Information completion and subtype analysis method for clinical data of multi-modal incomplete type 2 diabetes mellitus
By employing cross-view subtype prototype heterogeneous graphs and graph neural networks, the problem of information completion and subtype analysis in multimodal incomplete type 2 diabetes data was solved. Stable subtyping and interpretable subtype feature descriptions were achieved under high missing rates, supporting personalized diagnosis and treatment and large-scale dataset analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG UNIV
- Filing Date
- 2026-03-25
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies cannot effectively process multimodal and incomplete clinical data on type 2 diabetes, resulting in poor interpretability of subtype characteristics and insufficient fusion of information between views, which makes it impossible to achieve accurate stratification and personalized diagnosis and treatment for patients.
Through data preprocessing, view missing labeling and feature encoding, a cross-view subtype prototype heterogeneous graph is constructed and a relational graph convolutional network is used for feature updating. Missing values are filled in by combining granularity matching degree, and finally K-means clustering is performed to output subtype labels and clinical descriptions.
It achieves stable typing accuracy under high missing rates, provides interpretable subtype feature descriptions, supports personalized diagnosis and treatment, and is compatible with large-scale clinical dataset analysis.
Smart Images
Figure CN122050665A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of data mining and medical clinical data analysis, and in particular relates to a method for information completion and subtype analysis of multimodal incomplete clinical data of type 2 diabetes. Background Technology
[0002] Type 2 diabetes mellitus (T2DM) is a chronic metabolic disease caused by a combination of genetic and environmental factors, and has become a major public health problem in my country. The latest epidemiological data shows that the prevalence of T2DM in people aged 18 and above in my country has reached 11.2%, with a total of over 140 million patients. More than 60% of these patients have varying degrees of microvascular and macrovascular complications, making it one of the leading causes of end-stage renal disease, stroke, myocardial infarction, and blindness.
[0003] Type 2 diabetes exhibits highly heterogeneous clinical patterns; patients with the same blood glucose level can vary significantly in their risk of complications, rate of disease progression, and response to drug treatment. Traditional clinical classifications only categorize diabetes into type 1, type 2, special types, and gestational diabetes, failing to identify the biological subtypes within type 2 diabetes and hindering precise patient stratification and personalized treatment. Therefore, unsupervised subtyping of type 2 diabetes patients based on multimodal clinical data has become a research hotspot in the field of endocrinology.
[0004] In real clinical scenarios, the data of patients with type 2 diabetes are naturally multimodal and multimodal, with four core clinical views: (1) biochemical test view (fasting blood glucose, glycated hemoglobin, blood lipids, liver and kidney function, inflammatory markers, etc.); (2) complication marker view (urine microalbumin / creatinine ratio, fundus lesion grade, carotid intima-media thickness, etc.); (3) imaging view (abdominal ultrasound, cardiac ultrasound, fundus photography, etc.); (4) medical history and clinical scale view (diabetes course, medication history, history of comorbidities, quality of life scale, self-management ability scale, etc.). However, due to limitations such as the distribution of medical resources, patients' economic conditions, physical tolerance, and follow-up compliance, this multimodal data suffers from severe incompleteness: patients visiting primary hospitals generally lack imaging and scale view data, outpatient follow-up patients lack dynamic detection data of complication markers, and approximately 40% to 70% of patients have at least one view feature missing. Traditional multimodal clustering methods are all designed based on the assumption that "all view information of all samples is complete", which cannot directly adapt to the classification needs of such incomplete data. Summary of the Invention
[0005] Purpose of the invention: The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for information completion and subtype analysis of incomplete multimodal clinical data of type 2 diabetes, aiming to solve the problems of missing multimodal data, poor interpretability of subtype features, and insufficient information fusion between views in the prior art.
[0006] This invention specifically provides a method for information completion and subtype analysis of multimodal incomplete clinical data on type 2 diabetes, comprising the following steps:
[0007] Step S10, Data Preprocessing, View Missing Labeling and Feature Encoding: Input an incomplete multimodal clinical dataset of type 2 diabetes patients, complete data cleaning, view missing labeling, view-specific feature encoding and cross-view latent space alignment, and obtain feature representations of all samples in a unified low-dimensional latent space;
[0008] Step S20: Construction of view-specific subtype granular prototypes: Independently run fuzzy C-means clustering on the complete samples of each view to extract subtype membership, and construct super-rectangular granular prototypes of each subtype on each view based on weighted median and interval optimization.
[0009] Step S30, Construction and Evolution of Cross-View Subtype Prototype Heterogeneous Graph: Construct a heterogeneous graph containing sample nodes and prototype nodes, design a two-layer relational graph convolutional network to update node features, and fuse semantic associations between views;
[0010] Step S40, Granularity-based missing view imputation: For patients with missing views, the granularity matching degree between the complete features of the patients with missing views in other views and each subtype is calculated, the optimal subtype is selected and the missing value is filled with the midpoint of the optimal subtype prototype interval.
[0011] Step S50, Subtype Classification and Semantic Output: K-means clustering is performed using the imputed complete dataset, with the evolved prototype centers as the initial seeds, to output patient subtype labels and interpretable granular clinical descriptions.
[0012] Step S10 includes the following steps:
[0013] Step S11: Input an incomplete multimodal clinical dataset of type 2 diabetes patients. ,in The preset number of clinical views. These correspond to the biochemical test view, complication marker view, imaging view, and medical history and scale view, respectively. For the first The feature matrix of each view, where Represents the space of real numbers. The total number of patient samples. For the first Clinical feature dimensions of each view; if patient sample In the If a view has missing features, then record ;
[0014] Step S12: Generate the missing view index matrix , Indicates patient sample In the The clinical features of each view are complete. Indicates feature loss;
[0015] Step S13, Generation and The corresponding missing mask matrix is used to perform min-max normalization on the complete clinical features of each view, mapping all feature values to the [0,1] interval; at the same time, the classification features are one-hot encoded, and outliers are truncated using the quartile method to complete the clinical data cleaning.
