Bladder irrigation treatment effect prediction method and system based on multi-modal data

Abstract

The application provides a bladder perfusion treatment effect prediction method and system based on multi-modal data, relates to the technical field of multi-modal data, and comprises the following steps: acquiring image, test and pathological text data of a patient and extracting features, generating mode-invariant pure efficacy features by using a multi-modal anti-entanglement network, training a prediction model based on the pure efficacy features to obtain a predicted efficacy value, simulating efficacy changes under different perfusion parameter combinations by using a counterfactual reasoning engine, calculating parameter contribution degrees, and generating an individualized perfusion scheme. The application can effectively fuse multi-source heterogeneous data, eliminate mode bias, improve prediction accuracy, and provide a quantitative basis for formulating accurate individualized treatment schemes in clinical practice.

Description

Technical Field

[0001] This invention relates to multimodal data technology, and more particularly to a method and system for predicting the therapeutic effect of bladder instillation based on multimodal data. Background Technology

[0002] In the clinical treatment of bladder cancer, intravesical instillation therapy is an important postoperative adjuvant treatment aimed at eliminating residual tumor cells and reducing the risk of recurrence by instilling chemotherapy drugs or immunotherapies into the bladder. Currently, to assess and predict patient response to instillation therapy, clinical practice mainly relies on the analysis of single-modality or simply fused multimodal patient data. A common approach is to separately utilize medical imaging data, clinical laboratory reports, and pathological descriptions, employing independent analytical models or expert experience to assess tumor morphology, biomarker levels, and pathological grading, thereby comprehensively determining prognosis and developing treatment plans. These routine methods constitute the technical foundation for predicting efficacy and guiding treatment plans in this field.

[0003] However, these conventional approaches have significant limitations. A major drawback is the significant modal heterogeneity of data from different sources. For example, the pixel matrix of images, numerical sequences of tests, and descriptive language of text differ greatly in feature distribution. Existing methods typically employ simple feature concatenation or post-decision fusion, failing to effectively remove modality-specific noise from the data that is irrelevant to treatment efficacy. This results in extracted feature representations being contaminated with excessive modality bias. Consequently, the correlations learned by the model do not truly reflect the biological laws governing efficacy but instead rely excessively on the statistical characteristics of specific modalities, thus weakening the model's generalization ability and predictive robustness. Another drawback is that treatment optimization is often based on population statistical patterns or physician trial-and-error, lacking data-driven, systematic simulations tailored to individual patients. Conventional methods struggle to quantify the efficacy changes of different treatment parameter adjustments on specific patients and cannot proactively generate and compare multiple virtual intervention scenarios, thus falling short in achieving truly personalized and precise treatment parameter recommendations. Summary of the Invention

[0004] This invention provides a method and system for predicting the therapeutic effect of bladder instillation based on multimodal data, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for predicting the therapeutic effect of bladder instillation based on multimodal data, comprising: Acquire the patient's medical imaging data, clinical laboratory data, and pathological text data, and extract features from the medical imaging data, clinical laboratory data, and pathological text data respectively to obtain image feature vectors, laboratory feature vectors, and text feature vectors. A multimodal adversarial unentanglement network is constructed. By introducing a modality discriminator, the image feature vector, the test feature vector, and the text feature vector are trained adversarially. This enables the feature representation to retain efficacy-related information while eliminating the discriminability of modality sources, thereby generating modality-invariant pure efficacy features. The pure efficacy features are input into the prediction model for training. A mapping relationship is established between the pure efficacy features of historical patients and their corresponding real efficacy labels to obtain the trained prediction model. The predicted efficacy value of the current patient is output based on the trained prediction model. A counterfactual reasoning engine is constructed, which takes the pure efficacy characteristics and the predicted efficacy value as the baseline input to generate multiple virtual intervention scenarios. Based on each virtual intervention scenario, the efficacy change results are simulated under different combinations of infusion drug ratios, infusion time intervals and infusion doses. The efficacy changes in each virtual intervention scenario are compared with the predicted efficacy values. The counterfactual reasoning engine calculates the contribution of each parameter combination to the efficacy improvement, selects the parameter combination with the highest contribution as the optimization target, and generates an individualized perfusion plan based on the optimization target.

[0006] By introducing a modality discriminator to perform adversarial training on the image feature vector, the test feature vector, and the text feature vector, the feature representation can eliminate the discriminability of modality origin while retaining efficacy-related information, generating modality-invariant pure efficacy features, including: The image feature vector, the test feature vector, and the text feature vector are respectively input into the feature generator for encoding and conversion to obtain candidate features of the same dimension. The candidate features are then input into the modality discriminator for modality source identification. The modality confusion loss is calculated based on the modality classification results output by the modality discriminator. In the adversarial training process, a efficacy retention constraint is introduced. The correlation strength between the candidate features and the historical patient efficacy labels is calculated. The correlation strength is used as a quantitative indicator of the degree of efficacy information retention. A joint optimization objective function is constructed to maximize the modality confusion loss while maximizing the correlation strength. The parameters of the feature generator are updated by backpropagation, so that the candidate features maintain a strong correlation with the therapeutic effect while gradually losing modal discriminability. When the modal discriminator’s modal recognition accuracy for the candidate features drops below a preset accuracy threshold and the correlation strength stabilizes within a preset range, the candidate features at this time are determined as pure therapeutic effect features with unchanged modality.

[0007] Using the correlation strength as a quantitative indicator of the degree of retention of therapeutic information, a joint optimization objective function is constructed to maximize both the modality confusion loss and the correlation strength, including: The correlation strength is used as a quantitative indicator of the degree of retention of efficacy information. A joint optimization objective function is constructed based on the quantitative indicator. The joint optimization objective function includes a modality confusion loss term and an efficacy retention loss term. Positive optimization weights are configured for the modality confusion loss term to drive the feature generator to update parameters in the direction of maximizing the modality confusion loss. At the same time, positive optimization weights are configured for the efficacy retention loss term to drive the feature generator to update parameters in the direction of maximizing the correlation strength. By synchronously adjusting the parameters of the feature generator through a gradient optimization algorithm, the candidate features continuously reduce modality discriminability and continuously enhance efficacy relevance during the iteration process, thereby maximizing both the modality confusion loss and the relevance strength.

