Artificial intelligence-based medical data processing method, device, equipment and medium
By processing medical data through self-supervised learning, multi-task learning, and multi-objective optimization, a target medical analysis model is generated, which solves the problems of high cost and low efficiency in existing technologies and achieves efficient and accurate medical data evaluation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- PING AN TECH (SHENZHEN) CO LTD
- Filing Date
- 2023-06-20
- Publication Date
- 2026-07-14
AI Technical Summary
Existing biomolecular tasks typically employ rule-based or statistical methods, which require extensive artificial feature engineering and domain knowledge. They struggle to handle complex nonlinear relationships and multimodal data, resulting in high processing costs and low efficiency.
An artificial intelligence-based approach is used to collect initial data from multiple medical data sources. The medical analysis model is then trained and optimized through self-supervised learning, multi-task learning, multi-objective optimization, and feedback learning to generate a target medical analysis model for evaluating the processing of target medical data.
It improves the generalization ability, expressive ability, balance and adaptability of medical analysis models, enables the generation of rapid and accurate medical assessment results, and reduces assessment processing costs.
Smart Images
Figure CN116776110B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of artificial intelligence technology and digital healthcare, and in particular to a method, apparatus, device and medium for medical data processing based on artificial intelligence. Background Technology
[0002] Biomolecular tasks refer to the use of computer simulations and analyses of the structure and function of biomolecules such as compounds, proteins, and RNA (ribonucleic acid), as well as their interactions, to provide theoretical guidance and experimental basis for the biopharmaceutical field. Biomolecular tasks include structure prediction, property prediction, affinity prediction, drug design, and target discovery. These tasks are of great significance and value for new drug development, disease diagnosis, and personalized medicine.
[0003] Existing biomolecular tasks typically employ rule-based or statistical methods, which require extensive artificial feature engineering and domain knowledge. Furthermore, they struggle to handle complex nonlinear relationships and multimodal data, resulting in high costs and low efficiency in biomolecular task processing. Summary of the Invention
[0004] This application provides an artificial intelligence-based medical data processing method, device, computer equipment, and medium to address the technical problems of high cost and low efficiency in the processing of biomolecular tasks, which are usually carried out using rule-based or statistical methods, requiring a large amount of artificial feature engineering and domain knowledge, and are difficult to handle complex nonlinear relationships and multimodal data.
[0005] Firstly, an artificial intelligence-based medical data processing method is provided, including:
[0006] Initial medical data was collected from multiple medical data sources;
[0007] The initial medical data is preprocessed to obtain a medical sample set;
[0008] Based on the self-supervised learning method, the pre-set multimodal basic model is trained using the medical sample set to obtain an initial medical analysis model; wherein, the multimodal basic model is a model with an encoder and decoder architecture;
[0009] The initial medical analysis model is fine-tuned using a preset multi-task learning strategy to obtain the first medical analysis model;
[0010] The first medical analysis model is optimized using a pre-defined multi-objective optimization strategy to obtain a second medical analysis model;
[0011] The second medical analysis model is improved by using a pre-defined feedback learning method to obtain the improved target medical analysis model.
[0012] The target medical data to be processed is evaluated based on the target medical analysis model, and a medical evaluation result corresponding to the target medical data is generated.
[0013] Secondly, an artificial intelligence-based medical data processing device is provided, comprising:
[0014] The collection module is used to collect initial medical data from multiple medical data sources;
[0015] The first processing module is used to preprocess the initial medical data to obtain a medical sample set;
[0016] The training module is used to train a pre-defined multimodal basic model using the medical sample set based on a self-supervised learning method to obtain an initial medical analysis model; wherein the multimodal basic model is a model with an encoder and decoder architecture.
[0017] The second processing module is used to fine-tune the initial medical analysis model through a preset multi-task learning strategy to obtain the first medical analysis model.
[0018] The third processing module is used to optimize the first medical analysis model through a preset multi-objective optimization strategy to obtain a second medical analysis model;
[0019] The fourth processing module is used to improve the second medical analysis model through a preset feedback learning method to obtain the improved target medical analysis model.
[0020] The evaluation module is used to evaluate the target medical data to be processed based on the target medical analysis model, and generate medical evaluation results corresponding to the target medical data.
[0021] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described artificial intelligence-based medical data processing method.
[0022] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described artificial intelligence-based medical data processing method.
