Optimization method for acupuncture treatment of female stress urinary incontinence based on multi-modal data fusion
By generating a unified representation through multimodal data fusion technology, and combining dynamic temporal kernel and cross-attention mechanism, the problems of individualization and dynamic optimization in acupuncture treatment are solved, realizing individualized and precise treatment and efficacy prediction for female stress urinary incontinence, and improving the pertinence and predictability of treatment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING SHIJITAN HOSPITAL CAPITAL MEDICAL UNIVERSITY
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-24
AI Technical Summary
Existing acupuncture treatments for female stress urinary incontinence lack individualized considerations, lack dynamic adjustment mechanisms during treatment, make it difficult to predict efficacy, and hinder real-time prediction and program optimization.
Multimodal data fusion technology is employed to extract and fuse patient baseline features, pelvic floor images, and electromyographic signals through a representation learning subnetwork to generate a unified representation. This representation is then combined with a dynamic temporal kernel for efficacy prediction and a cross-attention mechanism to optimize acupuncture treatment parameters.
It enables personalized and precise treatment, improves the ability to predict efficacy and dynamically optimize the treatment process, supports personalized treatment recommendations, and promotes the intelligent development of integrated traditional Chinese and Western medicine.
Smart Images
Figure CN122455243A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent assisted medical technology, specifically relating to an optimized method for acupuncture treatment of female stress urinary incontinence based on multimodal data fusion. Background Technology
[0002] Stress urinary incontinence (SUI) is a common urinary tract disease in adult women, mainly characterized by involuntary leakage of urine when abdominal pressure increases, such as during coughing, sneezing, or exercise. It severely impacts patients' quality of life and social participation. Acupuncture, as an important non-surgical treatment for SUI in Traditional Chinese Medicine, has been widely used clinically due to its simplicity, minimal side effects, and high patient acceptance.
[0003] However, current clinical practice of acupuncture treatment for SUI still has many limitations. First, treatment plans are mostly based on standardized experience, lacking individualized consideration. Most patients receive the same acupoint selection, stimulation intensity, and treatment course, ignoring the heterogeneity of patients in terms of etiology, pathophysiological characteristics, physical condition, and treatment response, leading to poor efficacy in some patients. Second, the treatment process lacks a dynamic adjustment mechanism. Clinicians usually conduct efficacy assessments at fixed time points and decide on subsequent treatments based on experience, making it difficult to identify efficacy trends in the early stages of treatment and to optimize treatment plans for ineffective or poorly responding patients in a timely manner, resulting in wasted medical resources and decreased patient confidence. In addition, efficacy prediction is difficult. Existing methods struggle to answer key questions before or in the early stages of treatment, such as "Is this therapy suitable for this patient?" and "What is the optimal treatment cycle?", lacking forward-looking decision support.
[0004] In recent years, with the development of artificial intelligence and multimodal data fusion technology, some studies have attempted to apply data mining methods to the analysis of acupuncture efficacy. However, most of these studies remain limited to static, single-modal, post-hoc retrospective analysis, lacking the ability to deeply integrate multi-source heterogeneous data and dynamically model time series data, thus failing to achieve real-time prediction and closed-loop optimization during the treatment process. Therefore, there is an urgent need for an intelligent system that can integrate multi-dimensional patient information, achieve dynamic prediction of efficacy, and adaptive recommendation of treatment plans to promote the development of acupuncture treatment SUI towards precision and personalization. Summary of the Invention
[0005] To address the aforementioned technical issues, this invention provides an optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion. By constructing an intelligent diagnosis and treatment process of "multimodal feature fusion - efficacy trend prediction - personalized treatment plan recommendation," the method achieves dynamic optimization of the entire process of acupuncture treatment for stress urinary incontinence (SUI), thereby improving the targetedness, predictability, and clinical benefit rate of the treatment.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] An optimized approach to acupuncture treatment for female stress urinary incontinence based on multimodal data fusion includes:
[0008] Step S1: Extract and fuse multimodal data using a representation learning sub-network to generate a unified representation. The representation learning sub-network includes a modal difference feature extraction module, a cross-modal attention alignment module, and a fusion feature enhancement module connected in series.