[0016] Step S14, Design and Training of View-Specific Autoencoders: For each clinical view Design an independent view-specific autoencoder, which consists of two parts: an encoder and a decoder.
[0017] The encoder formula is:
[0018] (1),
[0019] in, No. The first clinical view Clinical feature vectors of each sample; For the first Each clinical view has its own dedicated encoder. For the first The first clinical view The latent feature vectors obtained by encoding each sample using an encoder; For dimension The real number space is used to limit the range of values for the latent feature vectors;
[0020] The decoder formula is:
[0021] (2),
[0022] in, For the first Each clinical view has its own dedicated decoder; For the first The first clinical view The reconstructed feature vector obtained by the decoder from each sample;
[0023] The autoencoder is trained with the goal of minimizing the view reconstruction loss.
[0024] (3),
[0025] in, For the first The reconstruction loss function for each clinical view; For the first Total number of samples for each clinical view; using The optimizer is trained iteratively with a learning rate of 0.001 and 150 iterations. After convergence, the trained view-specific encoder and decoder are obtained.
[0026] Step S20 includes the following steps:
[0027] Step S21: For each clinical view Extract the aligned latent feature set of the complete sample of the view. ;
[0028] in, For the first The complete set of latent features after alignment of each clinical view; For the first The first clinical view Aligned latent feature vectors corresponding to each patient sample; As an indicator of sample integrity, Indicates the first The patient sample in the first Each clinical view contains complete samples without missing data;
[0029] Step S22: Complete sample set for each view Independently run the fuzzy C-means clustering algorithm, with a preset number of subtypes. Iterative optimization yields the first... Membership matrix of each view and the membership degree of each patient sample to each subtype. ,in For patient sample indexing, For subtype index, For clinical view indexes, the membership degree satisfies the constraint. and ;
[0030] in, For the first A complete set of clinical features corresponding to each clinical view; K is the preset number of disease subtypes; For the first The membership matrix is obtained by fuzzy C-means clustering of each clinical view, and the matrix elements are the membership degree of each sample to each subtype. For the first In the first clinical view, the... The patient sample corresponds to the first The membership degree of a disease subtype is used to represent the probability that the sample belongs to the corresponding subtype.
[0031] Step S23: Calculate the weights for each subtype using membership degree as the weight. In view Each clinical feature dimension weighted median , serving as the interval anchor points for the feature dimensions; the weighted median is calculated by: assigning each clinical feature dimension to an interval anchor point; The clinical feature values are arranged in ascending order, and the cumulative membership degree and the feature value corresponding to 50% of the total membership degree are taken as the weighted median;
[0032] Step S24: Based on the principle of reasonable granularity, for each subtype In view Each clinical feature dimension Search optimization of the optimal interval The optimization objective is to maximize the joint metrics of coverage and specificity, and the optimization function is... for:
[0033] (4),
[0034] in, For the first The first clinical view, the first The disease subtype, the Coverage corresponding to clinical features; For the first The first clinical view, the first The disease subtype, the Subtype specificity corresponding to clinical characteristics;
[0035] The calculation formula is:
[0036] (5),
[0037] in This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. For patient samples In view No. Values of clinical features; For the first The first clinical view, the first The disease subtype, the The left endpoint of the optimal interval corresponding to the clinical characteristics; For the first The first clinical view, the first The disease subtype, the The right endpoint of the optimal interval corresponding to the clinical features;
[0038] For specificity, it is defined as an exponential decay function of the interval length, and the calculation formula is:
[0039] (6),
[0040] in is a hyperparameter balancing coverage and specificity; exp is the natural exponential function;
[0041] Step S25: Place each subtype c in the view The Cartesian product of the optimal intervals for all clinical feature dimensions yields subtype c in the view. Below 3D hyperrectangular granular prototype :
[0042] (7),
[0043] in, For the first The first clinical view The granular prototype corresponding to each disease subtype is a high-dimensional hyperrectangular space composed of the optimal intervals of each feature dimension. For the first The total dimensions of clinical features in each clinical view; This is the Cartesian product operation, used to combine multiple single-dimensional intervals into a high-dimensional hyperrectangular space;
[0044] Simultaneously, the interval anchor point The resulting vector serves as the central feature of the granular prototype and is used for initializing the node features of the subsequent heterogeneous graph.
[0045] Step S30 includes the following steps:
[0046] Step S31: Construct a cross-view subtype prototype heterogeneous graph :
[0047] in, For constructing a cross-view subtype prototype heterogeneous graph; It is a collection of nodes in a heterogeneous graph, containing patient sample nodes and subtype prototype nodes for each view; It is a set of edges of a heterogeneous graph, including two types: sample-prototype edges and prototype-prototype edges; This is the set of edge weights for a heterogeneous graph, used to represent the strength of the association between nodes;
[0048] Sample-prototype edge: connects sample nodes and view prototype node Weight ; For patient sample nodes With the The first view Edge weights between subtype prototype nodes; For the first In the first clinical view The patient sample corresponds to the first The membership degree of a disease subtype, the higher the weight, the stronger the association between the sample and the prototype;
[0049] Prototype-prototype edge: Connects the corresponding prototype nodes of different views, with weights initialized to... ;
[0050] in, For the first The first view The prototype node of the subtype and the first The first view Edge weights between subtype prototype nodes; This is the Gaussian kernel bandwidth parameter, used to adjust the rate at which the weights decay with feature distance; The square of the Euclidean distance between the two prototype center feature vectors is given. The closer the distance, the stronger the association between the same subtype prototypes across views and the higher the edge weight.