[0023] The aforementioned scheme, implemented using artificial intelligence-based medical data processing methods, devices, computer equipment, and storage media, involves: collecting initial medical data from multiple medical data sources; preprocessing the initial medical data to obtain a medical sample set; training a pre-defined multimodal basic model using the medical sample set based on a self-supervised learning method to obtain an initial medical analysis model; fine-tuning the initial medical analysis model using a pre-defined multi-task learning strategy to obtain a first medical analysis model; optimizing the first medical analysis model using a pre-defined multi-objective optimization strategy to obtain a second medical analysis model; improving the second medical analysis model using a pre-defined feedback learning method to obtain an improved target medical analysis model; and evaluating the target medical data to be processed based on the target medical analysis model to generate medical evaluation results corresponding to the target medical data. A medical sample set is obtained by preprocessing initial medical data collected from medical data sources. This sample set is then used to train a pre-defined multimodal base model, resulting in an initial medical analysis model. Subsequently, multi-task learning, multi-objective optimization, and feedback learning strategies are employed to optimize the initial model and generate a target medical analysis model. This effectively improves the generalization ability, expressive power, balance and coordination, adaptability, and overall performance of the generated target medical analysis model. The target medical analysis model is then used to evaluate the target medical data, enabling the rapid and accurate generation of medical evaluation results corresponding to the target medical data. This significantly improves the efficiency and reduces the cost of evaluating and processing the target medical data. Attached Figure Description
[0024] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 This is a schematic diagram of an application environment for an artificial intelligence-based medical data processing method according to an embodiment of this application;
[0026] Figure 2 This is a flowchart illustrating a medical data processing method based on artificial intelligence in one embodiment of this application;
[0027] Figure 3 This is a schematic diagram of a medical data processing device based on artificial intelligence according to one embodiment of this application;
[0028] Figure 4 This is a schematic diagram of the structure of a computer device according to one embodiment of this application;
[0029] Figure 5 This is another structural schematic diagram of a computer device in one embodiment of this application. Detailed Implementation
[0030] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0031] The artificial intelligence-based medical data processing method provided in this application can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can collect initial medical data from multiple medical data sources; preprocess the initial medical data to obtain a medical sample set; train a pre-defined multimodal basic model using the medical sample set based on a self-supervised learning method to obtain an initial medical analysis model; fine-tune the initial medical analysis model using a pre-defined multi-task learning strategy to obtain a first medical analysis model; optimize the first medical analysis model using a pre-defined multi-objective optimization strategy to obtain a second medical analysis model; improve the second medical analysis model using a pre-defined feedback learning method to obtain an improved target medical analysis model; evaluate the target medical data to be processed based on the target medical analysis model to generate medical evaluation results corresponding to the target medical data. In this application, a medical sample set is obtained by preprocessing initial medical data collected from medical data sources. This sample set is then used to train a pre-defined multimodal base model, resulting in an initial medical analysis model. Subsequently, a multi-task learning strategy, a multi-objective optimization strategy, and a feedback learning method are used to optimize the initial medical analysis model, generating a target medical analysis model. This effectively improves the generalization ability, expressive power, balance and coordination, adaptability, and overall performance of the generated target medical analysis model. The target medical analysis model is then used to evaluate the target medical data to be processed, achieving rapid and accurate generation of medical evaluation results corresponding to the target medical data. This effectively improves the efficiency and reduces the cost of evaluating and processing the target medical data. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. Specific embodiments of this application are described in detail below.
[0032] Please see Figure 2 As shown, Figure 2 A flowchart illustrating an artificial intelligence-based medical data processing method provided in this application embodiment includes the following steps:
[0033] Step S10: Collect initial medical data from multiple medical data sources.
[0034] The AI-based medical data processing method provided in this application can be applied to medical assessment applications in various scenarios. Data processing in medical assessment applications is typically implemented through a server-side component, which can receive initial medical data collected in real time from multiple medical data sources. These medical data sources can include public or private medical data sources. The initial medical data can include images, text, sound, molecular structures, biological sequences, and other data.
[0035] It should be noted that the implementation of this technical solution involves the collection, organization, and storage of case information and user information. All of the above-mentioned processing is carried out with the consent of the users and in strict compliance with national laws and regulations.
[0036] Step S20: Preprocess the initial medical data to obtain a medical sample set.
[0037] In this embodiment, the preprocessing described above may include at least cleaning, standardization, and encoding processes.
[0038] Step S30: Based on the self-supervised learning method, the preset multimodal basic model is trained using the medical sample set to obtain an initial medical analysis model; wherein, the multimodal basic model is a model with an encoder and decoder architecture.
[0039] In this embodiment, the aforementioned multimodal basic model is specifically a multi-level, multimodal, multi-task, and multi-objective basic model composed of an encoder E and a decoder D. The encoder E encodes data from different levels and modalities into a unified latent vector z, while the decoder D decodes the latent vector z into data from different levels and modalities. The multimodal basic model can simultaneously process medical data of different types and sources, improving the model's generalization and expressive capabilities. Furthermore, it can flexibly interpret different combinations of medical modalities, such as generating text descriptions based on molecular structures or vice versa, and can produce expressive outputs, such as free-text explanations, verbal suggestions, or image annotations, demonstrating advanced medical reasoning capabilities. The training process for the pre-defined multimodal basic model, based on the self-supervised learning method and using the medical sample set to obtain the initial medical analysis model, can refer to the training process of existing models and will not be elaborated upon here.
[0040] Step S40: Fine-tune the initial medical analysis model using a preset multi-task learning strategy to obtain the first medical analysis model.
[0041] In this embodiment, the initial medical analysis model is fine-tuned using a multi-task learning strategy, enabling the generated first medical analysis model to simultaneously perform multiple drug discovery-related tasks, such as molecular generation, optimization, classification, regression, clustering, alignment, and matching. It may also include tasks such as structure prediction, property prediction, affinity prediction, drug design, and target discovery for biomolecules such as compounds, proteins, and RNA. Furthermore, the first medical analysis model can share and transfer knowledge across different drug discovery tasks, improving generalization ability and efficiency, as well as enhancing the model's versatility and adaptability. The specific implementation process of fine-tuning the initial medical analysis model using a preset multi-task learning strategy to obtain the first medical analysis model will be further described in detail in subsequent embodiments and will not be elaborated upon here.
[0042] Step S50: Optimize the first medical analysis model using a preset multi-objective optimization strategy to obtain the second medical analysis model.
[0043] In this embodiment, a multi-objective optimization strategy is employed to optimize the first medical analysis model, enabling the generated second medical analysis model to consider multiple objectives and constraints during drug discovery, such as molecular activity, selectivity, stability, manufacturability, and deliverability. Furthermore, the second medical analysis model can balance and coordinate among different objectives, improving its balance and coordination. It also allows for drug design at different levels, enhancing the model's accuracy and innovation, and generating candidate molecules that meet multiple conditions. The specific implementation process of optimizing the first medical analysis model using a preset multi-objective optimization strategy to obtain the second medical analysis model will be further described in detail in subsequent embodiments and will not be elaborated upon here.
