Diagnostic methods for type 2 diabetes complications based on multi-source heterogeneity feature transfer

Abstract

This invention belongs to the field of diagnostic technology for diabetic complications, and relates to a diagnostic method for type 2 diabetes complications based on multi-source heterogeneous feature transfer. The method includes: extracting features from CGM sequence information, clinical indicators, and past medication information, and dividing them into training and testing sets; using a Gaussian mixture clustering algorithm on the training set to obtain K Gaussian models and the membership class of each sample based on a preset number of clusters K; pre-training samples from each class to obtain K diagnostic models; for each diagnostic model, weighting samples from other classes based on the membership matrix and then transferring the weighted samples from other classes for training to obtain K final models; and using the final models to diagnose complications of type 2 diabetes. Its beneficial effect is that it integrates multi-source information in information processing, solving the problem of fragmentation between sub-models of different subgroups, enabling the models to be fully trained, and improving the accuracy of the models in diagnosing complications.

Description

Technical Field

[0001] This invention relates to the field of diagnostic technology for diabetic complications, and in particular to a diagnostic method for type 2 diabetic complications based on multi-source heterogeneous feature migration. Background Technology

[0002] While hyperglycemia in type 2 diabetes doesn't directly threaten life like hypoglycemia, it's a common chronic metabolic disease with over a hundred complications. With increasing health awareness and advancements in medical technology, disease diagnosis and prevention have become crucial. Machine learning and deep learning, as powerful tools, are widely used in disease diagnosis. However, traditional machine learning typically expects data to follow a normal distribution during classification, as most analytical tools and algorithms assume this. Non-normal or multimodal distributions require further processing. Due to the differences in clinical phenotypes and progression among patients with type 2 diabetes, patient data is heterogeneous and exhibits a multimodal distribution, such as... Figure 1 As shown. Therefore, taking into account the heterogeneity of patients when building diagnostic models will be more helpful for accurate disease diagnosis.

[0003] Current diagnostic models often focus on single or multi-task modeling for one or more diseases, treating all available data as a whole to build a single model. During model training, all features and samples are considered. However, when heterogeneity or specific subgroups exist in the dataset, the overall model may fail to fully capture the characteristics of these subgroups. A common approach to address this issue is to build independent sub-models for each subtype of type 2 diabetes patients. Each sub-model can be trained for a specific patient group, better adapting to and capturing patterns within the group and handling heterogeneous data. However, this approach fails to utilize the interaction information between subgroups, resulting in fragmented models for each subgroup. This can lead to a situation where the overall sample size is sufficient, but the sample size for a specific subtype is insufficient, preventing the model from being adequately trained. Summary of the Invention

[0004] Technical problems to be solved

[0005] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a diagnostic method for complications of type 2 diabetes based on multi-source heterogeneous feature transfer, which solves the technical problems that the existing diagnostic models cannot utilize the interaction information between various subgroups, the sub-models of various subgroups are isolated from each other, and the overall sample is sufficient but the sample of the subtype is insufficient, so the model cannot be fully trained.

[0006] Technical solution

[0007] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0008] This invention provides a diagnostic method for type 2 diabetes complications based on multi-source heterogeneous feature transfer, comprising:

[0009] Acquire multi-source heterogeneous data from multiple samples, including: CGM sequence information, clinical indicators, past medication information, and complication information of the samples;

[0010] Based on the multi-source heterogeneous feature extraction and fusion network, features of CGM sequence information, clinical indicators and past medication information are extracted, and the extracted features are fused to obtain feature set H. The feature set is then divided into training set and test set.

[0011] Gaussian mixture clustering algorithm is used on the training set to obtain the Gaussian model of K classes and the membership matrix of each sample based on the preset number of clusters K;

[0012] The class with the highest probability in the membership matrix of each sample is taken as the membership class of that sample;

[0013] For each class of samples, pre-training is performed to obtain K diagnostic models;

[0014] For each diagnostic model, non-class samples are weighted according to the membership matrix, and then the non-class weighted samples are transferred for training to obtain K final models;

[0015] Determine the membership matrix of the test set to classify the test set, input the divided test set into the corresponding class of the final model according to the class, obtain the diagnostic results and evaluate the model performance;

[0016] The final model was used to diagnose complications of type 2 diabetes.