[0051] Step S32: Design a graph neural network (GNN) structure that includes a two-layer relational graph convolutional network (RGCN) as a prototype update. Here, RGCN is a relational graph convolutional network, which is a graph convolution operator that adapts to multiple types of edges in heterogeneous graphs. The graph neural network is used to propagate node information and update the feature representation of prototype nodes on heterogeneous graphs.
[0052] First-level message passing: For views prototype node Aggregate prototype nodes Information about the connected sample nodes:
[0053] (8),
[0054] in, For the first The first view The feature vectors of each subtype prototype node; AGG is an abbreviation for aggregation function, used to merge the information of multiple neighboring nodes into the update information of the target node. Here, the aggregation function uses a weighted sum. Represents the prototype node The set of connected sample nodes is aggregated using a weighted sum function; For the first In the first clinical view The patient sample corresponds to the first The membership degree of each disease subtype is used as the weight for information aggregation; For the first Feature vectors of each patient sample node; This means that the information obtained from the aggregation on the right is assigned to the feature of the prototype node on the left, thus completing the information update;
[0055] For sample nodes Aggregate sample nodes Information about the connected prototype nodes:
[0056] (9),
[0057] in To be related to sample nodes The set of all connected subtype prototype nodes; For the first The first view The feature vectors of each subtype prototype node; the aggregation function uses a weighted sum;
[0058] Second-level message passing: For views prototype node Aggregate information from the prototype nodes of other views:
[0059] (10)
[0060] in, For view With View Attention weights or fixed similarity weights are used to measure the strength of association between different views and adjust the weights of cross-view information aggregation. For the first The first view Feature vectors of each subtype prototype node; This indicates that the aggregation scope is excluding the current view. In addition, all other views have the same subtype prototype nodes; the meanings of the remaining symbols are consistent with the aforementioned formulas;
[0061] Step S33, Node Update: Update node features using residual connections, calculated as follows:
[0062] (11),
[0063] in, For the updated number The first view Feature vectors of each subtype prototype node; This is a node update function implemented using residual connections. It is used to fuse the aggregated neighbor node information with the original features of the node itself to complete the feature update. The original feature vector of the prototype node before the update; the aggregated information is the neighbor node information obtained by aggregation in formulas (8) and (10);
[0064] Step S34, Graph Neural Network (GNN) Training and Optimization: With the goal of minimizing the overall loss function, the Adam optimizer is used to iteratively optimize the learnable weight matrix of the GNN. During the training process, the sample node features, sample-prototype edge weights, and original membership matrix remain fixed throughout. Only the learnable weights and prototype node features of the GNN are updated. After iterative convergence, the final optimized global granular prototype is obtained.
[0065] Overall loss function The calculation formula is:
[0066] (12),
[0067] in, The balancing hyperparameter for cross-view clustering consistency loss; The balancing hyperparameter for L2 weighted regularization loss;
[0068] Cross-view clustering consistency loss The calculation formula is:
[0069] (13)
[0070] in, The total number of pairwise combinations of views. For patients In view The subtype membership vector, For patients In view Subtype membership vector;
[0071] L2 weighted regularization loss The calculation formula is:
[0072] (14)
[0073] in, For all learnable weight matrices in the graph encoder, Let f be the Frobenius norm of the matrix.
[0074] Step S40 includes the following steps:
[0075] Step S41: Iterate through all patient samples. If the patient is in the clinical view... Feature loss, i.e. Extract the patient's complete clinical features across all other views. ;
[0076] Step S42: Calculate patient samples With each subtype Granularity matching degree on view m :
[0077] (15)
[0078] in, For patients to have complete characteristics falling into the subtype The feature percentage within the corresponding view prototype range is specifically: statistics. Each feature dimension Does the value fall within the subtype? In view On the prototype interval Within, the percentage is the number of features that meet the conditions divided by The total feature dimension; For patients Except for views Complete feature splicing vectors from all views outside the scope; subtype The prototype center concatenation vector of the corresponding view. This is the overlap weighting coefficient. This is a distance adjustment parameter;
[0079] Step S43: Select the subtype with the highest granularity matching degree as the patient sample. Subtype of belonging on missing view m ;
[0080] Step S44, based on the subtype In view Granulated prototypes on Based on this, the midpoint of each dimension interval is taken as the fill value for the missing feature;
[0081] Step S45: Repeat steps S41 to S44 to impute all missing views for all patient samples, thus obtaining a complete multimodal clinical dataset of type 2 diabetes patients. .
[0082] Step S43: Obtain the subtype using the following formula. :
[0083] (16)
[0084] in, The disease subtype number to which the patient sample to be classified is ultimately assigned; To find the independent variable that maximizes the objective function The operation involves iterating through all preset disease subtypes. Find the subtype number that maximizes the matching score, and use it as the final subtype to which the sample belongs; For patient samples to be classified Corresponding to the A matching score function for each disease subtype; the higher the matching score, the higher the degree of matching between the sample and the corresponding disease subtype.
[0085] In step S44, for the d-th dimension feature of view m, the fill value is:
[0086] (17)
[0087] in and Subtypes The lower and upper bounds of the interval on the d-th dimension feature of view m.
[0088] Step S50 includes the following steps:
[0089] Step S51: Fill in the complete multimodal clinical dataset with the completed data. Input the K-means clustering algorithm, using the updated prototype node features obtained in step S30. As the initial seed for clustering, it is used to perform the final patient subtype classification;
[0090] Step S52: Output the final subtype label for each patient sample, and simultaneously generate a granular semantic description of the clinical features of each subtype, in the following format:
[0091] Subtype c: Glycated hemoglobin ∈ [value missing] in biochemical test view %, fasting blood glucose ∈ mmol / L; urinary microalbumin / creatinine ratio in the complication marker view ∈ mg / g; corresponding to a high, medium, or low risk of developing diabetic nephropathy;
[0092] in , They represent subtypes respectively. In view No. Lower and upper bounds of the granularity prototype interval on the dimensional feature;
[0093] Step S53: Based on the granular prototype features of each subtype, complete the risk stratification of diabetic microvascular and macrovascular complications, and use the subtype label, granular feature description and risk stratification results as the final output.