[0044] Step S60: Improve the second medical analysis model using a preset feedback learning method to obtain the improved target medical analysis model.
[0045] In this embodiment, a preset feedback learning method is used to improve the second medical analysis model, thereby dynamically adjusting and optimizing the target and strategy to obtain an improved target medical analysis model. The specific implementation process of improving the second medical analysis model using the preset feedback learning method to obtain the improved target medical analysis model will be described in more detail in subsequent embodiments of this application, and will not be elaborated upon here.
[0046] Step S70: Based on the target medical analysis model, evaluate the target medical data to be processed and generate a medical evaluation result corresponding to the target medical data.
[0047] In this embodiment, the target medical data to be processed can be biomolecular data corresponding to the biomolecular task to be processed. The specific implementation process of evaluating and processing the target medical data based on the target medical analysis model to generate medical evaluation results corresponding to the target medical data will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated upon here.
[0048] This application obtains a medical sample set by preprocessing initial medical data collected from medical data sources. This sample set is then used to train a pre-defined multimodal base model, resulting in an initial medical analysis model. Subsequently, a multi-task learning strategy, a multi-objective optimization strategy, and a feedback learning method are used to optimize the initial medical analysis model, generating a target medical analysis model. This effectively improves the generalization ability, expressive power, balance and coordination, adaptability, and overall performance of the generated target medical analysis model. The target medical analysis model is then used to evaluate the target medical data to be processed, enabling the rapid and accurate generation of medical evaluation results corresponding to the target medical data. This effectively improves the efficiency and reduces the cost of evaluating and processing the target medical data.
[0049] In some optional implementations of this embodiment, step S20 includes the following steps:
[0050] The initial medical data is cleaned to obtain the first medical data.
[0051] In this embodiment, data cleaning refers to the process of re-examining and validating data, aiming to remove duplicate information, correct existing errors, and ensure data consistency. Data cleaning may include consistency checks and the handling of invalid and missing values. Consistency checks examine whether the data meets requirements based on the reasonable value range and interrelationships of each variable, identifying data that exceeds normal ranges, is logically illogical, or contradicts itself. Additionally, due to survey, coding, and data entry errors, some invalid and missing values may exist in the data, requiring appropriate processing. Common processing methods include estimation, whole-case deletion, variable deletion, and pairwise deletion.
[0052] The first medical data is standardized to obtain the second medical data.
[0053] In this embodiment, standardization refers to transforming the original data according to a certain ratio using specific mathematical transformations, so that it falls within a small, specific interval, such as [0, 1] or [-1, 1]. This eliminates differences in the properties, dimensions, and orders of magnitude of different variables, transforming them into dimensionless relative values, i.e., standardized values. This ensures that the values of all indicators are at the same order of magnitude, facilitating comprehensive analysis and comparison of indicators with different units or orders of magnitude. Specifically, normalization and centralization methods can be used to standardize the first medical data to obtain the second medical data.
[0054] The second medical data is encoded to obtain the third medical data.
[0055] In this embodiment, the second medical data is encoded so that the resulting third medical data conforms to a format suitable for input into the medical analysis model.
[0056] The third medical data is used as the medical sample set.
[0057] This application cleanses initial medical data to obtain first medical data; then standardizes the first medical data to obtain second medical data; subsequently, it encodes the second medical data to obtain third medical data, which is then used as a medical sample set. This application achieves rapid and accurate preprocessing of the initial medical data through cleaning, standardization, and encoding, ensuring that the resulting medical sample set conforms to a format suitable for input into a medical analysis model. This facilitates the subsequent training and generation of the medical analysis model using the medical sample set, thereby improving the efficiency of medical analysis model generation.
[0058] In this embodiment, in some optional implementations, step S70 includes the following steps:
[0059] The target medical data is input into the input layer of the target medical analysis model, and the target medical data is embedded.
[0060] In this embodiment, the initial medical analysis model described above is a model architecture employing GNN+Transformer. Graph Convolutional Networks (GNNs) are a type of generalized neural network based on graph structures. Transformer is a deep learning model applicable to parallel computing that utilizes attention mechanisms to improve model training speed. Specifically, in the molecular application scenario of the medical field, the target medical data mentioned above can be a drug molecule graph.
[0061] The target medical data is encoded by the encoder in the target medical analysis model to obtain the first node feature corresponding to the target medical data.
[0062] In this embodiment, the encoder in the above-mentioned target medical analysis model is composed of multiple GNN and Transformer modules stacked alternately. Each module includes the following parts: (1) GNN part: using a message passing mechanism, it aggregates the features of itself and its neighboring nodes and updates the node features. Specifically, for each node The updated node characteristics are:
[0063] .in, These are activation functions, such as ReLU (Linear rectification function) and Tanh (hyperbolic tangent function). ReLU is an activation function in deep learning, which is a variant activation function that improves the linear rectification unit by considering Gaussian noise. and It is a learnable parameter matrix; It is a node The set of neighboring nodes. (2) Transformer part: Using the self-attention mechanism, capture the global dependencies between nodes and update node features. Specifically, for each node The updated node characteristics are: .in, It is a layer normalization operation; It is a multi-head attention mechanism. It is a learnable parameter matrix; It is the feature matrix of all nodes, where It is the number of nodes. (3) Skip the connection part: Add shallow node features to deep node features to enhance expressive power and avoid over-smoothing. Specifically, for each node The updated node characteristics are: .in, It is a layer normalization operation.
[0064] Based on the latent spatial layer in the target medical analysis model, the features of the first node are pooled and sampled to obtain the corresponding features of the second node.