[0008] A mapping relationship is established between the pure efficacy characteristics of historical patients and their corresponding real efficacy labels to obtain a trained prediction model. Based on the trained prediction model, the predicted efficacy value of the current patient is output, including: A training sample set is constructed by extracting pure efficacy features and real efficacy labels from a historical patient database. Cluster analysis is performed on the training sample set to identify patient subgroups with different efficacy response patterns. A dedicated sub-mapping model is constructed for each patient subgroup. A hierarchical mapping architecture is formed by integrating all sub-mapping models. A dynamic sample weighting mechanism is introduced during training to calculate the prediction residual of each training sample. The weight contribution of the corresponding sample in subsequent iterations is adaptively adjusted according to the prediction residual. Samples with prediction residuals are given weights to enhance the learning ability of difficult examples. Through iterative optimization, the hierarchical mapping architecture achieves stable prediction performance under different therapeutic response modes, and the trained prediction model is obtained. The pure efficacy characteristics of the current patient are input into the trained prediction model. The efficacy response subgroup to which the current patient belongs is identified by the patient subgroup attribution determination module. The sub-mapping model corresponding to the efficacy response subgroup is called and the predicted efficacy value of the current patient is calculated and output.

[0009] A counterfactual reasoning engine is constructed, using the pure efficacy features and the predicted efficacy values ​​as baseline inputs, to generate multiple virtual intervention scenarios, including: A counterfactual reasoning engine is constructed, which includes a structural causal model and a condition generator. The structural causal model learns the causal mechanism of infusion drug ratio, infusion time interval and infusion dose on efficacy by analyzing historical treatment data, and establishes a causal inference function from interventionable variables to changes in efficacy. Pure therapeutic efficacy features are input into the condition generator as the patient's inherent state benchmark. An intervention parameter adjustment space is constructed based on the causal inference function. Parameter adjustment combinations are sampled in the intervention parameter adjustment space. The perturbation effect of the parameter adjustment combinations on the pure therapeutic efficacy features is calculated through the causal inference function. The perturbation effect is superimposed on the pure therapeutic efficacy features to generate a virtual patient characteristic state. The predicted therapeutic effect value is associated with each virtual patient's characteristic state, and each virtual patient's characteristic state is labeled with its corresponding parameter adjustment combination and the predicted therapeutic effect value to form a virtual intervention scenario.

[0010] The results of efficacy changes simulated under different combinations of infusion drug ratios, infusion time intervals, and infusion doses for each of the aforementioned virtual intervention scenarios include: The virtual patient's characteristic state and corresponding parameter adjustment combination are extracted from each virtual intervention scenario to construct a therapeutic effect evolution simulation module. According to the therapeutic effect evolution simulation module, the changes in the infusion drug ratio, infusion time interval and infusion dose in the parameter adjustment combination are transformed into a perturbation vector in the feature space through the causal intervention propagation mechanism. The perturbation vector is propagated along the causal path of the virtual patient's characteristic state to simulate the physiological response cascade reaction caused by the adjustment of treatment parameters and generate a time-series therapeutic effect evolution curve. A parameter sensitivity decomposition algorithm is introduced to decompose the time-series efficacy evolution curve into multiple scales, extract the rapid response component and the delayed response component, calculate the contribution weight of each response component to the final efficacy value, and obtain the comprehensive efficacy prediction value of the corresponding virtual intervention scenario through weight reconstruction. The efficacy improvement index is calculated by comparing the comprehensive efficacy estimate with the predicted efficacy value. The contribution weight and the efficacy improvement index are combined to form the efficacy change result of the corresponding virtual intervention scenario.

[0011] The counterfactual reasoning engine calculates the contribution of each parameter combination to the improvement of therapeutic effect, selects the parameter combination with the highest contribution as the optimization target, and generates an individualized perfusion plan based on the optimization target, including: The efficacy change results of all virtual intervention scenarios are input into the counterfactual reasoning engine for contribution quantification analysis. Based on the counterfactual reasoning engine, the efficacy improvement index and corresponding parameter combination in each efficacy change result are extracted. The improvement of the efficacy improvement index relative to the predicted efficacy value is calculated. A causal attribution decomposition mechanism is introduced to decompose the improvement into the marginal contribution component and synergistic contribution component of the infusion drug ratio, infusion time interval and infusion dose. The contribution of the corresponding parameter combination to the efficacy improvement is obtained by weighted aggregation of the marginal contribution component and the synergistic contribution component. The contribution of all parameter combinations is sorted in descending order, and the parameter combination with the highest contribution is identified as the optimization target. The optimization target is then input into the scheme conversion module. Based on the scheme conversion module, the specific drug types and their proportions are determined according to the target setting value of the infusion drug ratio, the treatment cycle schedule is planned according to the target setting value of the infusion time interval, and the single-dose dosage specification is formulated according to the target setting value of the infusion dose, thus integrating them into an individualized infusion scheme that includes drug configuration scheme, treatment schedule and dosage specification.

[0012] A second aspect of the present invention provides a bladder instillation therapy efficacy prediction system based on multimodal data, comprising: The data acquisition unit is used to acquire the patient's medical imaging data, clinical test data, and pathological text data, and to extract features from the medical imaging data, the clinical test data, and the pathological text data to obtain image feature vectors, test feature vectors, and text feature vectors. The feature extraction unit is used to construct a multimodal adversarial unentanglement network. By introducing a modality discriminator, the image feature vector, the test feature vector, and the text feature vector are adversarially trained, so that the feature representation can eliminate the discriminability of modality source while retaining efficacy-related information, and generate modality-invariant pure efficacy features. The feature resolution unit is used to input the pure efficacy features into the prediction model for training, establish a mapping relationship between the pure efficacy features of historical patients and their corresponding real efficacy labels to obtain the trained prediction model, and output the predicted efficacy value of the current patient based on the trained prediction model. The model training unit is used to build a counterfactual reasoning engine. It takes the pure efficacy features and the predicted efficacy values ​​as benchmark inputs to generate multiple virtual intervention scenarios. Based on each virtual intervention scenario, it simulates the efficacy changes under different combinations of infusion drug ratios, infusion time intervals, and infusion doses. The counterfactual reasoning unit is used to compare the changes in efficacy in each virtual intervention scenario with the predicted efficacy value, calculate the contribution of each parameter combination to the improvement of efficacy through the counterfactual reasoning engine, select the parameter combination with the highest contribution as the optimization target, and generate an individualized perfusion plan based on the optimization target.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] The beneficial effects of this application are as follows: This method effectively integrates multimodal data such as medical imaging, clinical laboratory tests, and pathological texts, and obtains deep feature representations for each modality through feature extraction techniques. These feature vectors comprehensively cover the morphological information, physiological and biochemical indicators, and pathological descriptions of tumors, providing a rich and complementary information foundation for subsequent efficacy prediction and avoiding the one-sidedness and limitations of a single data source.

