Multi-constraint adaptive multi-modal fusion inflammatory bowel disease risk prediction method
By employing three-level bioinformatics filtering, modal orthogonal constraints, and hierarchical attention dynamic fusion, the problems of insufficient single-modal information, high-dimensional redundancy, and poor generalization in IBD diagnosis are solved, enabling early non-invasive and accurate risk prediction of IBD.
Patent Information
- Application Number
- CN202610276914.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-09
- Publication Date
- 2026-04-10
AI Technical Summary
Current IBD diagnostic techniques rely on endoscopy and pathological assessment, which are highly invasive and easily influenced by physician experience. Single-modal data has limited ability to capture disease features, and high-dimensional redundancy and noise characteristics lead to model overfitting. Multimodal fusion cannot adapt to pathological differences and has insufficient generalization ability.
Employing three-level bioinformatics filtering, modal orthogonal constraints, hierarchical attention dynamic fusion, and multi-objective optimization, this approach enhances early diagnosis of IBD through precise feature selection, intelligent modal fusion, and improved discrimination capabilities.
It achieves non-invasive and accurate risk prediction for early diagnosis of IBD, improves the stability and discrimination accuracy of the model, and solves the problems of insufficient single-modal information, high-dimensional redundancy and poor generalization.
Smart Images
Figure CN121839141A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of medical data processing and machine learning, specifically relating to a multi-constraint adaptive multimodal fusion method for predicting the risk of inflammatory bowel disease. Background Technology
[0002] Inflammatory bowel disease (IBD) is a group of idiopathic inflammatory bowel diseases, mainly including Crohn's disease (CD) and ulcerative colitis (UC). Clinical manifestations include diarrhea, abdominal pain, and bloody stools. Early diagnosis is crucial for treatment and prognosis. Traditional IBD diagnosis relies on endoscopic and pathological evaluation, which is invasive and easily influenced by physician experience, thus having significant limitations.
[0003] With the development of omics technology, multi-source data such as microbiome, enzyme function, and metabolome have provided new avenues for IBD diagnosis. However, existing technologies still face many problems: (1) Single-modal data can only capture some features of the disease and cannot reflect the multi-dimensional biological characteristics of IBD, resulting in limited predictive performance; (2) Modal data such as enzyme function have high-dimensional redundancy and high noise characteristics, and direct modeling is prone to "curse of dimensionality" and overfitting; (3) Existing multimodal fusion mostly adopts static weighting or vector splicing, which cannot adapt to the pathological dominant differences between CD and UC, and is difficult to characterize the complex nonlinear dependence between modalities; (4) The heterogeneity of multimodal data distribution leads to insufficient model generalization ability and poor stability in cross-population validation.
[0004] The disclosed multimodal prediction methods mostly focus on simple feature splicing or single loss optimization, failing to simultaneously address core issues such as high-dimensional feature compression, modality-specific preservation, dynamic fusion adaptation, and enhanced subtype discrimination. For example, some methods lack specific filtering strategies for the high redundancy of enzyme functional data, or lack modality orthogonality constraints leading to information assimilation, or fail to introduce contrastive learning to enhance subtype differentiation, making it difficult to meet the requirements for prediction accuracy, generalization, and clinical applicability. Summary of the Invention
[0005] This application provides a multi-constraint adaptive multimodal fusion method for predicting the risk of inflammatory bowel disease (IBD), addressing the problems of insufficient single-modal information, high-dimensional redundancy, inflexible static fusion, and poor generalization in existing IBD risk prediction methods. Through precise feature selection, intelligent modality fusion, and enhanced discriminative capabilities, it improves the early diagnosis of IBD, including:
[0006] S1. Collect multimodal omics data related to inflammatory bowel disease, and use a three-level bioinformatics filtering strategy for feature preprocessing to obtain low-dimensional, highly discriminative feature sets for each modality;
[0007] S2 extracts latent features from the three modal data of preprocessed microbiome, enzyme functional activity profile, and metabolic pathway abundance through independent encoders, and applies modal orthogonal constraints to preserve the modality-specific information.