[0009] Step S2: The unified representation is concatenated with the current reference acupuncture treatment parameter vector and then input into the dynamic temporal kernel generation network. Temporal features are aggregated based on the generated dynamic temporal kernel, and then the efficacy prediction results at multiple time points are output through multi-head attention and feedforward neural network.
[0010] Step S3: Based on the efficacy prediction results and the historical treatment protocol database, the optimized acupuncture treatment parameters are obtained through the cross-attention mechanism and the effect evaluation feedforward network.
[0011] Furthermore, the modal differential feature extraction module in step S1 includes multiple parallel feature extraction network branches, which are used to extract differential features from patient baseline features, pelvic floor image sequence data, electromyographic signal data and efficacy index data, respectively.
[0012] Furthermore, the patient baseline features are extracted using a multilayer perceptron network to represent static baseline features. The multilayer perceptron network includes a cascaded linear projection layer, a batch normalization layer, and a nonlinear activation layer. The pelvic floor image sequence data are extracted using a three-dimensional Swing Transformer encoder to represent multi-scale features in the spatial and temporal dimensions of the image sequence. The electromyographic signal data and therapeutic index data are obtained using a Transformer encoder structure, and their dynamic trends are captured through multi-head attention.
[0013] Furthermore, the cross-modal attention alignment module adopts a bidirectional cross-modal attention mechanism, using the patient's baseline features as anchors, to calculate the cross-attention between the pelvic floor image sequence data, electromyographic signal data, and efficacy index data and the patient's baseline features, thereby achieving semantic alignment between the pelvic floor image sequence data, electromyographic signal data, and efficacy index data and the patient's baseline features.
[0014] Furthermore, the fusion feature enhancement module concatenates the aligned modal features along the channel dimension to obtain aligned fusion features; through contrastive learning, it narrows the distance between different modal features of the same patient and widens the distance between features of different patients, optimizes the representation space distribution, and obtains the final unified representation; the optimization loss function of the contrastive learning uses cosine similarity calculation, and a learnable linear mapping and nonlinear activation function are used to amplify the differences between different samples.
[0015] Furthermore, step S2 specifically includes: concatenating the unified representation with the current reference acupuncture treatment parameter vector and inputting it into a dynamic temporal kernel generation network to generate a dynamic temporal kernel with normalized time as a parameter, wherein the dynamic temporal kernel is determined by a learnable parameter; using the generated dynamic temporal kernel to perform weighted aggregation of historical temporal features to obtain aggregated temporal features; and finally using a multi-head attention mechanism and a feedforward neural network to obtain the treatment state prediction results for multiple future time points.
[0016] Furthermore, step S3 specifically includes: introducing a state-scheme cross-attention mechanism, using the efficacy prediction result as the query and the historical treatment scheme parameter feature library as the key and value, calculating a learnable mapping matrix, and generating the scheme association feature that best matches the current patient state; training an efficacy evaluation feedforward network based on the treatment efficacy scoring system, and calculating optimized acupuncture treatment parameters with the goal of minimizing the treatment efficacy evaluation loss function.
[0017] Furthermore, the acupuncture treatment parameters include the combination of acupuncture points, the intensity of the manipulation, and the treatment cycle.
[0018] In a second aspect, the present invention provides an electronic device comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the aforementioned optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion.
[0019] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned optimized method for acupuncture treatment of female stress urinary incontinence based on multimodal data fusion.
[0020] The beneficial effects of this invention are as follows:
[0021] Achieving personalized and precise treatment: By integrating multimodal data such as patient baseline characteristics, pelvic floor imaging, and electromyography signals, a unified representation is constructed to overcome the homogenization problem of traditional treatments and provide each patient with highly targeted acupuncture points, manipulation intensity, and treatment cycle suggestions.
[0022] Enhancing efficacy prediction capabilities: By employing a dynamic temporal kernel fusion prediction method, the patient's status changes at multiple future time points can be accurately predicted based on current treatment parameters, helping doctors identify efficacy trends in the early stages of treatment and promptly adjust ineffective or poorly responded treatment plans.