[0094] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores program code that, when executed by the processor, causes the processor to perform the steps of the method.
[0095] The present invention also provides a storage medium storing a computer program or instructions that, when executed on a computer, perform the steps of the method.
[0096] The present invention has the following beneficial effects: it is adapted to scenarios with incomplete clinical data and has strong robustness in subtyping under high missing rates: the present invention addresses the common problem of missing views in clinical data of type 2 diabetes. It achieves end-to-end joint optimization of missing feature completion and patient subtyping through granular prototypes, avoiding the accumulation of errors in two-stage methods; the prototype based on reasonable granular optimization carries complete feature distribution information of subtypes, and its fault tolerance to noise, outliers and missing data in clinical data is much stronger than that of traditional single-point prototypes. It can still maintain stable subtyping accuracy in scenarios with high missing view rates of more than 50%.
[0097] The classification results have complete clinical interpretability and are well-suited to clinical diagnosis and treatment needs: This invention outputs the numerical range of each patient subtype in various clinical views and test indicators through granular prototype output, which can be directly converted into a subtype feature description that clinicians can understand. This solves the core pain point of traditional deep learning classification methods being "black box and uninterpretable". The classification results can be directly used to guide personalized clinical diagnosis and treatment.
[0098] Accurate identification of heterogeneous subtypes of type 2 diabetes patients and precise stratification of complication risks: This invention fully integrates complementary information from multiple dimensions, including patient biochemistry, imaging, medical history, and complication biomarkers, through cross-view prototype heterogeneous graphs and graph neural network (GNN) updates. It can accurately identify heterogeneous subtypes within type 2 diabetes, clarify the risk of complications in different subtypes, and provide an objective basis for early risk warning, personalized medication, and follow-up management of patients.
[0099] High computational efficiency and adaptability to large-scale clinical datasets: The method of this invention has controllable computational complexity, eliminates the need for repeated K-means clustering, and avoids computational redundancy in traditional methods; at the same time, it can realize the typing processing of ultra-large-scale diabetes cohort data through distributed computing, adapt to the analysis needs of multi-center clinical datasets across the country, and has strong clinical applicability. Attached Figure Description
[0100] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.
[0101] Figure 1 This is a schematic diagram of the overall process of the information completion and subtype analysis method for multimodal incomplete clinical data of type 2 diabetes of the present invention.
[0102] Figure 2 This is a diagram illustrating the overall data processing framework of the present invention's method for information completion and subtype analysis of multimodal incomplete clinical data of type 2 diabetes.
[0103] Figure 3 This is a framework diagram of the information completion and subtype analysis method for multimodal incomplete clinical data of type 2 diabetes provided by the present invention. Detailed Implementation
[0104] like Figure 1 , Figure 2 and Figure 3 As shown: This invention provides a method for information completion and subtype analysis of multimodal incomplete clinical data on type 2 diabetes, including the following steps:
[0105] S10. Data Preprocessing, View Missing Labeling and Feature Encoding: Input an incomplete multimodal clinical dataset of type 2 diabetes patients, complete data cleaning, view missing labeling, view-specific feature encoding and cross-view latent space alignment, and obtain feature representations of all samples in a unified low-dimensional latent space.
[0106] S20. Granular Initial Prototype Generation: Based on the preprocessed incomplete multimodal data, the membership matrix of the patient sample-subtype prototype is obtained through fuzzy C-means clustering. The weighted median of each clinical feature of each subtype is calculated as the interval anchor point using the membership degree as the weight. The upper and lower bounds of the intervals of each feature dimension are optimized based on the principle of reasonable granularity to generate the granular prototype of each patient subtype in each view.
[0107] S30. Construction of cross-view subtype prototype heterogeneous graph and updating of graph encoder: Construct a cross-view heterogeneous graph containing sample nodes and prototype nodes, design a two-layer graph neural network (GNN) as a graph encoder, perform message passing and feature updates only on prototype nodes, keep the features of sample nodes unchanged, fuse cross-view semantic associations, and obtain globally optimized granular prototypes.
[0108] S40. Missing view imputation based on granularity matching degree: For patient samples with missing views, calculate the granularity matching degree between their features on the existing views and the granular prototype of each subtype, select the optimal subtype with the highest matching degree, and use the midpoint of the prototype interval of the subtype on the missing view to imput the patient's missing clinical features.
[0109] S50. Final Subtyping and Output: Input the completed multimodal clinical data and the optimized granular prototype into the clustering algorithm, perform the final patient subtyping, and output the subtype label for each patient, the granular description of the clinical characteristics of each subtype, and the risk stratification results of diabetic complications for the corresponding subtype.
[0110] Step S10 specifically includes the following sub-steps:
[0111] S11. Input an incomplete multimodal clinical dataset of type 2 diabetes patients. ,in The total number of clinical views is preset in this invention. These correspond to the biochemical test view, complication marker view, imaging view, and medical history and scale view, respectively. For the first Feature matrix of each view The total number of patient samples. For the first Clinical feature dimensions of each view; if patient sample In the If a view has missing features, then record ;
[0112] S12. Generate the missing index matrix of the view. ,in Indicates patient sample In the The clinical features of each view are complete. Indicates feature loss;
[0113] S13, Generation and The corresponding missing mask matrix performs min-max normalization on the complete clinical features of each view, mapping all feature values to the [0,1] interval to eliminate the dimensional differences of different clinical test indicators; at the same time, the classification features are one-hot encoded, and outliers are truncated using the quartile method to complete the clinical data cleaning.
