Transfer to state S sub * The probability of R(S sub , b i ,S sub * ) represents the state S sub Take action b i Transfer to state S sub * The reward obtained, γ is expressed as a discount factor, which is used to balance the importance of current rewards and future rewards.

[0123] Action-value function Q π (S sub ,b i ) is:

[0124]

[0125] By continuously iteratively updating V π (S sub ) and Q π (S sub ,b i ), and according to Q π (S sub ,b i )’s value update strategy π, namely Ultimately, we find the optimal medical strategy. These formulas are based on the principle of Markov decision process, which optimizes the strategy by iteratively calculating the state value and action value.

[0126] After multiple iterations, when the value function converges, the obtained strategy is the optimized medical strategy.

[0127] The embodiment of the present invention comprehensively considers factors such as the correlation between the current causal tensor and the intervention action, the patient's physiological indicators and disease severity, the real-time status and disease change trend, and constructs a reinforcement learning objective based on the Markov decision process. It fully utilizes the information in multimodal medical data and accurately simulates the medical decision-making process. The resulting optimized medical strategy has higher accuracy and adaptability, which helps to improve the quality of medical care.

[0128] In an embodiment, the target reward function is obtained as described in steps 4041-4044:

[0129] Step 4041, determine the clinical target reward function based on the improvement of the patient's condition.

[0130] Optionally, the medical decision system selects a series of key clinical indicators to measure the improvement of the patient's condition, such as cure rate, symptom relief degree, etc. for a certain disease, to determine the clinical target reward function, i.e. short-term reward, which aims to reward the doctor for taking decisions that can improve the patient's physiological indicators. Taking septic patients as an example, the clinical indicators are blood pressure, heart rate, and blood oxygen, etc. The clinical target reward function is:

[0131]

[0132] ; wherein, represents the number of key physiological indicators (such as blood pressure, heart rate, blood oxygen, etc.); represents the physiological indicator value at time t ° ; b i represents the action taken; S sub represents the current state (i.e. the subspace determined by the current causal tensor); represents the physiological indicator value after taking b i action; η i represents the indicator importance weight (such as blood pressure is more important than heart rate for patients with low blood pressure).

[0133] Step 4042, determine the condition stability reward function based on the stability of the patient's condition.

[0134] Optionally, the medical decision system measures the stability of the patient's condition based on the fluctuation of the patient's vital signs (such as heart rate, blood pressure, etc.), thereby avoiding overly aggressive strategies in reinforcement learning that can cause the patient's condition to fluctuate dramatically (such as avoiding excessive fluctuations in blood pressure and blood sugar that can cause shock or coma; avoiding the use of excessive drugs that can cause side effects, etc.), and thereby determining the condition stability reward function. The condition stability reward function is:

[0135]

[0136] ; wherein, ε i represents the stability weight (such as for ICU patients, blood sugar stability is more important than body temperature stability); represents the change amplitude of the variable, and if the change is too large, it will be punished for reinforcement learning. And the condition stability reward is negative, so that reinforcement learning avoids high-risk strategies.

[0137] Step 4043, determine the causal correction reward function based on the true causal relationship of the patient's condition.

[0138] Alternatively, when determining the causal correction reward function, the medical decision-making system primarily aims to ensure that reinforcement learning optimizes causal relationships rather than correlations. For example, a wrong approach might be: reinforcement learning discovers that ICU patient survival is correlated with hyperglycemia, leading to the mistaken assumption that "increasing blood sugar" will improve survival. A correct approach might be: reinforcement learning discovers that hyperglycemia in sepsis patients is caused by insulin resistance, leading to the correct strategy of "optimizing antibiotic treatment." The final causal correction reward function is:

[0139]

[0140] ; where M1 represents the number of causal variables; λ causal Expressed as the weight controlling the causal reward; P(Y|do(B=b i )) means taking action b i The probability that the patient's condition will improve after i * )) represents taking action b i * (control group), the probability that the patient will improve; Expressed as the influence weight of the causal variable.

[0141] Step 4044: Perform weighted summation on the clinical target reward function, the disease stability reward function, and the causal correction reward function to obtain the comprehensive target reward function.

