CN-122025181-A - Method, device, equipment and storage medium for predicting myasthenia gravis drug treatment effect

CN122025181ACN 122025181 ACN122025181 ACN 122025181ACN-122025181-A

Abstract

The application discloses a prediction method, a device, equipment and a storage medium for the therapeutic effect of myasthenia gravis drugs, and relates to the technical field of computer-aided medical diagnosis, wherein the method comprises the steps of acquiring multi-mode data of clinic, image, gene and drugs, and converting the multi-mode data into tokens sequences through preprocessing; calculating the contribution degree of each mode, screening an effective mode set, carrying out dynamic routing pair selection, bidirectional attention fusion and three-way attention integration through a hierarchical multi-mode interaction attention mechanism to generate initial fusion characteristics, carrying out global average pooling and vector splicing on the characteristics to obtain target fusion characteristics, and inputting the target fusion characteristics and longitudinal follow-up data into a time sequence causal model to obtain a treatment effect prediction result and causal chain interpretation information. The method effectively overcomes the defects of the existing method in time sequence modeling and black box decision, and can realize longitudinal dynamic prediction of the therapeutic effect of the myasthenia gravis drug and promote the interpretation of the prediction.

Inventors

HUANG HUA
Tao Qiaodie
CHEN YAN
YANG HUAN
ZHENG ZEKAI

Assignees

湖南工商大学

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (10)