[0114] S14. Design and training of view-specific autoencoders: for each clinical view Design an independent view-specific autoencoder, which consists of two parts: an encoder and a decoder.
[0115] The encoder formula is:
[0116] (1),
[0117] The decoder formula is:
[0118] (2),
[0119] The autoencoder is trained with the goal of minimizing the view reconstruction loss.
[0120] (3),
[0121] in For view The number of complete samples; using The optimizer is trained iteratively with a learning rate of 0.001 and 150 iterations. After convergence, the trained view-specific encoder and decoder are obtained.
[0122] Step S20 includes the following steps:
[0123] S21. For each clinical view Extract the aligned latent feature set of the complete sample of the view. ;
[0124] S22, Complete sample set for each view Run the fuzzy C-means (FCM) clustering algorithm independently, with a preset number of subtypes. Iterative optimization yields the first... Membership matrix of each view and the membership degree of each patient sample to each subtype. ,in For patient sample indexing, For subtype index, For clinical view indexes, the membership degree satisfies the constraint. and ;
[0125] S23. Calculate the weights for each subtype using membership degree as the weight. In view Each clinical feature dimension weighted median The anchor point for this feature dimension is used as the interval anchor point; the weighted median is calculated as follows: the clinical feature values of this dimension are arranged in ascending order, and the feature value corresponding to the cumulative membership degree and the total membership degree of 50% is taken as the weighted median.
[0126] S24. Based on the principle of reasonable granularity, for each subtype... In view Each clinical feature dimension Search optimization of the optimal interval The optimization objective is to maximize the joint metrics of coverage and specificity, and the optimization function is:
[0127] (4),
[0128] in, Coverage is defined as the weighted sum of the membership degrees of patient samples within the interval [ac,dv,bc,dv], and is calculated using the following formula:
[0129] (5),
[0130] in This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. For patient samples In view No. Values of clinical features;
[0131] For specificity, it is defined as an exponential decay function of the interval length, and the calculation formula is:
[0132] (6),
[0133] In the formula This is a hyperparameter that balances coverage and specificity; the default value is 0.5.
[0134] S25. Place each subtype c in the view. The Cartesian product of the optimal intervals for all clinical feature dimensions yields the subtype under this view. 3D hyperrectangular granular prototype , represented as:
[0135] (7),
[0136] Simultaneously, the interval anchor point The resulting vector serves as the central feature of the granular prototype and is used for initializing the node features of the subsequent heterogeneous graph.
[0137] Step S30 includes the following steps:
[0138] S31. Construct a cross-view subtype prototype heterogeneous graph :
[0139] Sample-prototype edge: connects sample nodes and view prototype node Weight ;
[0140] Prototype-prototype edge: Connects the corresponding prototype nodes of different views, with weights initialized to... ;
[0141] S32. Design a graph neural network (GNN) structure that includes a two-layer relational graph convolutional network (RGCN) as the prototype update.
[0142] First-level message passing: For views prototype node Aggregate the information of the sample nodes it is connected to:
[0143] (8),
[0144] in Represents the prototype node The set of connected sample nodes is aggregated using a weighted sum function;
[0145] For sample nodes Aggregate the prototype node information it is connected to:
[0146] (9),
[0147] Aggregate functions use weighted sums;
[0148] Second-level message passing: For views prototype node Aggregate information from the prototype nodes of other views:
[0149] (10)
[0150] in For view With View Attention weights or fixed similarity weights between them;
[0151] S33. Node Update: Node features are updated using residual connections. The calculation formula is as follows:
[0152] (11),
[0153] S34, GNN Training Optimization: With the goal of minimizing the overall loss function, the Adam optimizer is used to iteratively optimize the learnable weight matrix of the GNN. During the training process, the sample node features, sample-prototype edge weights, and original membership matrix remain fixed throughout. Only the learnable weights and prototype node features of the GNN are updated. After the iteration converges, the final optimized global granular prototype is obtained.
[0154] Furthermore, the formula for the overall loss function is:
[0155] (12),
[0156] in, This is a balancing hyperparameter for cross-view clustering consistency loss, with a default value of 0.3; This is the balancing hyperparameter for L2 weighted regularization loss, with a default value of 0.001;
[0157] The formula for the cross-view clustering consistency loss is:
[0158] (13)
[0159] In the formula, The total number of pairwise combinations of views. For patients In view The subtype membership vector, For patient i in view The subtype membership vector remains fixed throughout the training process;
[0160] The formula for the L2 weighted regularization loss is:
[0161] (14)
[0162] in, For all learnable weight matrices in the GNN graph encoder, Let f be the Frobenius norm of the matrix.
[0163] Step S40 includes the following steps:
[0164] S41. Iterate through all patient samples, for each patient... If it is in clinical view Feature loss on (i.e.) Extract the patient's complete clinical features across all other views, and record them as follows: (Concatenating vectors);
[0165] S42, Calculate patient samples With each subtype The granularity matching degree on view m is calculated by combining the overlap between the patient's complete features and the subtype prototype region, and the Euclidean distance between the patient's features and the prototype center. The formula is:
[0166] (15)
[0167] In the formula, For patients to have complete characteristics falling into the subtype The feature percentage within the corresponding view prototype range is specifically: statistics. Each feature dimension d (belonging to the view) Does the value of ) fall within the prototype range of subtype k on this view? Within, the percentage is the number of features that meet the conditions divided by The total feature dimension. This is the complete feature splicing vector of patient i across all views except view m. This is the prototype center splicing vector of the view corresponding to subtype c. This is the overlap weighting coefficient, with a default value of 0.6. This is a distance adjustment parameter; the default value is 1.0.