[0023] Support for dynamic optimization of the treatment process: Introduce a "state-plan" cross-attention mechanism, combine real-time efficacy prediction with historical treatment plan characteristics, realize closed-loop dynamic recommendation of treatment plans, and promote the transformation of acupuncture treatment from static experience-based decision-making to dynamic data-driven optimization.
[0024] Enhance the intelligence of clinical decision-making: Construct an end-to-end intelligent diagnosis and treatment link from multimodal data fusion and efficacy trend prediction to personalized treatment plan recommendation, providing doctors with interpretable and actionable decision support, and improving treatment efficiency and patient compliance.
[0025] Promoting the intelligent development of integrated traditional Chinese and Western medicine: By deeply integrating modern artificial intelligence technology with traditional acupuncture therapy, a systematic approach is provided for SUI's intelligent acupuncture treatment, which has important clinical promotion and application value. Attached Figure Description
[0026] Figure 1 This is a flowchart of the optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion, as described in this invention.
[0027] Figure 2 Schematic diagram of the learning subnetwork;
[0028] Figure 3 A flowchart of the steps for predicting treatment efficacy;
[0029] Figure 4 A flowchart for optimizing acupuncture treatment parameters and steps. Detailed Implementation
[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0031] like Figure 1As shown, the proposed method for optimizing acupuncture treatment of female stress urinary incontinence based on multimodal data fusion first integrates heterogeneous data such as patient baseline characteristics, treatment parameters, pelvic floor imaging, and electromyography signals, generating a unified representation through multimodal fusion technology to bridge the semantic gap between multi-source data. Second, relying on temporal modeling technology, it accurately predicts the trend of patient status changes at multiple time points after treatment based on the unified representation. Finally, combining the trends of status changes at multiple time points, it outputs personalized acupuncture points, manipulation intensity, and treatment cycle suggestions through protocol adaptation technology, solving the problem of homogenized treatment protocols. The method proposed in this invention connects all aspects of data acquisition, efficacy prediction, and protocol optimization, constructing a clinically applicable intelligent diagnosis and treatment link, effectively improving the targeting and clinical benefit rate of individualized treatment. The method specifically includes:
[0032] Step S1: Extract and fuse multimodal data using a representation learning sub-network to generate a unified representation. The representation learning sub-network includes a modal difference feature extraction module, a cross-modal attention alignment module, and a fusion feature enhancement module connected in series.
[0033] The treatment of female stress urinary incontinence involves multi-source heterogeneous data, including baseline features (age, disease duration, weight, etc.) and dynamic sequence data of treatment status (pelvic floor imaging sequences, electromyographic signal time-series data, and efficacy indicator change sequences). These data exhibit significant differences in modal data distribution, dimensionality, and information density, making it difficult for traditional feature splicing or simple weighted fusion methods to capture deep intermodal relationships. This invention employs a representation learning sub-network to learn representations of the multimodal data in acupuncture treatment, obtaining unified representational features of the treatment status. This provides foundational input data for subsequent efficacy assessment and treatment plan recommendation.
[0034] like Figure 2 As shown, the representation learning sub-network architecture proposed in this invention includes the following three sequentially connected modules: a module for extracting features of different modes; a cross-modal attention alignment module; and a fusion feature enhancement module.
[0035] First, based on the modality difference feature extraction modules, differential features are extracted from the multimodal data. Each modality difference feature extraction module includes multiple parallel feature extraction network branches, each used to extract the aforementioned multimodal data; specifically, it includes:
[0036] For patient baseline features, a multilayer perceptron network is used to extract static feature representations of the patient baseline. This multilayer perceptron network consists of a linear projection layer, a batch normalization layer (BatchNorm), and a nonlinear activation layer (Gelu), which extracts the patient baseline feature data. Input, obtain patient baseline static feature representation The calculation formula is:
[0037] ,
[0038] In the formula , These are learnable parameters.