[0065] In this embodiment, in the latent space layer of the target medical analysis model, the feature representation of the entire drug molecule graph can be obtained by pooling the node features obtained from the encoder. Then, through a fully connected layer, the mean and variance of the latent variables are output, and the latent variables are sampled using a reparameterization technique. Specifically, for a drug molecule graph... Its latent variables are: in and These are the mean and variance of the latent variables, respectively. and It consists of learnable parameter matrices and vectors; It is a feature representation of the drug molecule map obtained by the encoder; It is a random noise vector; It refers to the dimension of the latent variable. MeanPooling refers to mean pooling, which is to average the value of feature points in the neighborhood.
[0066] The second node feature is updated based on the decoder in the target medical analysis model to obtain the corresponding third node feature.
[0067] In this embodiment, the decoder in the above-mentioned target medical analysis model is composed of multiple GNN and Transformer modules stacked alternately. Each module includes the following parts: (1) GNN part: using a message passing mechanism, it aggregates the features of itself and its neighboring nodes and updates the node features. Specifically, for each node The updated node characteristics are: .in These are activation functions, such as ReLU, Tanh, etc. and It is a learnable parameter matrix; It is a node The set of neighboring nodes. (2) Transformer part: Using the self-attention mechanism, capture the global dependencies between nodes and update node features. Specifically, for each node The updated node characteristics are: .in It is a layer normalization operation; It is a multi-head attention mechanism; It is a learnable parameter matrix; It is the feature matrix of all nodes, where It is the number of nodes. (3) Skip the connection part: Add shallow node features to deep node features to enhance expressive power and avoid over-smoothing. Specifically, for each node The updated node characteristics are: .in It is a layer normalization operation.
[0068] The third node features are evaluated and processed based on the output layer of the target medical analysis model to obtain the medical evaluation results corresponding to the target medical data.
[0069] In this embodiment, the aforementioned output layer can refer to the fully connected layer in the target medical analysis model. This layer outputs predicted values for nodes and edges by passing the node features obtained from the decoder through it, and generates medical evaluation results corresponding to the target medical data according to predefined rules. Specifically, for each node... Its predicted value is: .in and It consists of learnable parameter matrices and vectors; It is the number of node categories; It is a probability vector representing a node. The probability of belonging to each category. Softmax is a mathematical function commonly used to transform a set of arbitrary real numbers into real numbers representing a probability distribution. Essentially, it's a normalization function that converts a set of arbitrary real values into probability values between [0, 1]. Because softmax transforms them into values between 0 and 1, they can be interpreted as probabilities. If one of the inputs is small or negative, softmax transforms it into a low probability; if the input is large, it transforms it into a high probability, but it will always remain between 0 and 1. For each edge... Its predicted value is: .in and It consists of learnable parameter matrices and vectors; It is the number of edge types; It is a probability vector representing the edge The probability of belonging to each type. The predefined rules mentioned above may include rules such as maximum probability, threshold, and topology, to generate corresponding medical assessment results from the predicted values.
[0070] This application inputs target medical data into the input layer of a target medical analysis model and embeds the data. Then, the encoder in the model encodes the target medical data to obtain the first node feature corresponding to the data. Next, the first node feature is pooled and sampled based on the latent space layer of the model to obtain the corresponding second node feature. Subsequently, the decoder in the model updates the second node feature to obtain the corresponding third node feature. Finally, the output layer of the model evaluates the third node feature to obtain the medical evaluation result corresponding to the target medical data. This application uses a target medical analysis model trained on a medical sample set to evaluate target medical data, enabling rapid and accurate generation of medical evaluation results, improving the efficiency of target medical data evaluation, and ensuring the accuracy of the generated medical evaluation results.
[0071] In some optional implementations, the step of encoding the target medical data through the encoder in the target medical analysis model to obtain the first node feature corresponding to the target medical data includes the following steps:
[0072] The target data type of the target medical data is obtained.
[0073] In this embodiment, the target data type may include molecular diagrams, text sequences, images, etc. The text sequence may be medical text, which may be electronic medical records, electronic personal health records, specifically including medical records, electrocardiograms, medical images, and a series of electronic records with preservation and future reference value.
[0074] Determine the target encoding method that matches the target data type.
[0075] In this embodiment, for any input data It can be a molecular diagram. A text sequence An image Or any other type of data, encoder It can be encoded as a latent vector For different types of data, the encoder Different encoding methods can be used, for example, for molecular diagrams. Encoding is performed using graph neural networks (GNNs) or self-attention mechanisms; for text sequences Encoding is performed using recurrent neural networks (RNNs) or Transformers; for images Encoding can be performed using a convolutional neural network (CNN) or a Vision Transformer.
[0076] The target medical data is encoded using the encoder in the target medical analysis model and the target encoding method to obtain encoded data corresponding to the target medical data.
[0077] The encoded data is used as the feature of the first node.
[0078] This application obtains the target data type of the target medical data; then determines the target encoding method matching the target data type; subsequently, through the encoder in the target medical analysis model, the target medical data is encoded using the target encoding method to obtain coded data corresponding to the target medical data, and this coded data is used as the first node feature. This application, by obtaining the target data type of the target medical data and then intelligently using the encoder in the target medical analysis model to encode the target medical data using a target encoding method matching the target data type, achieves accurate encoding of the target medical data, improves the intelligence of the encoding process, and ensures the accuracy of the generated first node feature.