[0142] Optionally, after determining the clinical target reward function, the disease stability reward function, and the causal correction reward function, the medical decision system performs a weighted summation to obtain a comprehensive target reward function:

[0143] R(ψ,b i )=w1R clinical (ψ,b i )-w2R stability (ψ,b i )+w3R causal (ψ,b i )

[0144] ; Among them, w1, w2 and w3 represent weight coefficients, which are used to balance the importance of different reward items; and w1+w2+w3=1.

[0145] The embodiments of the present application comprehensively consider various factors of the patient's condition, so that the target reward function more comprehensively and accurately reflects the value of the medical decision. By respectively determining the reward functions of different aspects and weighted summation, the quantitative evaluation of the effect of complex medical decision is realized, which provides a more scientific guide for reinforcement learning, helps to improve the quality and effect of medical decision, and finally improves the treatment result of the patient.

[0146] Further, the artificial intelligence-based medical decision optimization system provided by the present application is described below, and the artificial intelligence-based medical decision optimization system described below can be correspondingly referred to the method of artificial intelligence-based medical decision optimization described above.

[0147] Optionally, referring to Figure 2 , Figure 2 is a structural schematic diagram of the artificial intelligence-based medical decision optimization system provided by the present application, and the artificial intelligence-based medical decision optimization system comprises.

[0148] The acquisition unit 210 is configured to acquire multi-modal medical data; the multi-modal medical data comprises structured data, unstructured data and time series data;

[0149] The causal tensor construction unit 220 is configured to determine an initial causal tensor based on the causal relationship among the structured data, the unstructured data and the time series data; the initial causal tensor is a three-order tensor, and the dimensions of the three-order tensor are respectively represented as the number of data modalities, the number of variables and the time step;

[0150] The state updating unit 230 is configured to perform low-rank tensor decomposition on the initial causal tensor to obtain a key causal pattern, and update the initial causal tensor based on the dynamic change of the patient's condition combined with the key causal pattern to obtain a current causal tensor;

[0151] The strategy optimization unit 240 is configured to perform decision optimization on the current causal tensor based on reinforcement learning to obtain an optimized medical strategy.

[0152] The embodiment of the present invention constructs an initial causal tensor through structured data, unstructured data and time series data in multimodal medical data, which can not only effectively integrate the causal relationship between multimodal data, clearly display the complex causal relationship between data, and avoid the problem of incorrect decision-making caused by the inability to understand the causal relationship, but also provide a more accurate basis for the initial causal tensor in the subsequent optimization of medical decisions through the fusion of multimodal data. Afterwards, by performing low-rank tensor decomposition on the initial causal tensor and combining it with the dynamic changes of the patient's condition, a current causal tensor that can reflect the condition in real time is obtained, and the dynamic update of the initial causal tensor is realized to adapt to the dynamic changes of the patient's condition, and further provide a more accurate basis for subsequent medical decisions. Finally, the dynamically updated current causal tensor is combined with reinforcement learning to obtain an optimized medical strategy, which solves the problem of difficulty in making comprehensive and accurate medical decisions and improves the accuracy of the medical strategy.

[0153] See also Figure 3 , Figure 3 This is a diagram of an embodiment of an electronic device provided by an embodiment of the present invention. Figure 3 As shown, an embodiment of the present invention provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, the following steps are implemented:

[0154] Acquiring multimodal medical data; the multimodal medical data includes structured data, unstructured data, and time series data;

[0155] Determining an initial causal tensor based on the causal relationship between the structured data, the unstructured data, and the time series data; wherein the initial causal tensor is a third-order tensor, and the dimensions of the third-order tensor are respectively represented by the number of data modes, the number of variables, and the time step;

[0156] Performing low-rank tensor decomposition on the initial causal tensor to obtain a key causal pattern, and updating the initial causal tensor based on the dynamic changes of the patient's condition in combination with the key causal pattern to obtain a current causal tensor;

[0157] Based on reinforcement learning, the current causal tensor is optimized for decision making to obtain an optimized medical strategy.