[0039] For the patient's pelvic floor imaging sequence data A 3D Swin Transformer encoder is used to input image sequences into a 3D convolutional network (Conv3D). Then, by alternately using windowed multi-head self-attention (WMSA) and shifted windowed multi-head self-attention (SWMSA) modules, the multi-scale features of the spatial and temporal dimensions of the image sequences are captured. The calculation formula is:
[0040] ,
[0041] For the patient's electromyography data and efficacy indicator data The Transformer encoder structure (TransEncoder) is adopted, and multi-head attention is used to capture the dynamic change trends of the two modalities, thereby obtaining the differentiated representation features of the two modalities. and The calculation method is as follows:
[0042] ,
[0043] ,
[0044] PosEmbed represents the position embedding of the time sequence.
[0045] Then, the differential features of the above four modalities are input into the cross-modal attention alignment module, which employs a bidirectional cross-modal attention mechanism (CrossAttention) based on the patient's baseline feature data. Using anchor points, calculate other modal features and baseline feature data. Cross-attention enables other modalities to interact with baseline feature data. Semantic alignment, a certain modality (can be) , and (Any of the) aligned features The calculation formula is:
[0046] ,
[0047] The fusion feature enhancement module is then used to concatenate the aligned modal features along the channel dimension to obtain the aligned fusion features. Contrastive learning optimizes the representation space distribution by narrowing the gap between different modalities of the same patient and widening the gap between features of different patients, making patient features more discriminative and ultimately achieving a unified representation. The optimized loss function is:
[0048] ,
[0049] Where sim is the cosine similarity, Linear is a learnable linear mapping, Gelu is a non-linear activation function, and i and j represent different patient samples, which amplify the differences between different samples in contrastive learning.
[0050] Step S2: The unified representation is concatenated with the current reference acupuncture treatment parameter vector and then input into the dynamic temporal kernel generation network. Temporal features are aggregated based on the generated dynamic temporal kernel, and then the efficacy prediction results at multiple time points are output through multi-head attention and feedforward neural network.
[0051] like Figure 3 As shown, after obtaining a unified representation of multimodal data, the patient is first represented as a unified entity. With acupuncture treatment parameter vector The concatenated vector is obtained after concatenation. The treatment parameters include: acupuncture point combinations, manipulation intensity, and treatment cycle. Without losing generality, the parameters mentioned here... This refers to a reference parameter vector, which can be provided by experts or based on clinical experience. Then, the concatenated vector... The dynamic temporal kernel is input to generate an MLP network. The dynamic temporal kernel is calculated using the following formula:
[0052] ,
[0053] Where t is the normalized time parameter. It is a dynamic timing core. , These are learnable parameters. Then, the generated dynamic temporal kernel is used to obtain the aggregated temporal features, calculated using the following formula:
[0054] ,
[0055] in This represents aggregated temporal features;
[0056] Finally, a multi-head attention mechanism (MultHeadAttn) and a feedforward neural network (FFN) are used to obtain the treatment effect, i.e., the prediction of future treatment status. The formula for calculating the efficacy prediction at multiple time points is as follows:
[0057] ,
[0058] Here, PosEmbed represents the temporal sequence position embedding. The objective function used for network training is the residual between the predicted and true values.
[0059] Step S3: Based on the efficacy prediction results and the historical treatment plan database, the optimized acupuncture treatment parameters are obtained through the cross-attention mechanism and the effect evaluation feedforward network, thereby providing an acupuncture treatment plan.
[0060] like Figure 4 As shown, firstly, a "state-scheme" cross-attention mechanism is introduced to learn the efficacy contribution weights of various acupuncture schemes under different patient conditions:
[0061] ,
[0062] in, To generate the associated features of the treatment plan that best matches the current patient status, This is a database of historical treatment protocol parameters. As a dimension of treatment plan characteristics, , , It is a learnable mapping matrix.
[0063] A treatment efficacy scoring system is used to evaluate treatment outcomes in different patient states by professional physicians. A feedforward network (FNN) is trained to assess treatment efficacy, and the loss function for treatment efficacy evaluation is:
[0064] ,
[0065] For the test patients, the optimized acupuncture treatment parameters are calculated by minimizing this loss function. The output of this method serves as supplementary decision-making information; the final treatment plan is determined by the physician.