[0016] By constructing a multimodal adversarial unentanglement network and introducing a modality discriminator for adversarial training, the network can be forced to learn feature representations that are highly correlated with therapeutic efficacy but independent of specific data modalities. This process eliminates redundant information related to the modality source in the features, generating modality-invariant, pure therapeutic efficacy features. These features effectively remove interference from non-clinical factors such as data acquisition equipment, testing methods, or textual description habits, enabling subsequent prediction models to focus on key factors essentially related to treatment efficacy, thus improving the model's generalization ability and robustness.

[0017] By training a predictive model using the pure efficacy characteristics and actual efficacy labels of historical patients, an accurate mapping relationship can be established from patient multimodal characteristics to treatment outcomes. The trained model can output objective and quantitative predicted efficacy values ​​based on the current patient's pure efficacy characteristics. This prediction provides important data support for clinical decision-making, helping physicians to pre-assess potential efficacy before treatment, thereby enabling early identification and focused attention of high-risk patients. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the method for predicting the therapeutic effect of bladder instillation based on multimodal data, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating the generation process of an individualized infusion scheme based on a counterfactual reasoning engine, as described in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0021] Figure 1 This is a flowchart illustrating the method for predicting the therapeutic effect of bladder instillation based on multimodal data according to an embodiment of the present invention. Figure 1 As shown, the method includes: Acquire the patient's medical imaging data, clinical laboratory data, and pathological text data, and extract features from the medical imaging data, clinical laboratory data, and pathological text data respectively to obtain image feature vectors, laboratory feature vectors, and text feature vectors. A multimodal adversarial unentanglement network is constructed. By introducing a modality discriminator, the image feature vector, the test feature vector, and the text feature vector are trained adversarially. This enables the feature representation to retain efficacy-related information while eliminating the discriminability of modality sources, thereby generating modality-invariant pure efficacy features. The pure efficacy features are input into the prediction model for training. A mapping relationship is established between the pure efficacy features of historical patients and their corresponding real efficacy labels to obtain the trained prediction model. The predicted efficacy value of the current patient is output based on the trained prediction model. A counterfactual reasoning engine is constructed, which takes the pure efficacy characteristics and the predicted efficacy value as the baseline input to generate multiple virtual intervention scenarios. Based on each virtual intervention scenario, the efficacy change results are simulated under different combinations of infusion drug ratios, infusion time intervals and infusion doses. The efficacy changes in each virtual intervention scenario are compared with the predicted efficacy values. The counterfactual reasoning engine calculates the contribution of each parameter combination to the efficacy improvement, selects the parameter combination with the highest contribution as the optimization target, and generates an individualized perfusion plan based on the optimization target.

[0022] In one optional implementation, adversarial training is performed on the image feature vector, the test feature vector, and the text feature vector by introducing a modality discriminator. This allows the feature representation to retain efficacy-related information while eliminating the discriminability of modality origin, generating modality-invariant pure efficacy features, including: The image feature vector, the test feature vector, and the text feature vector are respectively input into the feature generator for encoding and conversion to obtain candidate features of the same dimension. The candidate features are then input into the modality discriminator for modality source identification. The modality confusion loss is calculated based on the modality classification results output by the modality discriminator. In the adversarial training process, a efficacy retention constraint is introduced. The correlation strength between the candidate features and the historical patient efficacy labels is calculated. The correlation strength is used as a quantitative indicator of the degree of efficacy information retention. A joint optimization objective function is constructed to maximize the modality confusion loss while maximizing the correlation strength. The parameters of the feature generator are updated by backpropagation, so that the candidate features maintain a strong correlation with the therapeutic effect while gradually losing modal discriminability. When the modal discriminator’s modal recognition accuracy for the candidate features drops below a preset accuracy threshold and the correlation strength stabilizes within a preset range, the candidate features at this time are determined as pure therapeutic effect features with unchanged modality.

[0023] In the implementation of the multimodal adversarial unentanglement network, the feature generator adopts a fully connected neural network architecture, with independent encoding branches for image feature vectors, test feature vectors, and text feature vectors. Image feature vectors are typically 512-dimensional, derived from the last layer's global average pooling output of convolutional neural networks such as ResNet or DenseNet; test feature vectors are 128-dimensional, containing numerical indicators such as blood routine tests, urine routine tests, and tumor markers; and text feature vectors are 768-dimensional, generated by encoding pathology reports using the BERT model. Each branch encoder contains two fully connected layers, with 256 nodes in the intermediate layer, using ReLU activation, and the output layer maps the three modalities to a unified 128-dimensional candidate feature space.

[0024] The modality discriminator employs a three-class classification architecture, taking 128-dimensional candidate features as input and outputting probability distributions for the three modality classes through a three-layer fully connected network. The modality confusion loss is calculated using the negative form of cross-entropy, specifically... , where p_i is the one-hot encoding of the true modality and q_i is the predicted probability output by the discriminator. When the discriminator cannot accurately identify the source of the feature, this loss value approaches log(3), indicating that the candidate feature has eliminated modality specificity.

[0025] The efficacy maintenance constraint is achieved by calculating the Pearson correlation coefficient between candidate features and efficacy labels. A dimension-wise correlation analysis is performed between the candidate feature vector and the binary efficacy labels of historical patients (1 for effective, 0 for ineffective), resulting in 128 correlation coefficients. The average of their absolute values ​​is taken as the correlation strength index. The joint optimization objective function is designed as follows: Where L_correlation is the negative correlation strength, L_prediction is the binary cross-entropy loss of the prediction model, and the weight coefficients are... , , They were set to 0.3, 0.5, and 0.2 respectively.

[0026] Adversarial training employs an alternating optimization strategy. In each training batch, the feature generator parameters are first fixed, and the modality discriminator is updated five times using the standard classification loss to improve its modality recognition ability. Subsequently, the discriminator parameters are fixed, and the feature generator is updated once using the joint optimization objective function to cause its generated candidate features to confuse the modality source. An early stopping mechanism is implemented during training. When the modality discriminator's three-class classification accuracy on the validation set is below 0.4 (a preset accuracy threshold) for ten consecutive epochs, and the correlation strength is stable within the range of 0.65 to 0.75 (a preset range), the adversarial training is considered to have converged. At this point, the candidate features output by the feature generator have lost modality discriminability while retaining their essential association with the efficacy of bladder instillation, and are thus identified as pure efficacy features. The entire network uses the Adam optimizer with a learning rate of 0.0001 and a batch size of 32. Iterations are performed on a training set containing data from 800 historical patients for 150 to 200 epochs to achieve convergence.