[0168] S43. Select the subtype with the highest granularity matching degree as the subtype to which patient sample i belongs on the missing view m, as shown in the following formula:
[0169] (16)
[0170] S44, by subtype Granulated prototype on view m Based on this, the midpoint of each dimension interval is taken as the fill value for the missing feature. For the d-th dimension feature of view m, the fill value is:
[0171] (17)
[0172] in and Subtypes The lower and upper bounds of the interval on the d-th dimension feature of view m.
[0173] S45. Repeat steps S41-S44 to impute all missing views for all patient samples, obtaining a complete multimodal clinical dataset of type 2 diabetes patients. .
[0174] Step S50 includes the following steps:
[0175] S51. Complete the imputed multimodal clinical dataset Input the K-means clustering algorithm, using the updated prototype node features obtained in step S30. (or the original anchor vector) () is used as the initial seed for clustering, and the final patient subtype classification is performed;
[0176] S52. Output the final subtype label for each patient sample, and simultaneously generate a granular semantic description of the clinical features of each subtype, in the following format:
[0177] Subtype c: Glycated hemoglobin ∈ [value missing] in biochemical test view %, fasting blood glucose ∈ mmol / L; urinary microalbumin / creatinine ratio in the complication marker view ∈ mg / g; corresponding to a high / medium / low risk of developing diabetic nephropathy.
[0178] in , They represent subtypes respectively. In view No. Lower and upper bounds of the granularity prototype interval on the dimensional feature;
[0179] S53. Based on the granular prototype features of each subtype, complete the risk stratification of diabetic microvascular and macrovascular complications, and use the subtype label, granular feature description and risk stratification results as the final output.
[0180] Example 2: Based on Example 1, the specific content of the multimodal data is further defined. The biochemical test views include indicators such as glycated hemoglobin, fasting blood glucose, 2-hour postprandial blood glucose, triglycerides, total cholesterol, high-density lipoprotein cholesterol, and low-density lipoprotein cholesterol; the complication marker views include the urinary microalbumin / creatinine ratio, serum creatinine, estimated glomerular filtration rate, cystatin C, fundus examination results (such as diabetic retinopathy grading), and foot vibration sensation threshold; the imaging views include abdominal fat CT measurements (visceral fat area, subcutaneous fat area), liver ultrasound fat grading, and carotid intima-media thickness; the medical history and scale views include the duration of diabetes, body mass index, waist circumference, systolic blood pressure, diastolic blood pressure, smoking history, alcohol consumption history, family history of diabetes, diet score, and physical activity score. Each view's feature dimensions... It can be adjusted according to the actual data source. For example, the biochemical test view may contain 10 features, the complication marker view may contain 8 features, the imaging view may contain 5 features, and the medical history and scale view may contain 12 features.
[0181] Example 3: This example explains the hyperparameter α in step S24. α is a parameter that balances coverage and specificity, and is set to 0.5 by default. If a more compact interval (high specificity) is desired, α can be increased appropriately, for example, by setting it to 1.0; if more samples are covered (high coverage), α can be decreased, for example, by setting it to 0.2. In practical applications, the optimal α can be selected by evaluating subtype consistency on the validation set or by physician feedback.
[0182] To verify the regulatory effect of α, this embodiment uses a clinical dataset of 1268 patients with type 2 diabetes to test the coverage and specificity of subtype feature intervals under different α values. The results are shown in Table 1 below.
[0183] Table 1
[0184]
[0185] Example 4: This example explains the weight parameters ω and γ in step S42. ω controls the weight of overlap and distance, with a default value of 0.6, meaning more emphasis is placed on the overlap of feature intervals; γ controls the distance decay rate, with a default value of 1.0. Users can adjust these parameters according to data characteristics. For example, when feature noise is high, ω can be appropriately reduced to increase the weight of the distance term; if it is desired to emphasize the similarity of prototype centers more, γ can be reduced to make the distance penalty more gradual.
[0186] This embodiment uses a clinical dataset to test the mean absolute error (MAE) of missing value imputation under different parameter combinations. The results are shown in Table 2 below.
[0187] Table 2
[0188]
[0189] Example 5: This example extends the calculation of prototype-prototype edge weights in step S31. Besides initializing the weights using an exponential decay function, a learnable attention mechanism can be employed. This involves calculating the correlation score between two prototype nodes using a shared attention network, which serves as the edge weight. This attention network can take two prototype features as input and output a scalar weight. During training, it is jointly optimized with other losses, enabling the graph structure to adaptively adjust the cross-view prototype association strength.
[0190] This embodiment compares the model performance of the two weight calculation methods, and the results are shown in Table 3 below.
[0191] Table 3
[0192]
[0193] Example 6: This example provides a type 2 diabetes subtype clustering system for implementing the above method. The system includes:
[0194] Data acquisition module: Used to acquire multimodal clinical data of patients with type 2 diabetes.
[0195] Preprocessing module: Performs data cleaning, missing data labeling, and feature encoding in step S10.
[0196] Prototype building module: Performs fuzzy clustering and interval optimization in step S20 to generate granular prototypes for each view.
[0197] Graph Network Module: Performs heterogeneous graph construction and RGCN update in step S30.
[0198] Imputation module: Performs missing value imputation based on granularity matching in step S40.
[0199] Clustering output module: Executes step S50 to generate the final clustering and semantic description.
[0200] This invention provides a method for information completion and subtype analysis of multimodal incomplete clinical data on type 2 diabetes. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.