[0066] In a second aspect, the present invention provides an electronic device comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the aforementioned optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion.
[0067] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned optimized method for acupuncture treatment of female stress urinary incontinence based on multimodal data fusion.
[0068] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion, characterized in that, include: Step S1: Extract and fuse multimodal data using a representation learning sub-network to generate a unified representation. The representation learning sub-network includes a modal difference feature extraction module, a cross-modal attention alignment module, and a fusion feature enhancement module connected in series. Step S2: The unified representation is concatenated with the current reference acupuncture treatment parameter vector and then input into the dynamic temporal kernel generation network. Temporal features are aggregated based on the generated dynamic temporal kernel, and then the efficacy prediction results at multiple time points are output through multi-head attention and feedforward neural network. Step S3: Based on the efficacy prediction results and the historical treatment protocol database, the optimized acupuncture treatment parameters are obtained through the cross-attention mechanism and the effect evaluation feedforward network.
2. The optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion according to claim 1, characterized in that, The modal differential feature extraction module in step S1 includes multiple parallel feature extraction network branches, which are used to extract differential features from patient baseline features, pelvic floor image sequence data, electromyographic signal data and efficacy index data, respectively.
3. The optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion according to claim 2, characterized in that, The patient baseline features are extracted using a multilayer perceptron network to represent static baseline features. The multilayer perceptron network includes a cascaded linear projection layer, a batch normalization layer, and a nonlinear activation layer. The pelvic floor image sequence data are extracted using a three-dimensional Swing Transformer encoder to represent multi-scale features in the spatial and temporal dimensions of the image sequence. The electromyographic signal data and therapeutic index data are obtained using a Transformer encoder structure, and their dynamic trends are captured through multi-head attention.
4. The optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion according to claim 2, characterized in that, The cross-modal attention alignment module adopts a bidirectional cross-modal attention mechanism, using the patient's baseline features as anchors, to calculate the cross-attention between pelvic floor image sequence data, electromyographic signal data, and efficacy index data and the patient's baseline features, thereby achieving semantic alignment between the pelvic floor image sequence data, electromyographic signal data, and efficacy index data and the patient's baseline features.
5. The optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion according to claim 4, characterized in that, The fusion feature enhancement module concatenates the aligned modal features along the channel dimension to obtain aligned fusion features; through contrastive learning, it narrows the distance between different modal features of the same patient and widens the distance between features of different patients, optimizes the representation space distribution, and obtains the final unified representation; the optimization loss function of the contrastive learning uses cosine similarity calculation, and a learnable linear mapping and nonlinear activation function are used to amplify the differences between different samples.
6. The optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion according to claim 1, characterized in that, Step S2 specifically includes: concatenating the unified representation with the current reference acupuncture treatment parameter vector and inputting it into a dynamic temporal kernel generation network to generate a dynamic temporal kernel with normalized time as a parameter, wherein the dynamic temporal kernel is determined by a learnable parameter; using the generated dynamic temporal kernel to perform weighted aggregation of historical temporal features to obtain aggregated temporal features; and finally using a multi-head attention mechanism and a feedforward neural network to obtain the treatment state prediction results for multiple future time points.
7. The optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion according to claim 1, characterized in that, Step S3 specifically includes: introducing a state-plan cross-attention mechanism, using the efficacy prediction result as the query and the historical treatment plan parameter feature library as the key and value, calculating a learnable mapping matrix, and generating the plan association feature that best matches the current patient state; training an efficacy evaluation feedforward network based on the treatment efficacy scoring system, and calculating optimized acupuncture treatment parameters with the goal of minimizing the treatment efficacy evaluation loss function.
8. The optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion according to claim 7, characterized in that, The acupuncture treatment parameters include the combination of acupuncture points, the intensity of the manipulation, and the treatment cycle.
9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion as described in any one of claims 1-8.
10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the optimized acupuncture treatment method for female stress urinary incontinence based on multimodal data fusion as described in any one of claims 1-8.