[0027] In one optional implementation, the correlation strength is used as a quantitative indicator of the degree of retention of therapeutic information. Constructing a joint optimization objective function to maximize both the modality confusion loss and the correlation strength includes: The correlation strength is used as a quantitative indicator of the degree of retention of efficacy information. A joint optimization objective function is constructed based on the quantitative indicator. The joint optimization objective function includes a modality confusion loss term and an efficacy retention loss term. Positive optimization weights are configured for the modality confusion loss term to drive the feature generator to update parameters in the direction of maximizing the modality confusion loss. At the same time, positive optimization weights are configured for the efficacy retention loss term to drive the feature generator to update parameters in the direction of maximizing the correlation strength. By synchronously adjusting the parameters of the feature generator through a gradient optimization algorithm, the candidate features continuously reduce modality discriminability and continuously enhance efficacy relevance during the iteration process, thereby maximizing both the modality confusion loss and the relevance strength.

[0028] After the modality discriminator performs an initial evaluation of candidate features, a quantification mechanism is needed to ensure that the feature extraction process eliminates modality differences while retaining the predictive value of therapeutic efficacy. Specifically, the Pearson correlation coefficient is used to calculate the degree of linear association between candidate features and the true therapeutic efficacy label. This coefficient ranges from -1 to 1, with an absolute value closer to 1 indicating more sufficient retention of therapeutic efficacy information. The absolute value of the calculated correlation coefficient is defined as the correlation strength, serving as the core indicator for quantifying the degree of retention of therapeutic efficacy information.

[0029] When constructing the joint optimization objective function, a composite loss structure with two core components is designed. The first part is the modality confusion loss term, represented by the negative value of the cross-entropy loss output by the modality discriminator. The larger the value, the more difficult it is for the discriminator to distinguish the modality from which the feature originates. The second part is the efficacy retention loss term, defined as the difference between 1 and the correlation strength. When this difference approaches 0, it indicates that the association between the feature and the efficacy label has reached the optimal state. The two loss terms are integrated into a unified objective function through a weighted summation. The weight coefficients are dynamically adjusted based on the performance on the validation set. In a typical setting, the modality confusion weight is 0.4, and the efficacy retention weight is 0.6.

[0030] The parameter update strategy employs a bidirectional driving mechanism. A positive optimization coefficient is configured for the modality confusion loss term, causing the feature generator to adjust network weights along the direction that increases this loss value during backpropagation. Specifically, this manifests as the generated candidate features gradually exhibiting a uniform distribution across modalities. Simultaneously, a negative gradient propagation path is set for the efficacy preservation loss term, driving the generator to enhance the numerical components in the feature vector related to efficacy outcomes. In the high-dimensional feature space, this is reflected in the continuously shortening projection distance of candidate features onto the direction corresponding to the true efficacy label.

[0031] The Adam optimization algorithm was used to perform synchronous parameter adjustments, with a learning rate of 0.001 and a momentum parameter of 0.9. In each iteration, the feature generator received the comprehensive gradient signal from the joint objective function and calculated the partial derivatives of the two loss terms with respect to the parameters of each layer of the network using the chain rule. A gradient pruning mechanism was introduced during the gradient update process to limit the gradient norm to within 5, preventing numerical oscillations caused by adversarial training due to modality confusion loss. After 300 to 500 training epochs, the distribution of candidate features in the feature space exhibited modality boundary ablation and enhanced efficacy discrimination. At this point, the classification accuracy of the modality discriminator dropped to around 55%, approaching the level of random guessing, while the efficacy correlation strength stabilized above 0.82, achieving the dual optimization goal of modality confusion and efficacy preservation.

[0032] In one optional implementation, a mapping relationship is established between the pure efficacy characteristics of historical patients and their corresponding real efficacy labels to obtain a trained prediction model. The predicted efficacy value for the current patient is then output based on the trained prediction model, including: A training sample set is constructed by extracting pure efficacy features and real efficacy labels from a historical patient database. Cluster analysis is performed on the training sample set to identify patient subgroups with different efficacy response patterns. A dedicated sub-mapping model is constructed for each patient subgroup. A hierarchical mapping architecture is formed by integrating all sub-mapping models. A dynamic sample weighting mechanism is introduced during training to calculate the prediction residual of each training sample. The weight contribution of the corresponding sample in subsequent iterations is adaptively adjusted according to the prediction residual. Samples with prediction residuals are given weights to enhance the learning ability of difficult examples. Through iterative optimization, the hierarchical mapping architecture achieves stable prediction performance under different therapeutic response modes, and the trained prediction model is obtained. The pure efficacy characteristics of the current patient are input into the trained prediction model. The efficacy response subgroup to which the current patient belongs is identified by the patient subgroup attribution determination module. The sub-mapping model corresponding to the efficacy response subgroup is called and the predicted efficacy value of the current patient is calculated and output.

[0033] To leverage the multimodal feature data accumulated in the historical patient database, a hierarchical mapping architecture is first constructed. Historical cases with completed perfusion therapy and clear efficacy evaluations are extracted from the database. For each patient, a purified efficacy feature vector (typically 128 to 256 dimensions) generated after processing by a multimodal adversarial deentanglement network is paired with its corresponding true efficacy label. The efficacy label uses a five-level classification system, including five levels: markedly effective, effective, stable, progressive, and ineffective. The K-means clustering algorithm is applied to the constructed training sample set, with the number of clusters K set to 3 to 5. The clustering quality is evaluated by calculating the silhouette coefficient, classifying patients into high-response, moderate-response, and low-response subgroups. For each identified patient subgroup, a sub-mapping model based on a gradient boosting decision tree is constructed, with a model depth of 6 to 10 layers and a learning rate of 0.01 to 0.05. Each sub-model is trained independently to capture the efficacy response patterns of a specific subgroup.

[0034] A dynamic sample weight adjustment mechanism is introduced during the training phase. Initially, all training sample weights are set to 1.0. After the first round of training, the prediction residual for each sample is calculated, defined as the absolute difference between the predicted efficacy value and the true efficacy label. For difficult samples with residuals greater than 0.5, their weights are adjusted according to the formula... An improvement is performed, where r_i represents the normalized residual of the i-th sample, and α is an adjustment coefficient ranging from 1.5 to 2.5. In subsequent iterations, high-weight samples contribute more to the loss function calculation, prompting the model to focus on optimizing its fitting ability to difficult examples. After 15 to 30 rounds of training, when the mean absolute error on the validation set decreases by less than 0.001 for five consecutive rounds, the hierarchical mapping architecture is considered to have converged, and the parameter combinations of each sub-mapping model are saved to form the final post-trained prediction model.