[0079] In some alternative implementations, for the decoder in the target medical analysis model, for any latent vector It can be made by an encoder The decoder, whether generated or otherwise. It can be decoded into any type of data. For different types of data, the decoder Different decoding methods can be used, for example, for molecular diagrams. Decoding can be performed using a graph generative network or an autoregressive model; for text sequences... Decoding is performed using a recurrent neural network (RNN) or a Transformer; for images Decoding can be performed using deconvolutional neural networks or GANs (Generative Adversarial Networks).
[0080] In some alternative implementations, step S40 includes the following steps:
[0081] Obtain the loss functions for various pre-defined drug discovery tasks;
[0082] In this embodiment, the drug discovery task may include tasks such as molecule generation, optimization, classification, regression, clustering, alignment, and matching. A task-specific loss function is predefined for each drug discovery task. ,in Input data The latent vector, It is the predicted value output by the model.
[0083] All the loss functions are combined to obtain the corresponding total loss function.
[0084] In this embodiment, the loss functions of different drug discovery tasks are combined into a single total loss function. ,in These are the weights of different tasks, which can be adjusted based on the importance or difficulty of the drug discovery task.
[0085] The total loss function is optimized using gradient descent to obtain the optimized target total loss function.
[0086] In this embodiment, the total loss function is optimized using gradient descent. To obtain the optimized target total loss function
[0087] The model parameters of the initial medical analysis model are updated based on the target total loss function to obtain the first medical analysis model.
[0088] In this embodiment, the model parameters of the initial medical analysis model are updated based on the target total loss function. Thus, the first medical analysis model is obtained.
[0089] This application obtains the loss functions of various pre-defined drug discovery tasks; then combines all the loss functions to obtain the corresponding total loss function; subsequently, it optimizes the total loss function using gradient descent to obtain the optimized target total loss function; finally, it updates the model parameters of the initial medical analysis model based on the target total loss function to obtain the first medical analysis model. This application uses a multi-task learning strategy to fine-tune the initial medical analysis model on various drug discovery-related tasks, based on the total loss function obtained by combining the loss functions of various drug discovery tasks, to obtain the first medical analysis model. This achieves the generation of a first medical analysis model with the ability to share and transfer knowledge, improving the model's generalization and expressive capabilities.
[0090] In some optional implementations of this embodiment, step S50 includes the following steps:
[0091] Obtain predefined task objectives corresponding to various drug discovery tasks.
[0092] In this embodiment, the drug discovery tasks described above may include tasks such as molecule generation, optimization, classification, regression, clustering, alignment, and matching. Predefined task objectives corresponding to various drug discovery tasks may include, for example, molecular activity, selectivity, stability, manufacturability, and deliverability. Task objectives can also be referred to as task goals and constraints.
[0093] Obtain the objective function corresponding to each of the stated task objectives.
[0094] In this embodiment, for the task objective of drug discovery, an objective function corresponding to the task objective is pre-constructed based on actual usage requirements. ,in, Input data The implicit vector.
[0095] All the objective functions are combined to obtain the corresponding overall objective function.
[0096] In this embodiment, the objective functions of all drug discovery tasks can be combined into a single overall objective function. ,in, These are the weights for different objectives, which can be adjusted based on user preferences or experimental results.
[0097] The second medical analysis model is obtained by optimizing the input data, output data, latent vectors, and overall objective function of the first medical analysis model based on a preset multi-objective evolutionary algorithm.
[0098] In this embodiment, the first medical analysis model can be simply referred to as the model. The optimal set of solutions that satisfies the objectives and constraints of multiple drug discovery tasks can be found using a multi-objective evolutionary algorithm (MOEA). ,in It is the latent vector space. We will use the model's input data. Treating the model's output data as an individual Treating the individual's phenotype as the model's latent vectors Treating an individual's genotype as the overall objective function This is considered as the fitness of an individual. We define population initialization, selection, crossover, mutation, and update operations according to different MOEA algorithms (Multi-Objective Evolutionary Algorithms) to generate a new generation of the population, which gradually converges to the optimal solution set. .
[0099] This application obtains predefined task objectives corresponding to various drug discovery tasks; then obtains the objective functions corresponding to each task objective; subsequently, it combines all objective functions to obtain the corresponding overall objective function; and then optimizes the input data, output data, latent vectors, and overall objective function of the first medical analysis model based on a preset multi-objective evolutionary algorithm to obtain the second medical analysis model. By using a multi-objective optimization strategy, this application optimizes the first medical analysis model for various drug discovery-related tasks based on the overall objective function obtained by combining the objective functions of the task objectives corresponding to each drug discovery task, thereby generating a second medical analysis model with the ability to balance and coordinate among different objectives. This allows the second medical analysis model to consider multiple objectives and constraints during the drug discovery process, improving the model's balance and coordination.
[0100] In some alternative implementations, context learning methods can also be used, enabling the medical analysis model to perform tasks that have never been explicitly trained on, using text prompts or a small number of examples. Specifically, this is achieved by encoding the text prompts or examples into a latent vector. ,in, This is a text hint or example. It will include implicit vectors. With input data latent vectors By concatenating or weighting the vectors, a new latent vector is obtained. ,in It is a concatenation or weighted average function. It will generate a new latent vector. via decoder Decode into corresponding output data For example, if you want to design an antibody against the novel coronavirus, you could provide a text prompt such as "Design an antibody against the RBD binding domain of the novel coronavirus S protein," or an example such as "REGN-COV2 is an antibody mixture composed of two monoclonal antibodies that bind to epitopes at different positions of the S protein RBD binding domain, preventing the virus from binding to the ACE2 receptor." This is achieved by encoding the text prompt or example into a latent vector. and input data Latent vectors (which can be data of any type and source) By concatenating or weighting the vectors, a new latent vector is obtained. The new hidden vector will be used subsequently. via decoder Decoded into an antibody molecule structure And output it to the user.