1. A method of predicting the therapeutic effect of a myasthenia gravis drug, the method comprising: acquiring multi-modal data comprising clinical data, image data, gene data and drug data, and preprocessing the multi-modal data to obtain tokens sequences; calculating the contribution degree of each mode in the tokens sequences, and screening the tokens sequences according to the contribution degree to obtain an effective mode set; performing dynamic routing pair, bidirectional attention fusion and three-way attention integration on the effective modal set through a hierarchical multi-modal interaction attention mechanism to obtain initial fusion characteristics; Performing global average pooling and vector splicing on the initial fusion features to obtain target fusion features; And inputting the target fusion characteristics and the longitudinal follow-up data into a time sequence causal model to obtain causal chain interpretation information and a treatment effect prediction result, wherein the time sequence causal model is constructed according to a graph neural network and a structural causal model.
2. The method of claim 1, wherein the step of dynamically routing pairs, bi-directional attention fusion and tri-directional attention integration of the active modality set through a hierarchical multi-modality interactive attention mechanism to obtain initial fusion features comprises: Performing global average calculation on tokens sequences of each mode in the effective mode set to obtain global average representation of each mode; According to the global average representation of any two modes, a cosine similarity set is calculated, and two modes corresponding to the maximum value in the cosine similarity set are used as optimal mode pairs; Respectively carrying out linear transformation and dimension mapping on two modes in the optimal mode pair to obtain respective first query matrixes, first key matrixes and first value matrixes; Performing bidirectional fusion on the first query matrix, the first key matrix and the first value matrix through sparse self-attention and cross-modal attention to obtain a modal pair fusion representation; Performing linear transformation and dimension mapping on the fusion representation of the modal pairs and residual modalities respectively to obtain a second query matrix, a second key matrix and a second value matrix of each modality, wherein the residual modalities refer to the modalities except the optimal modal pair in the effective modal set; And three-way integration is carried out on the second query matrix, the second key matrix and the second value matrix through the sparse self-attention and the multi-group cross-modal attention, so that initial fusion characteristics are obtained.
3. The method of claim 1, wherein the step of constructing the time-series causal model comprises: determining causal directed graph nodes according to preset node types, and defining characteristic attributes corresponding to the nodes; Based on causal constraint of causal results, setting directed edges among the nodes, and setting weights of the directed edges as learnable parameters to form an initial causal directed graph; Determining node embedding updating rules of a causal graph rolling unit based on the node and edge structures of the initial causal graph, and setting normalization constants of aggregation neighbor nodes, learnable weights of edges between nodes and self-loop weights; Determining a parameterized function according to the node characteristic propagation logic of the graph neural network and the causal inference structure of the structural causal model; Taking the current moment characteristics, the historical curative effect state and the structural information of the initial causal directed graph as inputs, taking the predicted curative effect value at the next moment as output, and constructing a prediction function by combining the parameterized function; determining a classification unit based on the output dimension of the prediction function and a preset efficacy class, the classification unit being configured with the learnable parameters and a softmax function; Determining a causal interpretation unit based on the nodes and edge weight parameters of the initial causal directed graph; And integrating the initial causal directed graph, the causal graph convolution unit, the prediction function, the classification unit and the causal interpretation unit to obtain a time sequence causal model.
4. The method of claim 3, wherein the propagation formula of the node embedded update rule is: Wherein, the A feature embedding vector representing node v in layer i +1, The activation function is represented as a function of the activation, Representing the converged neighboring node, Representing the said normalization constant(s), The learnable weights representing the edges of node u to node v, Representing the weight of the self-loop, A feature embedding vector representing node u in the i-th layer, A feature embedding vector representing a node v in the i-th layer; The formula of the prediction function is as follows: Wherein, the Representing the predicted efficacy value at the next time, Representing the parameterized function, the parameterized function being a differentiable mapping function that jointly encodes the node feature propagation logic and the causal inference structure, the parameterized function internally comprising graph convolution operations, causal constraint activation, and temporal recurrence mechanisms, The feature embedding vector at time t is represented, And (5) representing the historical efficacy state at the time t, and G representing the structural information of the initial causal directed graph.
5. A method according to claim 3, wherein the time-series causal model comprises a causal graph convolution unit, an initial causal directed graph, a prediction function, a classification unit, and a causal interpretation unit; The step of inputting the target fusion characteristic and the longitudinal follow-up data into a time sequence causal model to obtain causal chain interpretation information and a treatment effect prediction result comprises the following steps of: Combining the target fusion characteristic with longitudinal follow-up data according to time sequence to obtain a multi-time-point input sequence, wherein the multi-time-point input sequence comprises characteristic information and historical curative effect states at all times; Distributing corresponding node attributes for the multi-time-point input sequence through the causal graph construction unit; inputting the multi-time point input sequence into the causal graph convolution unit, wherein the causal graph convolution unit is based on the node embedding updating rule, iteratively converging the neighbor node characteristics by combining the node and the edge structure of the initial causal directed graph, and updating the characteristic embedding vector of each node; The updated feature embedded vector, the historical curative effect state and the structural information of the initial causal directed graph are input into the prediction function to obtain a predicted curative effect value at the next moment; Inputting the predicted curative effect value at the next moment into the classification unit, dividing the classification unit according to a first preset threshold value and a second preset threshold value and descending amplitude of a symptom scale to obtain class results comprising three types of significant improvement, partial improvement and invalidation, and converting the class results into probability distribution of corresponding classes through a softmax function; inputting the edge weight information of the initial causal directed graph and the updated characteristic embedded vector into the causal interpretation unit, and resolving the weight of the edge and a causal path between nodes by the causal interpretation unit to generate causal chain interpretation information; Recursively calling the prediction function to perform multi-step iterative prediction on the categories corresponding to the significant improvement and the partial improvement, and generating a longitudinal curative effect trend in a time dimension; and integrating the category results, the probability distribution and the longitudinal curative effect trend to obtain a treatment effect prediction result.
6. The method of claim 1, wherein the steps of calculating a contribution of each modality in the tokens sequence, and screening the tokens sequence for an effective set of modalities based on the contribution comprise: Inputting each mode data in the tokens sequences into corresponding feature extraction units respectively to obtain initial feature vectors of each mode; Calculating the sensitivity of the initial feature vector to the comprehensive representation of the patient based on a preset loss function to obtain the contribution degree of each mode; marking the mode with the contribution degree lower than a preset contribution degree threshold as unavailable, and marking the mode with the contribution degree higher than the preset contribution degree threshold as available; and according to the tokens sequences corresponding to all the modalities marked as available, an effective modality set is formed.
7. The method of claim 6, wherein the contribution is calculated as: Wherein, the The degree of contribution is indicated by the degree of contribution, Representing the said preset loss function, Representing the initial feature vector.
8. A device for predicting the therapeutic effect of a myasthenia gravis drug, said device comprising: The preprocessing module is used for acquiring multi-mode data comprising clinical data, image data, gene data and drug data, and preprocessing the multi-mode data to obtain tokens sequences; the contribution calculation module is used for calculating the contribution of each mode in the tokens sequence, and screening the tokens sequence according to the contribution to obtain an effective mode set; The multi-mode fusion module is used for carrying out dynamic routing pair, bidirectional attention fusion and three-way attention integration on the effective mode set through a layering multi-mode interaction attention mechanism to obtain initial fusion characteristics; the feature output module is used for carrying out global average pooling and vector splicing on the initial fusion features to obtain target fusion features; The prediction module is used for inputting the target fusion characteristics and the longitudinal follow-up data into a time sequence causal model to obtain causal chain interpretation information and a treatment effect prediction result, and the time sequence causal model is constructed according to a graph neural network and a structural causal model.
9. A device for predicting the effect of a myasthenia gravis drug treatment, characterized in that it comprises a memory, a processor and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the method for predicting the effect of a myasthenia gravis drug treatment according to any one of claims 1 to 7.
10. A storage medium, characterized in that the storage medium is a computer-readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, realizes the steps of the method for predicting the effect of a myasthenia gravis drug treatment according to any one of claims 1 to 7.