[0035] When predicting the efficacy of treatment for current patients, the patient's pure efficacy features are input into the subgroup classification module. This module uses a nearest neighbor classification strategy to calculate the Euclidean distance between the current patient's feature vector and the centroids of each patient subgroup, selecting the subgroup with the smallest distance as the classification category. After determining the classification, the corresponding submapping model is invoked to perform forward propagation. The model output is a continuous value between 0 and 1, representing the probability estimate of the effectiveness of perfusion therapy. This value is mapped to a five-level efficacy classification system to generate the final predicted efficacy value, while also outputting the predicted confidence interval, providing a quantitative reference for clinical decision-making.

[0036] In one alternative implementation, a counterfactual reasoning engine is constructed, using the pure efficacy features and the predicted efficacy values ​​as baseline inputs to generate multiple virtual intervention scenarios, including: A counterfactual reasoning engine is constructed, which includes a structural causal model and a condition generator. The structural causal model learns the causal mechanism of infusion drug ratio, infusion time interval and infusion dose on efficacy by analyzing historical treatment data, and establishes a causal inference function from interventionable variables to changes in efficacy. Pure therapeutic efficacy features are input into the condition generator as the patient's inherent state benchmark. An intervention parameter adjustment space is constructed based on the causal inference function. Parameter adjustment combinations are sampled in the intervention parameter adjustment space. The perturbation effect of the parameter adjustment combinations on the pure therapeutic efficacy features is calculated through the causal inference function. The perturbation effect is superimposed on the pure therapeutic efficacy features to generate a virtual patient characteristic state. The predicted therapeutic effect value is associated with each virtual patient's characteristic state, and each virtual patient's characteristic state is labeled with its corresponding parameter adjustment combination and the predicted therapeutic effect value to form a virtual intervention scenario.

[0037] The counterfactual reasoning engine is built upon causal inference theory, employing a structural causal model to describe the causal dependency between treatment parameters and therapeutic efficacy. In its implementation, the structural causal model uses a directed acyclic graph to represent the causal path between treatment parameter nodes and efficacy nodes. By analyzing the infusion drug ratios, infusion intervals, infusion doses, and corresponding therapeutic outcomes of patients in historical treatment datasets, structural equation modeling is used to learn the direct causal strength and indirect transmission paths of each parameter on therapeutic efficacy. This model introduces exogenous variables to control for confounding factors, ensuring that the identified causal relationships are not affected by selection bias in the observational data.

[0038] The condition generator employs a variational autoencoder architecture, embedding pure therapeutic features as conditional inputs into the latent space. This feature vector carries immutable fundamental information such as the patient's inherent physiological state and tumor characteristics, remaining unchanged during subsequent parameter interventions. Based on the causal inference function learned from the structural causal model, a three-dimensional intervention parameter adjustment space is constructed, encompassing the range of infusion drug ratios, candidate time intervals, and dose adjustment magnitudes. Specifically, the ratio range can be set as a continuous interval from 1:0 to 0:1 for the ratio of mitomycin to pirarubicin; the time interval covers candidate regimens from weekly to monthly; and the dose adjustment magnitude is set as a variation range of ±30% relative to the standard dose.

[0039] In the intervention parameter adjustment space, a Latin hypercube sampling method is used to generate parameter adjustment combinations, ensuring that the sampling points are uniformly distributed in the parameter space. For each sampled parameter adjustment combination, its perturbation effect on the pure efficacy characteristics is calculated using a causal inference function. This perturbation effect is represented as an offset vector in the feature space, the direction and magnitude of which are determined by the causal action coefficients and interaction effects of each parameter in the causal model. The perturbation effect vector is then vector-added with the original pure efficacy characteristics to obtain a virtual patient characteristic state reflecting the parameter adjustments. This characteristic state simulates the characteristic distribution exhibited by the patient under hypothetical intervention.

[0040] To establish a complete virtual intervention scenario, the predicted efficacy values ​​output by the prediction model are structurally correlated with each virtual patient's characteristic state. Each virtual characteristic state is labeled with triplet information, including the corresponding parameter adjustment combination, the original predicted efficacy value, and the simulated efficacy value after intervention. The post-intervention efficacy value is obtained by re-inputting the virtual patient's characteristic state into the trained prediction model, reflecting the expected treatment effect under that parameter combination. All labeled virtual characteristic states and their associated information together constitute the virtual intervention scenario set, providing a decision-making basis for subsequent parameter optimization. The entire process ensures the causal interpretability of counterfactual reasoning, enabling the generated virtual scenario to accurately simulate the effects of real clinical interventions.

[0041] In one optional implementation, the simulation results of efficacy changes under different combinations of infusion drug ratios, infusion time intervals, and infusion doses for each virtual intervention scenario include: The virtual patient's characteristic state and corresponding parameter adjustment combination are extracted from each virtual intervention scenario to construct a therapeutic effect evolution simulation module. According to the therapeutic effect evolution simulation module, the changes in the infusion drug ratio, infusion time interval and infusion dose in the parameter adjustment combination are transformed into a perturbation vector in the feature space through the causal intervention propagation mechanism. The perturbation vector is propagated along the causal path of the virtual patient's characteristic state to simulate the physiological response cascade reaction caused by the adjustment of treatment parameters and generate a time-series therapeutic effect evolution curve. A parameter sensitivity decomposition algorithm is introduced to decompose the time-series efficacy evolution curve into multiple scales, extract the rapid response component and the delayed response component, calculate the contribution weight of each response component to the final efficacy value, and obtain the comprehensive efficacy prediction value of the corresponding virtual intervention scenario through weight reconstruction. The efficacy improvement index is calculated by comparing the comprehensive efficacy estimate with the predicted efficacy value. The contribution weight and the efficacy improvement index are combined to form the efficacy change result of the corresponding virtual intervention scenario.

[0042] In each virtual intervention scenario generated by the counterfactual reasoning engine, the scenario parameters are first parsed to obtain the virtual patient's characteristic state and the corresponding parameter adjustment combination. The virtual patient's characteristic state consists of a pure efficacy feature vector, representing the patient's current physiological and pathological state. The parameter adjustment combination includes specific values ​​for three dimensions: infusion drug ratio, infusion time interval, and infusion dose.