[0101] In some optional implementations of this embodiment, step S60 includes the following steps:
[0102] Obtain experimental results and expert opinions corresponding to the pre-set drug discovery tasks.
[0103] In this embodiment, the drug discovery task may include tasks such as molecule generation, optimization, classification, regression, clustering, alignment, and matching.
[0104] The experimental results and expert opinions are encoded to obtain corresponding feedback information.
[0105] In this embodiment, experimental results or expert opinions can be encoded into a feedback signal. ,in It is a feedback function. It is the actual value. This is a predicted value.
[0106] The output data and feedback information of the second medical analysis model are optimized using reinforcement learning methods to update the model parameters of the second medical analysis model and obtain the target medical analysis model.
[0107] In this embodiment, the second medical analysis model can be simply referred to as the model. The model parameters can be updated using reinforcement learning (RL) methods. This allows the model to maximize the accumulated feedback signal. ,in It is a discount factor. This refers to the time step. Then, policy-based or value-based reinforcement learning algorithms are used, such as Reinforce, Actor-Critic, Q-learning, or Deep Q-Network. Reinforce is a policy gradient-based update algorithm that uses the future total reward Gt instead of the Q-value; Actor-Critic is a reinforcement learning algorithm that combines value and policy functions; Q-learning is a value-based reinforcement learning algorithm; and Deep Q-network (DQN) is a highly practical reinforcement learning algorithm. This is achieved by using the model's output data... The behavior (action) of the model is considered as experimental results or expert opinions. This is considered a reward from the environment. Subsequently, the model's policy is defined based on different RL algorithms. OR value function or And use gradient descent to optimize the corresponding objective function. or or .
[0108] This application obtains experimental results and expert opinions corresponding to a pre-defined drug discovery task; then encodes the experimental results and expert opinions to obtain corresponding feedback information; subsequently, it optimizes the output data and feedback information of a second medical analysis model based on reinforcement learning methods to update the model parameters of the second medical analysis model, thereby obtaining a target medical analysis model. This application uses a pre-defined feedback learning method to improve the second medical analysis model based on the experimental results and expert opinions corresponding to the drug discovery task, thus obtaining an improved target medical analysis model. This achieves the goal and strategy of dynamically adjusting and optimizing the second medical analysis model, effectively improving the model accuracy and innovativeness of the generated target medical analysis model.
[0109] In one embodiment, an artificial intelligence-based medical data processing device is provided, which corresponds one-to-one with the artificial intelligence-based medical data processing methods described in the above embodiments. For example... Figure 3 As shown, the AI-based medical data processing device 100 includes a collection module 101, a first processing module 102, a training module 103, a second processing module 104, a third processing module 105, a fourth processing module 106, and an evaluation module 107. Detailed descriptions of each functional module are as follows:
[0110] Collection module 101 is used to collect initial medical data from multiple medical data sources;
[0111] The first processing module 102 is used to preprocess the initial medical data to obtain a medical sample set;
[0112] Training module 103 is used to train a preset multimodal basic model using the medical sample set based on a self-supervised learning method to obtain an initial medical analysis model; wherein the multimodal basic model is a model with an encoder and decoder architecture.
[0113] The second processing module 104 is used to fine-tune the initial medical analysis model through a preset multi-task learning strategy to obtain the first medical analysis model.
[0114] The third processing module 105 is used to optimize the first medical analysis model through a preset multi-objective optimization strategy to obtain a second medical analysis model;
[0115] The fourth processing module 106 is used to improve the second medical analysis model through a preset feedback learning method to obtain the improved target medical analysis model.
[0116] The evaluation module 107 is used to evaluate the target medical data to be processed based on the target medical analysis model, and generate a medical evaluation result corresponding to the target medical data.
[0117] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the product recommendation method in the aforementioned embodiments, and will not be repeated here.
[0118] In one embodiment, the first processing module 102 specifically includes:
[0119] The first processing submodule is used to clean the initial medical data to obtain the first medical data.
[0120] The second processing submodule is used to standardize the first medical data to obtain the second medical data.
[0121] The third processing submodule is used to encode the second medical data to obtain the third medical data;
[0122] The first determining submodule is used to use the third medical data as the medical sample set.
[0123] In one embodiment, the evaluation module 107 specifically includes:
[0124] The input submodule is used to input the target medical data into the input layer of the target medical analysis model and to perform embedding processing on the target medical data;
[0125] The encoding submodule is used to encode the target medical data through the encoder in the target medical analysis model to obtain the first node feature corresponding to the target medical data;
[0126] The fourth processing submodule is used to perform pooling and sampling processing on the first node features based on the latent space layer in the target medical analysis model to obtain the corresponding second node features;
[0127] The first update submodule is used to perform node update processing on the second node feature based on the decoder in the target medical analysis model to obtain the corresponding third node feature;
[0128] The evaluation submodule is used to evaluate the features of the third node based on the output layer of the target medical analysis model, and obtain the medical evaluation result corresponding to the target medical data.
[0129] In one embodiment, the encoding submodule specifically includes:
[0130] The acquisition unit is used to acquire the target data type of the target medical data;
[0131] The first determining unit is used to determine the target encoding method that matches the target data type;
[0132] The processing unit is used to encode the target medical data using the encoder in the target medical analysis model and the target encoding method to obtain encoded data corresponding to the target medical data.
[0133] The second determining unit is used to use the encoded data as the feature of the first node.