[0008] S3 constructs a hierarchical attention mechanism, which filters key features of a single modality through intramodal attention and adaptively allocates the weights of each modality through intermodal attention, thereby achieving two-stage dynamic fusion.
[0009] S4 introduces the maximum mean difference (MMD) loss to align the multimodal feature distribution, and combines supervised contrastive learning to enhance inter-class separability, constructing a multi-objective joint optimization loss function;
[0010] S5, based on the fused features and optimized model, outputs the inflammatory bowel disease risk prediction results for the samples.
[0011] Furthermore, the multimodal omics data collection and preprocessing in step S1 specifically includes: (1) Data sources: collecting three types of omics data from IBD patients and healthy individuals, including microbial community classification, enzyme functional activity profile, and metabolic pathway abundance. The samples include healthy control group and IBD group; (2) Three-level bioinformatics filtering: ① Variance threshold screening: calculating the variance distribution of each feature in all samples, removing the low-fluctuation features in the bottom 15% of variance ranking, and retaining disease-related high-variability features; ② Graph clustering dimensionality reduction: treating features such as enzyme function as nodes, constructing edge weights based on Pearson correlation coefficient (|r|>0.7), dividing functional modules through Louvain algorithm, calculating the KL divergence between features and IBD phenotype in each module, and retaining the representative feature with the highest divergence value; ③ Discriminative screening: calculating the JS divergence between the healthy group and IBD group for the screened features, and retaining features with JS divergence value > 0.25 as the final candidate markers.
[0012] Furthermore, the single-modal encoding and modal orthogonality constraint in step S2 specifically include: (1) The single-modal encoder uses a 3-layer fully connected network, with the input being the preprocessed modal features. The hidden layer dimension gradually decreases from 512 to 128, the activation function is ReLU, and the dropout rate is 0.3, used to extract the low-dimensional latent representation of each modality; (2) The modal orthogonality constraint is achieved through Frobenius norm penalty, which forces the encoder parameter matrices of the three modalities of microbiome, enzyme function, and metabolic pathway to remain approximately orthogonal in the representation space, avoiding modal information assimilation. The loss function is defined as: ,in, These are the parameter matrices for the nth and mth modal encoders, respectively. This represents the Frobenius norm.
[0013] Furthermore, the hierarchical attention dynamic fusion in step S3 specifically includes: (1) Intramodal attention: For the 128-dimensional latent features of each modality, the feature importance weights are calculated through the attention layer: ,in Let be the encoding vector of the k-th feature within the modality. , For attention parameters; , Here are the learnable parameters, σ is the sigmoid activation function, and the weighted intra-modal features are: (2) Intermodal attention: Based on the disease classification and inflammatory activity indicators of the samples, calculate the mutual information between each modality and the prediction target, construct the modal correlation matrix, and generate sample-specific weights through the Softmax function;
[0014] Furthermore, the multi-objective joint optimization in step S4 specifically includes: (1) Weighted cross-entropy loss (main loss): used to optimize the consistency between the predicted probability and the true label, defined as: ,in The true labels for the samples are (0 = healthy, 1 = IBD). (2) MMD loss: used to align the distribution of multimodal fusion features and reduce the impact of data heterogeneity, defined as: (3) Supervised contrastive loss: used to enhance the inter-class separability of IBD subtypes, defined as: ,in For the same type of sample set (same subtype and same state). For heterogeneous sample sets, The normalized embedding vector is generated by the 2-layer MLP projection head, τ=0.07 is the temperature coefficient, and sim(⋅,⋅) is the cosine similarity; (4) Total loss function: The total loss is minimized using the AdamW optimizer.
[0015] Furthermore, the risk prediction output in step S5 specifically includes: (1) inputting the fused feature H into the output layer and outputting the predicted probability of each category through the sigmoid activation function; (2) determining the sample category (healthy / IBD) based on the maximum predicted probability, and outputting a risk score (0-100 points), where a higher score indicates a higher risk of IBD; (3) model evaluation indicators include accuracy, recall, precision, F1 score and AUC value, which are used to quantify prediction performance.