[0043] The efficacy evolution simulation module employs a causal intervention propagation mechanism to map parameter changes to the feature space. This mechanism normalizes the change in each parameter in the parameter adjustment combination, converting changes in the infusion drug ratio into a ratio offset coefficient, changes in the infusion time interval into a time scale factor, and changes in the infusion dose into a dose intensity coefficient. Using a pre-established causal propagation matrix, these three coefficients are encoded into a high-dimensional perturbation vector, the dimension of which is consistent with the pure efficacy feature vector.

[0044] The causal propagation process unfolds based on the pathological mechanism of bladder instillation therapy. The perturbation vector first acts on the feature dimensions related to drug absorption, triggering local feature responses. Subsequently, it propagates downstream through a predefined causal relationship map, gradually influencing immune response features, tissue repair features, and tumor suppression features. The propagation process employs a decay function to control the intensity of the influence, ensuring that the perturbation effect conforms to physiological laws. After multiple rounds of iterative propagation, a feature state sequence containing multiple time nodes is generated, with each time node corresponding to a feature vector. These feature vectors are input into a pre-trained efficacy evaluation network, which outputs the predicted efficacy values ​​for each time point, forming a time-series efficacy evolution curve.

[0045] The parameter sensitivity decomposition algorithm processes the time-series efficacy evolution curve, using wavelet transform to decompose the curve into different frequency components. High-frequency components correspond to the rapid response of the efficacy, reflecting the direct effects of the drug in the initial stage of action, including immediate bladder mucosal response and peak drug concentration effect. Low-frequency components correspond to the delayed response of the efficacy, reflecting the long-term processes of immune regulation and tissue remodeling. Energy integrals are calculated for each decomposed frequency component; the energy integral value of the rapid response component is denoted as E_fast, and the energy integral value of the delayed response component is denoted as E_slow.

[0046] The contribution weights were calculated using a normalization method. The contribution weight of the fast response component was E_fast divided by the total energy, and the contribution weight of the delayed response component was E_slow divided by the total energy. The weight reconstruction process involved weighted summation of the efficacy values ​​of each response component at the final time point to obtain a comprehensive efficacy prediction. This value comprehensively reflects the overall efficacy level of the parameter adjustment combination in the virtual intervention scenario.

[0047] The difference between the estimated overall efficacy and the baseline predicted efficacy is used as the efficacy improvement index, with positive values ​​indicating improved efficacy and negative values ​​indicating decreased efficacy. The contribution weight vector is concatenated with the efficacy improvement index to form a complete descriptive vector encompassing response mechanism analysis and the degree of efficacy improvement, representing the efficacy change result for this virtual intervention scenario. This result not only provides the numerical change in efficacy but also reveals the specific ways in which parameter adjustments affect the final efficacy through both rapid and delayed response paths, providing a multi-dimensional decision-making basis for the selection of subsequent optimization targets.

[0048] In one optional implementation, the counterfactual reasoning engine calculates the contribution of each parameter combination to the improvement of therapeutic effect, selects the parameter combination with the highest contribution as the optimization target, and generates an individualized perfusion plan based on the optimization target, including: The efficacy change results of all virtual intervention scenarios are input into the counterfactual reasoning engine for contribution quantification analysis. Based on the counterfactual reasoning engine, the efficacy improvement index and corresponding parameter combination in each efficacy change result are extracted. The improvement of the efficacy improvement index relative to the predicted efficacy value is calculated. A causal attribution decomposition mechanism is introduced to decompose the improvement into the marginal contribution component and synergistic contribution component of the infusion drug ratio, infusion time interval and infusion dose. The contribution of the corresponding parameter combination to the efficacy improvement is obtained by weighted aggregation of the marginal contribution component and the synergistic contribution component. The contribution of all parameter combinations is sorted in descending order, and the parameter combination with the highest contribution is identified as the optimization target. The optimization target is then input into the scheme conversion module. Based on the scheme conversion module, the specific drug types and their proportions are determined according to the target setting value of the infusion drug ratio, the treatment cycle schedule is planned according to the target setting value of the infusion time interval, and the single-dose dosage specification is formulated according to the target setting value of the infusion dose, thus integrating them into an individualized infusion scheme that includes drug configuration scheme, treatment schedule and dosage specification.

[0049] like Figure 2 As shown, the method includes: After simulating the efficacy of multiple virtual intervention scenarios, the efficacy changes from all scenarios are batch-fed into the counterfactual reasoning engine. The engine first performs structured analysis on each efficacy change result, extracting the efficacy improvement index and its corresponding parameter combination. The efficacy improvement index is typically characterized by quantitative indicators such as the percentage decrease in bladder tumor recurrence rate, improvement in cystoscopy scores, or reduction in inflammatory marker concentrations. The difference between this index and the baseline predicted efficacy value is calculated to obtain the absolute improvement, which is then normalized by dividing by the predicted efficacy value to form the relative improvement.

[0050] The core of the causal attribution decomposition mechanism lies in breaking down the overall improvement into interpretable sub-components. In practice, a structural causal model is used to establish a causal relationship diagram between three variables—the infusion drug ratio, the infusion time interval, and the infusion dose—and the improvement in efficacy. Using the do-calculus operator, two of these variables are fixed one by one, while only the third variable is changed. The change in efficacy improvement is observed; this change represents the marginal contribution component of that variable. For example, if the time interval and dose are fixed, and only the drug ratio is adjusted, and the efficacy improves by 0.15 units, then the marginal contribution component of the drug ratio is 0.15. The synergistic contribution component is obtained by subtracting the sum of the three marginal contribution components from the actual total improvement; this value reflects the gain effect produced by the interaction between variables.

[0051] When weighting and aggregating the marginal contribution component and the synergistic contribution component, weighting coefficients are set based on clinical practice experience. Typically, the marginal contribution weight is set to 0.7, and the synergistic contribution weight is set to 0.3. The weighted sum of these two values ​​yields the overall contribution level of the parameter combination. After calculating the contribution level of all parameter combinations, a mapping table between parameter combinations and contribution levels is established, and the combinations are sorted in descending order of contribution level. The parameter combination that ranks first in the sorting is selected as the optimization target, as this combination possesses the optimal potential for improving therapeutic efficacy.