[0158] See also Figure 4 , Figure 4 Detailed description of an embodiment of a computer-readable storage medium provided by an embodiment of the present invention. Figure 4 As shown, this embodiment provides a computer-readable storage medium 400 on which a computer program 311 is stored. When the computer program 311 is executed by a processor, the following steps are implemented:

[0159] Acquiring multimodal medical data; the multimodal medical data includes structured data, unstructured data, and time series data;

[0160] Determining an initial causal tensor based on the causal relationship between the structured data, the unstructured data, and the time series data; wherein the initial causal tensor is a third-order tensor, and the dimensions of the third-order tensor are respectively represented by the number of data modes, the number of variables, and the time step;

[0161] Performing low-rank tensor decomposition on the initial causal tensor to obtain a key causal pattern, and updating the initial causal tensor based on the dynamic changes of the patient's condition in combination with the key causal pattern to obtain a current causal tensor;

[0162] Based on reinforcement learning, the current causal tensor is optimized for decision making to obtain an optimized medical strategy.

[0163] In another aspect, the present invention further provides a computer program product, comprising a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform the method for optimizing medical decision-making based on artificial intelligence provided by the above methods, the method comprising:

[0164] Acquiring multimodal medical data; the multimodal medical data includes structured data, unstructured data, and time series data;

[0165] Determining an initial causal tensor based on the causal relationship between the structured data, the unstructured data, and the time series data; wherein the initial causal tensor is a third-order tensor, and the dimensions of the third-order tensor are respectively represented by the number of data modes, the number of variables, and the time step;

[0166] Performing low-rank tensor decomposition on the initial causal tensor to obtain a key causal pattern, and updating the initial causal tensor based on the dynamic changes of the patient's condition in combination with the key causal pattern to obtain a current causal tensor;

[0167] Based on reinforcement learning, the current causal tensor is optimized for decision making to obtain an optimized medical strategy.

[0168] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0169] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or certain parts of the embodiments.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for optimizing medical decision-making based on artificial intelligence, characterized in that: include: Acquiring multimodal medical data; the multimodal medical data includes structured data, unstructured data, and time series data; The structured data includes the patient's basic information, diagnosis results, vital signs data and test results; the unstructured data includes medical imaging data and medical record text data; the time series data includes the patient's vital signs data collected regularly or periodically; Determining an initial causal tensor based on the causal relationship between the structured data, the unstructured data, and the time series data; wherein the initial causal tensor is a third-order tensor, and the dimensions of the third-order tensor are respectively represented by the number of data modes, the number of variables, and the time step; Performing low-rank tensor decomposition on the initial causal tensor to obtain a key causal pattern, and updating the initial causal tensor based on the dynamic changes of the patient's condition in combination with the key causal pattern to obtain a current causal tensor; including: analyzing the causal flow between the variables in the key causal pattern based on the dynamic changes of the condition to obtain a causal flow analysis result; Determining a time window for causal flow analysis based on the speed and trend of the dynamic changes in the disease condition; Performing intensity calculation processing on the causal flow analysis result based on the transfer entropy function to obtain the causal flow intensity; updating the initial causal tensor based on the causal flow strength and the time window to obtain an updated causal tensor; Performing low-rank reconstruction on the updated causal tensor to obtain the current causal tensor; Optimizing the current causal tensor based on reinforcement learning to obtain an optimized medical strategy; including: Determining a state transition probability based on the correlation between the current causal tensor and a preset clinical intervention action set; The current causal tensor is used as a state space, and the state space is divided based on the patient's physiological indicator range and disease severity to obtain multiple subspaces; Analyze the patient's real-time status and disease change trends to determine the causal distribution weights in each subspace; A Markov decision process is constructed based on the causal distribution weight, the state transition probability, the multiple subspaces, the preset clinical intervention action set and the preset comprehensive target reward function, and the Markov decision process is analyzed and processed to obtain an optimized medical strategy.

2. The method for optimizing medical decision-making based on artificial intelligence according to claim 1, characterized in that: The low-rank tensor decomposition of the initial causal tensor to obtain a key causal pattern includes: Performing feature division on the initial causal tensor to obtain a plurality of local features, and performing feature analysis on each of the local features to obtain a local feature analysis result; Adjusting kernel function parameters based on the local feature analysis result to obtain an adjusted kernel function; Performing a low-rank decomposition on the initial causal tensor based on the adjusted kernel function to obtain a low-rank tensor factor; Error verification is performed on the low-rank tensor factor and the initial causal tensor, and analysis and judgment are performed based on the error verification results to obtain the key causal pattern.