[0016] The beneficial effects of this invention are:
[0017] By addressing issues such as insufficient single-modal information, redundancy of high-dimensional features, rigidity of multimodal static fusion, and poor model generalization in IBD risk prediction, this method achieves feature denoising and dimensionality reduction through three-level bioinformatics filtering to avoid overfitting; retains the specificity of each modality by leveraging modal orthogonal constraints; and completes dynamic adaptive fusion by combining hierarchical attention. Combined with MMD distribution alignment and supervised contrastive learning, it significantly improves the model's generalization ability and IBD subtype discrimination accuracy, enabling non-invasive and accurate early IBD risk prediction and subtyping, providing effective technical support for clinical non-invasive screening and risk stratification. Attached Figure Description
[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0019] Figure 1 This is a structural diagram of a multi-constraint adaptive multimodal fusion method for predicting the risk of inflammatory bowel disease according to the present invention;
[0020] Figure 2 This is a flowchart of a multi-constraint adaptive multimodal fusion method for predicting the risk of inflammatory bowel disease according to the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the following description will be made in conjunction with the accompanying drawings of the present invention (…). Figure 1 This is a structural diagram of a multi-constraint adaptive multimodal fusion method for predicting the risk of inflammatory bowel disease. Figure 2 This is a flowchart of a multi-constraint adaptive multimodal fusion method for predicting the risk of inflammatory bowel disease. The technical solutions of this invention are clearly and completely described. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.
[0022] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein.
[0023] In the field of inflammatory bowel disease (IBD) risk prediction, existing technologies suffer from problems such as insufficient single-modal information, high-dimensional feature redundancy leading to model overfitting, static cross-modal fusion failing to adapt to disease subtype differences, and heterogeneous multimodal feature distribution resulting in poor model generalization ability. To address these issues, this invention proposes a multi-constraint adaptive multimodal fusion method for IBD risk prediction. Through three-level bioinformatics filtering, modal orthogonal constraints, hierarchical attention dynamic fusion, and multi-objective optimization strategies, it achieves accurate IBD risk prediction, improving model stability and discrimination accuracy. The specific implementation of this method is described in detail below with reference to the accompanying drawings.
[0024] Figure 2 This is a flowchart illustrating a multi-constraint adaptive multimodal fusion method for predicting the risk of inflammatory bowel disease provided in an embodiment of the present invention, as shown below. Figure 2 As shown in this embodiment, the inflammatory bowel disease risk prediction method based on adaptive multimodal fusion includes:
[0025] S1: Start the training and prediction process of the inflammatory bowel disease risk prediction model, complete the hardware and software environment deployment before the model runs, including building the model network based on the PyTorch framework, configuring hardware resources such as processors and video memory, and setting the basic parameters for model training.
[0026] S2, Collect IBD multimodal omics data (microbes / enzyme function / metabolic pathways). This step corresponds to... Figure 1 The input layer obtains three types of multimodal omics raw data related to inflammatory bowel disease: microbial community taxonomy, enzyme functional activity profile, and metabolic pathway abundance data.
[0027] S3, Level 3 Bioinformatics Filtering Preprocessing. This step corresponds to... Figure 1 The preprocessing layer performs a three-level bioinformatics filtering process on the raw multimodal omics data collected in step S2. The core purpose is to streamline high-dimensional redundant features, remove noisy features, and retain high-discriminative features related to IBD disease, thereby avoiding overfitting problems in subsequent model training. The specific three-level filtering operations are as follows:
[0028] Variance threshold screening: Calculate the variance distribution of each feature in all samples, and remove the low-fluctuation features in the bottom 15% of variance ranking. These features mostly correspond to the basal metabolic process, which does not change significantly in the IBD disease state and has no disease discrimination value.
[0029] Graph clustering dimensionality reduction: Each enzyme functional feature is regarded as a node. Edge weights are constructed based on the Pearson correlation coefficient of the feature expression profile (|r|>0.7). Functional modules are divided by the Louvain algorithm. In each module, the KL divergence between the feature and the IBD phenotype is calculated. Only the representative features with the highest divergence value and the strongest association with the disease are retained, thus compressing the feature space while preserving the biological significance of the pathway.
[0030] JS divergence screening: The Jensen-Shannon divergence metric was used to measure the distribution distance of each feature between the healthy group and the IBD group. Features with a divergence value > 0.25 were retained as the final candidate biomarkers to enhance the disease discriminative power of the features.