[0052] After optimizing the target input scheme conversion module, the module first parses the target settings for the infusion drug ratio. Assuming the target setting is a gemcitabine to mitomycin ratio of 3:1, the module determines the gemcitabine concentration to be 40 mg / mL and the mitomycin concentration to be 13.3 mg / mL based on the standard drug concentration database, and generates drug preparation operation guidelines. Regarding the target setting for the infusion interval, if it is set to be once a week for 6 weeks, the module automatically generates a treatment cycle schedule, marking the specific date and time window for each infusion, while reserving a 72-hour adjustment buffer period to handle unexpected patient situations. If the target setting for the infusion dose is 50 mL per dose, the module specifies the dosage and associates it with the infusion device model, recommending the use of a 50 mL bladder infusion catheter with a precision infusion pump to control the flow rate.

[0053] The protocol conversion module integrates and encapsulates drug preparation plans, treatment schedules, and dosage specifications into a structured, individualized infusion protocol document. This document includes drug name, concentration ratio, single dose, infusion frequency, total treatment cycle length, specific execution date for each infusion, a list of required medical devices, and operational precautions. The protocol also highlights key quality control points, such as the stability testing timeframe after drug preparation, bladder pressure monitoring thresholds during infusion, and laboratory indicators requiring follow-up during treatment. This provides a complete execution basis for clinical implementation, ensuring the operability and safety of the individualized protocol in practical application.

[0054] A second aspect of the present invention provides a bladder instillation therapy efficacy prediction system based on multimodal data, comprising: The data acquisition unit is used to acquire the patient's medical imaging data, clinical test data, and pathological text data, and to extract features from the medical imaging data, the clinical test data, and the pathological text data to obtain image feature vectors, test feature vectors, and text feature vectors. The feature extraction unit is used to construct a multimodal adversarial unentanglement network. By introducing a modality discriminator, the image feature vector, the test feature vector, and the text feature vector are adversarially trained, so that the feature representation can eliminate the discriminability of modality source while retaining efficacy-related information, and generate modality-invariant pure efficacy features. The feature resolution unit is used to input the pure efficacy features into the prediction model for training, establish a mapping relationship between the pure efficacy features of historical patients and their corresponding real efficacy labels to obtain the trained prediction model, and output the predicted efficacy value of the current patient based on the trained prediction model. The model training unit is used to build a counterfactual reasoning engine. It takes the pure efficacy features and the predicted efficacy values ​​as benchmark inputs to generate multiple virtual intervention scenarios. Based on each virtual intervention scenario, it simulates the efficacy changes under different combinations of infusion drug ratios, infusion time intervals, and infusion doses. The counterfactual reasoning unit is used to compare the changes in efficacy in each virtual intervention scenario with the predicted efficacy value, calculate the contribution of each parameter combination to the improvement of efficacy through the counterfactual reasoning engine, select the parameter combination with the highest contribution as the optimization target, and generate an individualized perfusion plan based on the optimization target.

[0055] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0056] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0057] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0058] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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 or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the therapeutic effect of bladder instillation based on multimodal data, characterized in that, include: Acquire the patient's medical imaging data, clinical laboratory data, and pathological text data, and extract features from the medical imaging data, clinical laboratory data, and pathological text data respectively to obtain image feature vectors, laboratory feature vectors, and text feature vectors. A multimodal adversarial unentanglement network is constructed. By introducing a modality discriminator, the image feature vector, the test feature vector, and the text feature vector are trained adversarially. This enables the feature representation to retain efficacy-related information while eliminating the discriminability of modality sources, thereby generating modality-invariant pure efficacy features. The pure efficacy features are input into the prediction model for training. A mapping relationship is established between the pure efficacy features of historical patients and their corresponding real efficacy labels to obtain the trained prediction model. The predicted efficacy value of the current patient is output based on the trained prediction model. A counterfactual reasoning engine is constructed, which takes the pure efficacy characteristics and the predicted efficacy value as the baseline input to generate multiple virtual intervention scenarios. Based on each virtual intervention scenario, the efficacy change results are simulated under different combinations of infusion drug ratios, infusion time intervals and infusion doses. The efficacy changes in each virtual intervention scenario are compared with the predicted efficacy values. The counterfactual reasoning engine calculates the contribution of each parameter combination to the efficacy improvement, selects the parameter combination with the highest contribution as the optimization target, and generates an individualized perfusion plan based on the optimization target.

2. The method according to claim 1, characterized in that, By introducing a modality discriminator to perform adversarial training on the image feature vector, the test feature vector, and the text feature vector, the feature representation can eliminate the discriminability of modality origin while retaining efficacy-related information, generating modality-invariant pure efficacy features, including: The image feature vector, the test feature vector, and the text feature vector are respectively input into the feature generator for encoding and conversion to obtain candidate features of the same dimension. The candidate features are then input into the modality discriminator for modality source identification. The modality confusion loss is calculated based on the modality classification results output by the modality discriminator. In the adversarial training process, a efficacy retention constraint is introduced. The correlation strength between the candidate features and the historical patient efficacy labels is calculated. The correlation strength is used as a quantitative indicator of the degree of efficacy information retention. A joint optimization objective function is constructed to maximize the modality confusion loss while maximizing the correlation strength. The parameters of the feature generator are updated by backpropagation, so that the candidate features maintain a strong correlation with the therapeutic effect while gradually losing modal discriminability. When the modal discriminator’s modal recognition accuracy for the candidate features drops below a preset accuracy threshold and the correlation strength stabilizes within a preset range, the candidate features at this time are determined as pure therapeutic effect features with unchanged modality.

3. The method according to claim 2, characterized in that, Using the correlation strength as a quantitative indicator of the degree of retention of therapeutic information, a joint optimization objective function is constructed to maximize both the modality confusion loss and the correlation strength, including: The correlation strength is used as a quantitative indicator of the degree of retention of efficacy information. A joint optimization objective function is constructed based on the quantitative indicator. The joint optimization objective function includes a modality confusion loss term and an efficacy retention loss term. Positive optimization weights are configured for the modality confusion loss term to drive the feature generator to update parameters in the direction of maximizing the modality confusion loss. At the same time, positive optimization weights are configured for the efficacy retention loss term to drive the feature generator to update parameters in the direction of maximizing the correlation strength. By synchronously adjusting the parameters of the feature generator through a gradient optimization algorithm, the candidate features continuously reduce modality discriminability and continuously enhance efficacy relevance during the iteration process, thereby maximizing both the modality confusion loss and the relevance strength.