[0134] In one embodiment, the second processing module 104 specifically includes:
[0135] The first acquisition submodule is used to acquire the loss functions of various preset drug discovery tasks respectively;
[0136] The combination submodule is used to combine all the loss functions to obtain the corresponding total loss function;
[0137] The first optimization submodule is used to optimize the total loss function using the gradient descent method to obtain the optimized target total loss function;
[0138] The second update submodule is used to update the model parameters of the initial medical analysis model based on the target total loss function to obtain the first medical analysis model.
[0139] In one embodiment, the third processing module 105 includes:
[0140] The second acquisition submodule is used to acquire predefined task objectives corresponding to various drug discovery tasks.
[0141] The third acquisition submodule is used to acquire the objective function corresponding to each of the task objectives;
[0142] The second combination submodule is used to combine all the objective functions to obtain the corresponding total objective function;
[0143] The second optimization submodule is used to optimize the input data, output data, latent vectors, and overall objective function of the first medical analysis model based on a preset multi-objective evolutionary algorithm to obtain the second medical analysis model.
[0144] In one embodiment, the fourth processing module 106 includes:
[0145] The fourth acquisition submodule is used to acquire experimental results and expert opinions corresponding to the preset drug discovery tasks;
[0146] The fifth processing submodule is used to encode the experimental results and the expert opinions to obtain corresponding feedback information;
[0147] The sixth processing submodule is used to optimize the output data of the second medical analysis model and the feedback information based on the reinforcement learning method, so as to update the model parameters of the second medical analysis model and obtain the target medical analysis model.
[0148] This application provides an artificial intelligence-based medical data processing method and apparatus. It preprocesses initial medical data collected from medical data sources to obtain a medical sample set, then uses this sample set to train a pre-defined multimodal basic model to obtain an initial medical analysis model. Subsequently, it uses multi-task learning strategies, multi-objective optimization strategies, and feedback learning methods to optimize the initial medical analysis model, generating a target medical analysis model. This effectively improves the generalization ability, expressive power, balance and coordination, adaptability, and model performance of the generated target medical analysis model. The target medical analysis model is then used to evaluate the target medical data to be processed, achieving rapid and accurate generation of medical evaluation results corresponding to the target medical data. This effectively improves the efficiency and reduces the cost of evaluating and processing the target medical data. Specific limitations of the artificial intelligence-based medical data processing apparatus can be found in the limitations of the artificial intelligence-based medical data processing method described above, and will not be repeated here. Each module in the above-described artificial intelligence-based medical data processing apparatus can be implemented entirely or partially through software, hardware, or a combination thereof. The above modules can be embedded in the processor of a computer device in hardware form or independent of it, or they can be stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each of the above modules.
[0149] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side method for medical data processing based on artificial intelligence.
[0150] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing 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 stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements client-side functions or steps of an artificial intelligence-based medical data processing method.
[0151] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:
[0152] Initial medical data was collected from multiple medical data sources;
[0153] The initial medical data is preprocessed to obtain a medical sample set;
[0154] Based on the self-supervised learning method, the pre-set multimodal basic model is trained using the medical sample set to obtain an initial medical analysis model; wherein, the multimodal basic model is a model with an encoder and decoder architecture;
[0155] The initial medical analysis model is fine-tuned using a preset multi-task learning strategy to obtain the first medical analysis model;
[0156] The first medical analysis model is optimized using a pre-defined multi-objective optimization strategy to obtain a second medical analysis model;
[0157] The second medical analysis model is improved by using a pre-defined feedback learning method to obtain the improved target medical analysis model.
[0158] The target medical data to be processed is evaluated based on the target medical analysis model, and a medical evaluation result corresponding to the target medical data is generated.
[0159] In this embodiment, a medical sample set is obtained by preprocessing initial medical data collected from medical data sources. This sample set is then used to train a pre-defined multimodal base model, resulting in an initial medical analysis model. Subsequently, a multi-task learning strategy, a multi-objective optimization strategy, and a feedback learning method are used to optimize the initial medical analysis model, generating a target medical analysis model. This effectively improves the generalization ability, expressive power, balance and coordination, adaptability, and overall performance of the generated target medical analysis model. The target medical analysis model is then used to evaluate the target medical data to be processed, enabling the rapid and accurate generation of medical evaluation results corresponding to the target medical data. This effectively improves the efficiency of target medical data evaluation and processing while reducing the cost.
[0160] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0161] Initial medical data was collected from multiple medical data sources;
[0162] The initial medical data is preprocessed to obtain a medical sample set;
[0163] Based on the self-supervised learning method, the pre-set multimodal basic model is trained using the medical sample set to obtain an initial medical analysis model; wherein, the multimodal basic model is a model with an encoder and decoder architecture;
[0164] The initial medical analysis model is fine-tuned using a preset multi-task learning strategy to obtain the first medical analysis model;
[0165] The first medical analysis model is optimized using a pre-defined multi-objective optimization strategy to obtain a second medical analysis model;
[0166] The second medical analysis model is improved by using a pre-defined feedback learning method to obtain the improved target medical analysis model.
[0167] The target medical data to be processed is evaluated based on the target medical analysis model, and a medical evaluation result corresponding to the target medical data is generated.
[0168] In this embodiment, a medical sample set is obtained by preprocessing initial medical data collected from medical data sources. This sample set is then used to train a pre-defined multimodal base model, resulting in an initial medical analysis model. Subsequently, a multi-task learning strategy, a multi-objective optimization strategy, and a feedback learning method are used to optimize the initial medical analysis model, generating a target medical analysis model. This effectively improves the generalization ability, expressive power, balance and coordination, adaptability, and overall performance of the generated target medical analysis model. The target medical analysis model is then used to evaluate the target medical data to be processed, enabling the rapid and accurate generation of medical evaluation results corresponding to the target medical data. This effectively improves the efficiency of target medical data evaluation and processing while reducing the cost.