[0031] Meanwhile, feature standardization was performed on the microbial community taxonomy and metabolic pathway abundance data to eliminate the dimensional influence between different features, so that the preprocessed data met the model input requirements.
[0032] S4, Single-modal independent coding feature extraction. This step corresponds to... Figure 1 The single-modal encoder group in the coding layer performs feature extraction and low-dimensional mapping on the three types of modal data after preprocessing in step S3 using independent single-modal encoders. Each single-modal encoder employs a 3-layer fully connected network structure with ReLU activation function, a Dropout rate of 0.3, and hidden layer dimensions gradually decreasing from 512 to 128. Through nonlinear transformation of the multi-layer neural network, the high-dimensional preprocessed features are mapped to a 128-dimensional low-dimensional space, obtaining the single-modal potential features of the microbiome, enzyme function, and metabolic pathways, achieving feature compression and representation, and providing a standardized feature foundation for subsequent cross-modal fusion.
[0033] S5, Apply modal orthogonality constraints to preserve specificity. This step corresponds to... Figure 1 The modality orthogonality constraint module of the encoding layer applies orthogonality constraints to the parameter matrices of each modality encoder after single-modality feature encoding is completed in step S3. This constraint is achieved through Frobenius norm penalty, forcing the feature representation spaces of the three modalities—microbiome, enzyme function, and metabolic pathway—to maintain an approximately orthogonal relationship. The core purpose of this operation is to prevent the feature information of different modalities from assimilizing each other during model training, to retain the independent and specific information of each modality to the maximum extent, and to ensure that multimodal data can achieve information complementarity during fusion, rather than being dominated or covered by information from a single modality.
[0034] S6, Hierarchical attention dynamic fusion of multimodal features. This step corresponds to... Figure 1 The fusion layer performs hierarchical dynamic attention fusion on the single-modal latent features constrained by orthogonality in step S5. This fusion process is divided into two stages: intra-modal attention filtering and inter-modal attention weight allocation, achieving intelligent filtering and adaptive integration of multimodal features.
[0035] Intramodal attention: For the 128-dimensional latent features of each modality, the importance weight of the features is calculated through the attention layer. The weight is obtained by normalization by the sigmoid function. Based on the weight, the biomarker features that play a key role in IBD risk prediction in a single modality are selected, and the secondary features within the modality are removed.
[0036] Intermodal attention: Based on information such as disease classification and inflammatory activity indicators of the samples, the mutual information between each modality and the prediction target is calculated, a modality correlation matrix is constructed, and sample-specific modality fusion weights are generated through the Softmax function. For example, the weight of the microbiome is increased for CD patient samples, and the weight of metabolic pathways is increased for UC patient samples. Finally, the key features of the selected single modality are integrated according to the weights to obtain the fused global features.
[0037] S7, MMD loss alignment feature distribution + supervised contrastive learning optimization of this step Figure 1 The optimization layer performs dual optimization on the global fusion features obtained in step S6, adjusting model parameters through a multi-objective loss function to improve the model's generalization ability and IBD subtype discrimination accuracy. Specifically, this includes:
[0038] MMD loss aligns feature distribution: The maximum mean difference (MMD) loss is introduced to calculate the distribution difference after the fusion of different modal features. By optimizing this loss, the multimodal feature distribution tends to be aligned, which alleviates the problem of insufficient model generalization ability caused by data distribution heterogeneity.
[0039] Supervised contrastive learning optimization: Introducing supervised contrastive loss forces the features of similar samples (IBD samples of the same subtype and disease state) to cluster together in the latent feature space, while the features of dissimilar samples (such as healthy samples and CD samples, CD samples and UC samples) are far apart, thereby strengthening the inter-class separability of features and improving the discrimination confusion problem of different IBD subtypes.
[0040] In this step, the model integrates the weighted cross-entropy loss, modal orthogonality loss, MMD loss, and supervised contrastive loss into a total loss function. The AdamW optimizer is used to minimize the total loss, thereby achieving iterative updates of the model parameters.