Claims
1. A method for information completion and subtype analysis of multimodal incomplete clinical data on type 2 diabetes, characterized in that, Includes the following steps: Step S10, Data Preprocessing, View Missing Labeling and Feature Encoding: Input an incomplete multimodal clinical dataset of type 2 diabetes patients, complete data cleaning, view missing labeling, view-specific feature encoding and cross-view latent space alignment, and obtain feature representations of all samples in a unified low-dimensional latent space; Step S20: Construction of view-specific subtype granular prototypes: Independently run fuzzy C-means clustering on the complete samples of each view to extract subtype membership, and construct super-rectangular granular prototypes of each subtype on each view based on weighted median and interval optimization. Step S30, Construction and Evolution of Cross-View Subtype Prototype Heterogeneous Graph: Construct a heterogeneous graph containing sample nodes and prototype nodes, design a two-layer relational graph convolutional network to update node features, and fuse semantic associations between views; Step S40, Granularity-based missing view imputation: For patients with missing views, the granularity matching degree between the complete features of the patients with missing views in other views and each subtype is calculated, the optimal subtype is selected and the missing value is filled with the midpoint of the optimal subtype prototype interval. Step S50, Subtype Classification and Semantic Output: K-means clustering is performed using the imputed complete dataset, with the evolved prototype centers as the initial seeds, to output patient subtype labels and interpretable granular clinical descriptions.
2. The method according to claim 1, characterized in that, Step S10 includes the following steps: Step S11: Input an incomplete multimodal clinical dataset of type 2 diabetes patients. ,in Total number of clinical views; For the first The feature matrix of each view, where Represents the space of real numbers. The total number of patient samples. For the first Clinical feature dimensions of each view; if patient sample In the If a view has missing features, then record ; Step S12: Generate the missing view index matrix , Indicates patient sample In the The clinical features of each view are complete. Indicates feature loss; Step S13, Generation and The corresponding missing mask matrix is used to perform min-max normalization on the complete clinical features of each view, mapping all feature values to the [0,1] interval; at the same time, the classification features are one-hot encoded, and outliers are truncated using the quartile method to complete the clinical data cleaning. Step S14, Design and Training of View-Specific Autoencoders: For each clinical view Design an independent view-specific autoencoder, which consists of two parts: an encoder and a decoder. The encoder formula is: （1）， in, No. The first clinical view Clinical feature vectors of each sample; For the first Each clinical view has its own dedicated encoder. For the first The first clinical view The latent feature vectors obtained by encoding each sample using an encoder; For dimension The real number space; The decoder formula is: （2）， in, For the first Each clinical view has its own dedicated decoder; For the first The first clinical view The reconstructed feature vector obtained by the decoder from each sample; The autoencoder is trained with the goal of minimizing the view reconstruction loss. （3）， in, For the first The reconstruction loss function for each clinical view; For the first Total number of samples for each clinical view; using The optimizer is trained iteratively, and after convergence, the trained view-specific encoder and decoder are obtained.
3. The method according to claim 2, characterized in that, Step S20 includes the following steps: Step S21: For each clinical view Extract the aligned latent feature set of the complete sample of the view. ; in, For the first The complete set of latent features after alignment of each clinical view; For the first The first clinical view Aligned latent feature vectors corresponding to each patient sample; As an indicator of sample integrity, Indicates the first The patient sample in the first Each clinical view contains complete samples without missing data; Step S22: Complete sample set for each view Independently run the fuzzy C-means clustering algorithm, with a preset number of subtypes. Iterative optimization yields the first... Membership matrix of each view and the membership degree of each patient sample to each subtype. ,in For patient sample indexing, For subtype index, For clinical view indexes, the membership degree satisfies the constraint. and ; in, For the first A complete set of clinical features corresponding to each clinical view; K is the preset number of disease subtypes; For the first Membership matrix obtained by fuzzy C-means clustering of clinical views; For the first In the first clinical view, the... The patient sample corresponds to the first Membership degree of each disease subtype; Step S23: Calculate the weights for each subtype using membership degree as the weight. In view Each clinical feature dimension weighted median ; Step S24: Based on the principle of reasonable granularity, for each subtype In view Each clinical feature dimension Search optimization of the optimal interval The optimization objective is to maximize the joint metrics of coverage and specificity, and the optimization function is... for: （4）， in, For the first The first clinical view, the first The disease subtype, the Coverage corresponding to clinical features; For the first The first clinical view, the first The disease subtype, the Subtype specificity corresponding to clinical characteristics; The calculation formula is: （5）， in This is an indicator function; it takes the value 1 if the condition within the parentheses is true, and 0 otherwise. For patient samples In view No. Values of clinical features; For the first The first clinical view, the first The disease subtype, the The left endpoint of the optimal interval corresponding to the clinical characteristics; For the first The first clinical view, the first The disease subtype, the The right endpoint of the optimal interval corresponding to the clinical features; For specificity, it is defined as an exponential decay function of the interval length, and the calculation formula is: （6）， in is a hyperparameter balancing coverage and specificity; exp is the natural exponential function; Step S25: Place each subtype c in the view The Cartesian product of the optimal intervals for all clinical feature dimensions yields subtype c in the view. Below 3D hyperrectangular granular prototype : （7）， in, For the first The first clinical view Granulated prototypes corresponding to each disease subtype; For the first The total dimensions of clinical features in each clinical view; For Cartesian product operations; Simultaneously, the interval anchor point The vector formed serves as the central feature of the granulation prototype.