Description

Method, device, equipment and storage medium for predicting myasthenia gravis drug treatment effect Technical Field The application relates to the technical field of computer-aided medical diagnosis, in particular to a method, a device, equipment and a storage medium for predicting the therapeutic effect of myasthenia gravis drugs. Background The current clinical judgment on the curative effect of the medicine mainly depends on clinical scale scores during follow-up, such as Quantitative myasthenia gravis Score (QMG) and myasthenia gravis daily life capacity scale (MYASTHENIA GRAVIS ACTIVITIES of DAILY LIVING SCALE, MG-ADL), and the method can reflect a certain curative effect change trend. In addition, some scholars have tried to introduce imaging data, immunological detection results or genetic data into efficacy evaluation and develop predictive studies in combination with statistical analysis or machine learning methods. The exploration improves the objectivity and accuracy of curative effect judgment to a certain extent, and provides auxiliary value for individuation treatment. However, the existing method still has the following defects that (1) in most researches, a prediction result still mainly depends on single-mode data, especially static analysis based on scale scores, clinical, image, gene and medicine multi-mode information cannot be effectively integrated, so that information utilization is insufficient, (2) the existing prediction method is mainly concentrated on effect judgment at a single treatment time point, lacks longitudinal modeling capability for curative effect evolution in a follow-up process, and is difficult to truly reflect long-term change trend of diseases, and (3) part of researches are introduced into machine learning or deep learning methods, but the model is mostly in a 'black box' characteristic, so that causal mechanisms behind the curative effect of medicines are difficult to explain, and clinical credibility is insufficient. Therefore, how to realize the longitudinal dynamic prediction of the therapeutic effect of myasthenia gravis drugs and to improve the interpretability of the prediction is a problem to be solved. Disclosure of Invention The application aims to provide a method, a device, equipment and a storage medium for predicting the therapeutic effect of myasthenia gravis drugs, and aims to solve the technical problem of how to realize longitudinal dynamic prediction of the therapeutic effect of myasthenia gravis drugs and improve the interpretation of the prediction. To achieve the above object, the present application provides a method for predicting the therapeutic effect of a myasthenia gravis drug, the method comprising: acquiring multi-modal data comprising clinical data, image data, gene data and drug data, and preprocessing the multi-modal data to obtain tokens sequences; calculating the contribution degree of each mode in the tokens sequences, and screening the tokens sequences according to the contribution degree to obtain an effective mode set; performing dynamic routing pair, bidirectional attention fusion and three-way attention integration on the effective modal set through a hierarchical multi-modal interaction attention mechanism to obtain initial fusion characteristics; Performing global average pooling and vector splicing on the initial fusion features to obtain target fusion features; And inputting the target fusion characteristics and the longitudinal follow-up data into a time sequence causal model to obtain causal chain interpretation information and a treatment effect prediction result, wherein the time sequence causal model is constructed according to a graph neural network and a structural causal model. In an embodiment, the step of obtaining initial fusion characteristics includes performing global average computation on tokens sequences of each mode in the active mode set to obtain global average representation of each mode, calculating a cosine similarity set according to the global average representation of any two modes, using two modes corresponding to the maximum value in the cosine similarity set as optimal mode pairs, performing linear transformation and dimension mapping on the two modes in the optimal mode pairs to obtain respective first query matrix, first key matrix and first value matrix, performing bidirectional fusion on the first query matrix, the first key matrix and the first value matrix through sparse self-attention and cross-mode attention, obtaining a mode pair fusion representation, performing linear transformation and dimension mapping on the fusion representation and the residual modes respectively to obtain a second query matrix, a second key matrix and a second value matrix of each mode, wherein the second query matrix and the second value matrix refer to the optimal mode pair, and the second value matrix are integrated in the sparse self-attention matrix and the initial value matrix. In an embodiment, the t