[0041] S8, Output IBD risk prediction results (healthy / IBD). This step corresponds to... Figure 1The output layer of the model takes the global fusion features optimized in step S7 as input and maps the fusion features to the predicted probabilities of the sample in the healthy and IBD categories through the sigmoid activation function. The disease classification result of the sample is determined based on the maximum value of the predicted probability. At the same time, a risk score of 0-100 can be generated based on the predicted probability value. The higher the score, the higher the risk of IBD in the sample. Finally, the disease classification result and risk score are output simultaneously to complete the prediction of the inflammatory bowel disease risk of the sample.
[0042] S9 completes the single inflammatory bowel disease risk prediction process. If it is the model training stage, return to step S2 to continue iterative training until the model converges to the preset accuracy. If it is the actual prediction stage, output the prediction result and terminate the process.
[0043] The inflammatory bowel disease risk prediction method based on adaptive multimodal fusion provided in this invention, through the orderly execution of the above steps, combined with... Figure 1 This hierarchical model structure enables end-to-end processing of multimodal omics data from raw input to risk outcome output, addressing issues such as insufficient single-modal information, redundancy of high-dimensional features, and static fusion in existing IBD risk prediction technologies. This method balances the specificity and complementarity of multimodal features, effectively improving the model's accuracy and generalization ability for early IBD diagnosis and subtype differentiation, providing a reliable technical solution for non-invasive clinical screening and risk stratification of inflammatory bowel disease.
[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A multi-constraint adaptive multimodal fusion method for predicting the risk of inflammatory bowel disease, characterized in that, Includes the following steps: (a) Data collection: Collect multimodal omics data related to inflammatory bowel disease, wherein the multimodal omics data are omics data related to the pathogenesis of inflammatory bowel disease; (b) Data preprocessing: The multimodal omics data are preprocessed using a three-level bioinformatics filtering strategy to remove redundant and outlier data and obtain standardized data for each modality. (c) Modal encoding and constraints: Feature encoding is performed on each modal omics data after preprocessing, and modal orthogonality constraints are applied to each encoded modal feature to preserve the specificity of each modal data; (d) Adaptive multimodal fusion: The encoded modal features are dynamically fused in two stages through a hierarchical attention mechanism, and the weights of each modal feature are adaptively allocated to obtain the fused features; (e) Feature optimization and prediction: Supervised contrastive learning and maximum mean difference (MMD) loss are introduced to optimize the fusion features, enhance feature discriminativeness and distribution consistency, and output inflammatory bowel disease risk prediction results based on the optimized fusion features. The prediction results include two categories: healthy and IBD.
2. The method according to claim 1, characterized in that, The multimodal omics data mentioned in step (a) includes at least two of the following: microbiome data, enzyme functional activity profile data, metabolome data, and proteome data.
3. The method according to claim 1, characterized in that, The three-level bioinformatics filtering strategy described in step (b) is as follows: the first level of filtering removes data samples with a missing value ratio of more than 30%; the second level of filtering removes outlier samples; and the third level of filtering removes redundant features through variance analysis algorithm or mutual information method.
4. The method according to claim 1, characterized in that, The feature encoding in step (c) is implemented using any one of the deep neural network models, such as convolutional neural network (CNN), recurrent neural network (RNN), or Transformer model; the modal orthogonality constraint is implemented by applying an orthogonal regularization term to each modal encoded feature to ensure the linear independence between different modal features.
5. The method according to claim 1, characterized in that, The two-stage dynamic fusion described in step (d) specifically involves: in the first stage, preliminary fusion of the features of each modality is performed to obtain intermediate fused features; The second stage uses a hierarchical attention mechanism to perform secondary fusion of intermediate fused features and original modality coding features, and dynamically adjusts the weights of each feature.
6. The method according to claim 1, characterized in that, The supervised contrastive learning in step (e) uses a supervised contrastive loss function to narrow the feature distance between similar samples and widen the feature distance between dissimilar samples; the maximum mean difference (MMD) loss is used to minimize the difference between the fused features and the true label distribution, thereby improving the consistency of feature distribution; the risk prediction result is output through a softmax classifier, which outputs the probability value of each category, and the final risk classification result is determined based on the maximum probability value.