4. The method according to claim 3, characterized in that, Step S30 includes the following steps: Step S31: Construct a cross-view subtype prototype heterogeneous graph : in, For constructing a cross-view subtype prototype heterogeneous graph; It is the set of nodes in a heterogeneous graph; It is a set of edges of a heterogeneous graph, including two types: sample-prototype edges and prototype-prototype edges; This is the set of edge weights for a heterogeneous graph; Sample-prototype edge: connects sample nodes and view prototype node Weight ; For patient sample nodes With the The first view Edge weights between subtype prototype nodes; For the first In the first clinical view The patient sample corresponds to the first Membership degree of each disease subtype; Prototype-prototype edge: Connects the corresponding prototype nodes of different views, with weights initialized to... ; in, For the first The first view The prototype node of the subtype and the first The first view Edge weights between subtype prototype nodes; This refers to the Gaussian kernel bandwidth parameter; The square of the Euclidean distance between the two prototype center eigenvectors; Step S32: Design a graph neural network (GNN) structure that includes a two-layer relational graph convolutional network (RGCN) as a prototype update, where RGCN is a relational graph convolutional network; the graph neural network is used to propagate node information and update the feature representation of prototype nodes on heterogeneous graphs. First-level message passing: For views prototype node Aggregate prototype nodes Information about the connected sample nodes: （8）， in, For the first The first view Feature vectors of each subtype prototype node; AGG is the aggregation function; Represents the prototype node A set of connected sample nodes; For the first In the first clinical view The patient sample corresponds to the first Membership degree of each disease subtype; For the first Feature vectors of each patient sample node; This means that the information obtained from the aggregation on the right is assigned to the feature of the prototype node on the left, thus completing the information update; For sample nodes Aggregate sample nodes Information about the connected prototype nodes: （9）， in To be related to sample nodes The set of all connected subtype prototype nodes; For the first The first view Feature vectors of each subtype prototype node; Second-level message passing: For views prototype node Aggregate information from the prototype nodes of other views: （10）， in, For view With View Attention weights or fixed similarity weights between them; For the first The first view Feature vectors of each subtype prototype node; This indicates that the aggregation scope is excluding the current view. In addition, all other views have the same subtype prototype nodes; the meanings of the remaining symbols are consistent with the aforementioned formulas; Step S33, Node Update: Update node features using residual connections, calculated as follows: （11）， in, For the updated number The first view Feature vectors of each subtype prototype node; Update the node function; The original feature vector of the prototype node before the update; the aggregated information is the neighbor node information obtained by aggregation in formulas (8) and (10); Step S34, Graph Neural Network (GNN) Training and Optimization: With the goal of minimizing the overall loss function, the Adam optimizer is used to iteratively optimize the learnable weight matrix of the GNN. During the training process, the sample node features, sample-prototype edge weights, and original membership matrix remain fixed throughout. Only the learnable weights and prototype node features of the GNN are updated. After iterative convergence, the final optimized global granular prototype is obtained. Overall loss function The calculation formula is: （12）， in, The balancing hyperparameter for cross-view clustering consistency loss; The balancing hyperparameter for L2 weighted regularization loss; Cross-view clustering consistency loss The calculation formula is: （13）， in, The total number of pairwise combinations of views. For patients In view The subtype membership vector, For patients In view Subtype membership vector; L2 weighted regularization loss The calculation formula is: （14）， in, For all learnable weight matrices in the graph encoder, Let f be the Frobenius norm of the matrix.
5. The method according to claim 4, characterized in that, Step S40 includes the following steps: Step S41: Iterate through all patient samples. If the patient is in the clinical view... Feature loss, i.e. Extract the patient's complete clinical features across all other views. ; Step S42: Calculate patient samples With each subtype Granularity matching degree on view m : （15）， in, For patients to have complete characteristics falling into the subtype The feature percentage within the corresponding view prototype range is specifically: statistics. Each feature dimension Does the value fall within the subtype? In view On the prototype interval Within, the percentage is the number of features that meet the conditions divided by The total feature dimension; For patients Except for views Complete feature splicing vectors from all views outside the scope; subtype The prototype center concatenation vector of the corresponding view. This is the overlap weighting coefficient. This is a distance adjustment parameter; Step S43: Select the subtype with the highest granularity matching degree as the patient sample. Subtype of belonging on missing view m ; Step S44, based on the subtype In view Granulated prototypes on Based on this, the midpoint of each dimension interval is taken as the fill value for the missing feature; Step S45: Repeat steps S41 to S44 to impute all missing views for all patient samples, thus obtaining a complete multimodal clinical dataset of type 2 diabetes patients. .
6. The method according to claim 5, characterized in that, Step S43: Obtain the subtype using the following formula. : （16）， in, The disease subtype number to which the patient sample to be classified is ultimately assigned; To find the independent variable that maximizes the objective function The operation; For patient samples to be classified Corresponding to the A matching score function for each disease subtype.
7. The method according to claim 6, characterized in that, In step S44, for the d-th dimension feature of view m, the fill value is: （17）， in and Subtypes The lower and upper bounds of the interval on the d-th dimension feature of view m.
8. The method according to claim 7, characterized in that, Step S50 includes the following steps: Step S51: Fill in the complete multimodal clinical dataset with the completed data. Input the K-means clustering algorithm, using the updated prototype node features obtained in step S30. As the initial seed for clustering, it is used to perform the final patient subtype classification; Step S52: Output the final subtype label for each patient sample, and simultaneously generate a granular semantic description of the clinical features of each subtype, in the following format: Subtype c: Glycated hemoglobin ∈ [value missing] in biochemical test view %, fasting blood glucose ∈ mmol / L; urinary microalbumin / creatinine ratio in the complication marker view ∈ mg / g; corresponding to a high, medium, or low risk of developing diabetic nephropathy; in , They represent subtypes respectively. In view No. Lower and upper bounds of the granularity prototype interval on the dimensional feature; Step S53: Based on the granular prototype features of each subtype, complete the risk stratification of diabetic microvascular and macrovascular complications, and use the subtype label, granular feature description and risk stratification results as the final output.
9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.
10. A storage medium, characterized in that, It stores a computer program or instructions that, when executed on a computer, perform the steps of the method as described in any one of claims 1 to 8.