[0169] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.
[0170] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0171] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.
[0172] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A medical data processing method based on artificial intelligence, characterized in that, include: Initial medical data was collected from multiple medical data sources; The initial medical data is preprocessed to obtain a medical sample set; Based on a self-supervised learning method, the pre-defined multimodal basic model is trained using the medical sample set to obtain an initial medical analysis model. The multimodal basic model employs an encoder and decoder architecture. The encoder consists of multiple graph neural network (GNN) modules and transformer modules stacked alternately, used to encode the target medical data to obtain node features. The decoder also consists of multiple GNN modules and transformer modules stacked alternately, used to update nodes based on the features sampled from the latent space. The multimodal basic model further includes a latent space layer, used to perform pooling and sampling processing on the node features output by the encoder to obtain latent variables. The initial medical analysis model is fine-tuned using a preset multi-task learning strategy to obtain the first medical analysis model; The first medical analysis model is optimized using a pre-defined multi-objective optimization strategy to obtain a second medical analysis model; The second medical analysis model is improved by using a pre-defined feedback learning method to obtain the improved target medical analysis model. The target medical data to be processed is evaluated based on the target medical analysis model, and a medical evaluation result corresponding to the target medical data is generated.
2. The artificial intelligence-based medical data processing method as described in claim 1, characterized in that, The step of preprocessing the initial medical data to obtain a medical sample set includes: The initial medical data is cleaned to obtain the first medical data; The first medical data is standardized to obtain the second medical data. The second medical data is encoded to obtain the third medical data; The third medical data is used as the medical sample set.
3. The artificial intelligence-based medical data processing method as described in claim 1, characterized in that, The step of evaluating and processing the target medical data to be processed based on the target medical analysis model, and generating a medical evaluation result corresponding to the target medical data, includes: The target medical data is input into the input layer of the target medical analysis model, and the target medical data is embedded. The target medical data is encoded by the encoder in the target medical analysis model to obtain the first node feature corresponding to the target medical data; Based on the latent space layer in the target medical analysis model, the first node features are pooled and sampled to obtain the corresponding second node features. Based on the decoder in the target medical analysis model, the second node feature is updated to obtain the corresponding third node feature; The third node features are evaluated and processed based on the output layer of the target medical analysis model to obtain the medical evaluation results corresponding to the target medical data.
4. The artificial intelligence-based medical data processing method as described in claim 3, characterized in that, The step of encoding the target medical data using the encoder in the target medical analysis model to obtain the first node feature corresponding to the target medical data includes: Obtain the target data type of the target medical data; Determine the target encoding method that matches the target data type; The target medical data is encoded using the encoder in the target medical analysis model and the target encoding method to obtain encoded data corresponding to the target medical data. The encoded data is used as the feature of the first node.
5. The artificial intelligence-based medical data processing method as described in claim 1, characterized in that, The step of fine-tuning the initial medical analysis model using a preset multi-task learning strategy to obtain the first medical analysis model includes: Obtain the loss functions for various pre-defined drug discovery tasks; By combining all the aforementioned loss functions, the corresponding total loss function is obtained. The total loss function is optimized using gradient descent to obtain the optimized target total loss function. The model parameters of the initial medical analysis model are updated based on the target total loss function to obtain the first medical analysis model.
6. The artificial intelligence-based medical data processing method as described in claim 1, characterized in that, The step of optimizing the first medical analysis model using a preset multi-objective optimization strategy to obtain the second medical analysis model includes: Obtain predefined task objectives corresponding to various drug discovery tasks; Obtain the objective function corresponding to each of the stated task objectives; By combining all the aforementioned objective functions, the corresponding overall objective function is obtained. The second medical analysis model is obtained by optimizing the input data, output data, latent vectors, and overall objective function of the first medical analysis model based on a preset multi-objective evolutionary algorithm.
7. The artificial intelligence-based medical data processing method as described in claim 1, characterized in that, The step of improving the second medical analysis model using a preset feedback learning method to obtain the improved target medical analysis model includes: Obtain experimental results and expert opinions corresponding to the pre-set drug discovery tasks; The experimental results and expert opinions are encoded to obtain corresponding feedback information; The output data and feedback information of the second medical analysis model are optimized using reinforcement learning methods to update the model parameters of the second medical analysis model and obtain the target medical analysis model.
8. A medical data processing device based on artificial intelligence, characterized in that, include: The collection module is used to collect initial medical data from multiple medical data sources; The first processing module is used to preprocess the initial medical data to obtain a medical sample set; The training module is used to train a pre-defined multimodal base model using the medical sample set based on a self-supervised learning method, thereby obtaining an initial medical analysis model. The multimodal base model employs an encoder and decoder architecture. The encoder consists of multiple graph neural network (GNN) modules and transformer modules stacked alternately, used to encode the target medical data to obtain node features. The decoder also consists of multiple GNN modules and transformer modules stacked alternately, used to update nodes based on the features sampled from the latent space. The multimodal base model further includes a latent space layer, used to perform pooling and sampling processing on the node features output by the encoder to obtain latent variables. The second processing module is used to fine-tune the initial medical analysis model through a preset multi-task learning strategy to obtain the first medical analysis model. The third processing module is used to optimize the first medical analysis model through a preset multi-objective optimization strategy to obtain a second medical analysis model; The fourth processing module is used to improve the second medical analysis model through a preset feedback learning method to obtain the improved target medical analysis model. The evaluation module is used to evaluate the target medical data to be processed based on the target medical analysis model, and generate medical evaluation results corresponding to the target medical data.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence-based medical data processing method as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the artificial intelligence-based medical data processing method as described in any one of claims 1 to 7.