3. The method for optimizing medical decision-making based on artificial intelligence according to claim 2, characterized in that: The low-rank reconstruction of the updated causal tensor to obtain the current causal tensor includes: Performing singular value decomposition on the updated causal tensor to obtain a first singular value matrix; updating the singular values ​​in the first singular value matrix based on a preset singular value threshold to obtain a second singular value matrix; Constructing geometric constraints based on the spatial geometric characteristics of the multimodal medical data; The second singular value matrix is ​​low-rank reconstructed based on the geometric constraint to obtain the current causal tensor.

4. The method for optimizing medical decision-making based on artificial intelligence according to claim 1, characterized in that: The determining of an initial causal tensor based on the causal relationship among the structured data, the unstructured data, and the time series data includes: Preprocessing and feature extraction are performed on the structured data, the unstructured data, and the time series data to obtain structured features, unstructured features, and time series features; An initial Bayesian network is constructed by taking the structured features, the unstructured features, and the time series features as nodes and the independent relationships between the nodes as edges; Performing information entropy processing on each node in the initial Bayesian network based on a preset information entropy function to obtain node information entropy; Performing mutual information processing on the nodes at both ends of each edge in the initial Bayesian network based on a preset mutual information function to obtain node mutual information between the nodes at both ends of each edge; The weight of each edge in the initial Bayesian network is adjusted based on the node information entropy and the node mutual information, a current Bayesian network is determined, and a tensor transformation is performed on the current Bayesian network to obtain the initial causal tensor.

5. The method for optimizing medical decision-making based on artificial intelligence according to claim 1, characterized in that: The steps for obtaining the comprehensive objective reward function are: Determine the clinical objective reward function based on the improvement of the patient's condition; Determine a reward function for the stability of the patient's condition based on the stability of the patient's condition; Determine a causal correction reward function based on the true causal relationship of the patient's condition; The clinical target reward function, the disease stability reward function and the causal correction reward function are weightedly summed to obtain the comprehensive target reward function.

6. A medical decision-making system based on artificial intelligence optimization, characterized in that: The method for optimizing medical decision-making based on artificial intelligence according to any one of claims 1 to 5; the system for optimizing medical decision-making based on artificial intelligence comprises: Acquisition unit: used to acquire multimodal medical data; the multimodal medical data includes structured data, unstructured data and time series data; the structured data includes the patient's basic information, diagnosis results, vital signs data and test results; the unstructured data includes medical imaging data and medical record text data; the time series data includes the patient's vital signs data collected regularly or periodically; A causal tensor construction unit is configured to determine an initial causal tensor based on the causal relationship between the structured data, the unstructured data, and the time series data; the initial causal tensor is a third-order tensor, and the dimensions of the third-order tensor are respectively represented by the number of data modes, the number of variables, and the time step; A state updating unit is configured to perform a low-rank tensor decomposition on the initial causal tensor to obtain a key causal pattern, and update the initial causal tensor based on the dynamic changes of the patient's condition in combination with the key causal pattern to obtain a current causal tensor; including: analyzing the causal flow between the variables in the key causal pattern based on the dynamic changes of the condition to obtain a causal flow analysis result; Determining a time window for causal flow analysis based on the speed and trend of the dynamic changes in the disease condition; Performing intensity calculation processing on the causal flow analysis result based on the transfer entropy function to obtain the causal flow intensity; updating the initial causal tensor based on the causal flow strength and the time window to obtain an updated causal tensor; Performing low-rank reconstruction on the updated causal tensor to obtain the current causal tensor; A strategy optimization unit is used to optimize the decision of the current causal tensor based on reinforcement learning to obtain an optimized medical strategy; it includes: Determining a state transition probability based on the correlation between the current causal tensor and a preset clinical intervention action set; The current causal tensor is used as a state space, and the state space is divided based on the patient's physiological indicator range and disease severity to obtain multiple subspaces; Analyze the patient's real-time status and disease change trends to determine the causal distribution weights in each subspace; A Markov decision process is constructed based on the causal distribution weight, the state transition probability, the multiple subspaces, the preset clinical intervention action set and the preset comprehensive target reward function, and the Markov decision process is analyzed and processed to obtain an optimized medical strategy.

7. An electronic device comprising: Memory for storing computer software programs; A processor for reading and executing the computer software program, wherein when the processor executes the computer software program, it implements the method for optimizing medical decision-making based on artificial intelligence as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium storing a computer software program, characterized in that: When the computer software program is executed by a processor, the method for optimizing medical decision-making based on artificial intelligence as described in any one of claims 1 to 5 is implemented.

Images