4. The method according to claim 1, characterized in that, A mapping relationship is established between the pure efficacy characteristics of historical patients and their corresponding real efficacy labels to obtain a trained prediction model. Based on the trained prediction model, the predicted efficacy value of the current patient is output, including: A training sample set is constructed by extracting pure efficacy features and real efficacy labels from a historical patient database. Cluster analysis is performed on the training sample set to identify patient subgroups with different efficacy response patterns. A dedicated sub-mapping model is constructed for each patient subgroup. A hierarchical mapping architecture is formed by integrating all sub-mapping models. A dynamic sample weighting mechanism is introduced during training to calculate the prediction residual of each training sample. The weight contribution of the corresponding sample in subsequent iterations is adaptively adjusted according to the prediction residual. Samples with prediction residuals are given weights to enhance the learning ability of difficult examples. Through iterative optimization, the hierarchical mapping architecture achieves stable prediction performance under different therapeutic response modes, and the trained prediction model is obtained. The pure efficacy characteristics of the current patient are input into the trained prediction model. The efficacy response subgroup to which the current patient belongs is identified by the patient subgroup attribution determination module. The sub-mapping model corresponding to the efficacy response subgroup is called and the predicted efficacy value of the current patient is calculated and output.

5. The method according to claim 1, characterized in that, A counterfactual reasoning engine is constructed, using the pure efficacy features and the predicted efficacy values ​​as baseline inputs, to generate multiple virtual intervention scenarios, including: A counterfactual reasoning engine is constructed, which includes a structural causal model and a condition generator. The structural causal model learns the causal mechanism of infusion drug ratio, infusion time interval and infusion dose on efficacy by analyzing historical treatment data, and establishes a causal inference function from interventionable variables to changes in efficacy. Pure therapeutic efficacy features are input into the condition generator as the patient's inherent state benchmark. An intervention parameter adjustment space is constructed based on the causal inference function. Parameter adjustment combinations are sampled in the intervention parameter adjustment space. The perturbation effect of the parameter adjustment combinations on the pure therapeutic efficacy features is calculated through the causal inference function. The perturbation effect is superimposed on the pure therapeutic efficacy features to generate a virtual patient characteristic state. The predicted therapeutic effect value is associated with each virtual patient's characteristic state, and each virtual patient's characteristic state is labeled with its corresponding parameter adjustment combination and the predicted therapeutic effect value to form a virtual intervention scenario.

6. The method according to claim 1, characterized in that, The results of efficacy changes simulated under different combinations of infusion drug ratios, infusion time intervals, and infusion doses for each of the aforementioned virtual intervention scenarios include: The virtual patient's characteristic state and corresponding parameter adjustment combination are extracted from each virtual intervention scenario to construct a therapeutic effect evolution simulation module. According to the therapeutic effect evolution simulation module, the changes in the infusion drug ratio, infusion time interval and infusion dose in the parameter adjustment combination are transformed into a perturbation vector in the feature space through the causal intervention propagation mechanism. The perturbation vector is propagated along the causal path of the virtual patient's characteristic state to simulate the physiological response cascade reaction caused by the adjustment of treatment parameters and generate a time-series therapeutic effect evolution curve. A parameter sensitivity decomposition algorithm is introduced to decompose the time-series efficacy evolution curve into multiple scales, extract the rapid response component and the delayed response component, calculate the contribution weight of each response component to the final efficacy value, and obtain the comprehensive efficacy prediction value of the corresponding virtual intervention scenario through weight reconstruction. The efficacy improvement index is calculated by comparing the comprehensive efficacy estimate with the predicted efficacy value. The contribution weight and the efficacy improvement index are combined to form the efficacy change result of the corresponding virtual intervention scenario.

7. The method according to claim 1, characterized in that, The counterfactual reasoning engine calculates the contribution of each parameter combination to the improvement of therapeutic effect, selects the parameter combination with the highest contribution as the optimization target, and generates an individualized perfusion plan based on the optimization target, including: The efficacy change results of all virtual intervention scenarios are input into the counterfactual reasoning engine for contribution quantification analysis. Based on the counterfactual reasoning engine, the efficacy improvement index and corresponding parameter combination in each efficacy change result are extracted. The improvement of the efficacy improvement index relative to the predicted efficacy value is calculated. A causal attribution decomposition mechanism is introduced to decompose the improvement into the marginal contribution component and synergistic contribution component of the infusion drug ratio, infusion time interval and infusion dose. The contribution of the corresponding parameter combination to the efficacy improvement is obtained by weighted aggregation of the marginal contribution component and the synergistic contribution component. The contribution of all parameter combinations is sorted in descending order, and the parameter combination with the highest contribution is identified as the optimization target. The optimization target is then input into the scheme conversion module. Based on the scheme conversion module, the specific drug types and their proportions are determined according to the target setting value of the infusion drug ratio, the treatment cycle schedule is planned according to the target setting value of the infusion time interval, and the single-dose dosage specification is formulated according to the target setting value of the infusion dose, thus integrating them into an individualized infusion scheme that includes drug configuration scheme, treatment schedule and dosage specification.

8. A bladder instillation therapy efficacy prediction system based on multimodal data, used to implement the method of any one of claims 1-7, characterized in that, include: The data acquisition unit is used to acquire the patient's medical imaging data, clinical test data, and pathological text data, and to extract features from the medical imaging data, the clinical test data, and the pathological text data to obtain image feature vectors, test feature vectors, and text feature vectors. The feature extraction unit is used to construct a multimodal adversarial unentanglement network. By introducing a modality discriminator, the image feature vector, the test feature vector, and the text feature vector are adversarially trained, so that the feature representation can eliminate the discriminability of modality source while retaining efficacy-related information, and generate modality-invariant pure efficacy features. The feature resolution unit is used to input the pure efficacy features into the prediction model for training, establish a mapping relationship between the pure efficacy features of historical patients and their corresponding real efficacy labels to obtain the trained prediction model, and output the predicted efficacy value of the current patient based on the trained prediction model. The model training unit is used to build a counterfactual reasoning engine. It takes the pure efficacy features and the predicted efficacy values ​​as benchmark inputs to generate multiple virtual intervention scenarios. Based on each virtual intervention scenario, it simulates the efficacy changes under different combinations of infusion drug ratios, infusion time intervals, and infusion doses. The counterfactual reasoning unit is used to compare the changes in efficacy in each virtual intervention scenario with the predicted efficacy value, calculate the contribution of each parameter combination to the improvement of efficacy through the counterfactual reasoning engine, select the parameter combination with the highest contribution as the optimization target, and generate an individualized perfusion plan based on the optimization target.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images