[0017] Optionally, extracting CGM sequence information features includes:

[0018] Define the CGM sequence dataset C={ There are N sample data points in total. Let i be the CGM sequence of the i-th patient sample, i∈[1,N], for dimensional vector, Determined based on the duration the patient wears the CGM device;

[0019] Construct an encoder with a one-dimensional CNN and LSTM combined structure, and the decoder structure is symmetrical to the encoder;

[0020] Deep features of dataset C are extracted based on the encoder. Q∈[1,N], where, , Represents the coding layer network. This is the weight parameter vector of the coding layer network. This represents the bias vector of the coding layer network;

[0021] Deep features The reconstructed data is obtained through the decoder. { }, ,in, Indicates the decoding layer network. This is the weight parameter vector of the decoding layer network. This represents the bias vector of the decoding layer network;

[0022] Based on error function Combining deep features Obtain the reconstructed data with the minimum error compared to dataset C. , ;

[0023] deep features The diagnostic output is obtained through a feedforward neural network. { Diagnostic error Added to network training, among which, , , This is the weight parameter vector of the feedforward neural network. This represents the bias vector of the feedforward neural network. The function will output the diagnostic results with probabilities. The loss function uses cross-entropy loss, and the final loss function, Loss, is a linear combination of the reconstruction error and the diagnostic error. .

[0024] Optionally, extracting clinical indicator features includes:

[0025] Define clinical indicator feature dataset }, corresponding to the features of N samples, the i-th sample Dimensions Depending on the number of clinical indicators included, the dataset will be... Weight vectors are generated through a feedforward neural network. For each sample, and Having the same dimensions , and Multiplication yields weighted clinical indicator characteristics ;

[0026] in, , and This is the weight matrix for the first and second layers of the feedforward neural network. and The bias vectors of the first and second layers of the feedforward neural network, the number of neurons in the FNN network, and the sample dimension. same, , Let be the weight of the j-th feature in the i-th sample. .

[0027] Optionally, extracting features from past medication information includes:

[0028] The past medication information is converted into a treatment description vector using an encoding method. Each position in the treatment description vector represents the use of a drug. The positions are, in order, metformin, sulfonylureas, thiazolidinediones, meglitinides, DPP-4 inhibitors, AGI and insulin. The position is assigned a value of 1 for the drug used and 0 for the drug not used.

[0029] Define a dataset of past medication information },in The medication status of the i-th patient;

[0030] Dataset based on a one-dimensional CNN network Different convolutional kernel sizes were used to extract features from combinations of different treatment methods. Then, a feedforward neural network was used to extract deep features. The convolutional kernels were used to extract features from combinations of different treatment methods. , and The size of the feature extraction is determined by the size of the feature extraction process, and finally, the feature extraction is performed through two layers of FNN network to obtain the deep feature. .

[0031] Optionally, the extracted features are fused to obtain a feature set H, including:

[0032] The extracted CGM sequence information, clinical indicators, past medication information, and GSP features were concatenated. The GSP features were obtained from the CGM time series based on blood glucose levels and fluctuations. A joint training approach was used to optimize the multi-source heterogeneous feature extraction and fusion network. The concatenated features were then trained using a Free Network Neural Network (FNN), and a Dropout layer was used to enhance the model's generalization ability. After training iterations, the feature set was obtained. , The loss function Loss is a linear combination of reconstruction error and diagnostic error.

[0033] Optionally, a Gaussian mixture clustering algorithm is used on the training set to obtain a Gaussian model of K classes and a membership matrix of each sample based on a preset number of clusters K, including:

[0034] Initialize the initial parameters of K Gaussian models based on the number of clusters K. and ;

[0035] The membership degree of each sample to each Gaussian model is determined according to the following formula. : , Let X be the i-th sample data in the training set, and let X = { }, i∈[1,N];

[0036] Based on membership degree The unknown parameters of each Gaussian model are determined using the following formula. and :

[0037] , , ;

[0038] Determine whether the Gaussian model converges using the following formula:

[0039] ,in, and Let be the membership degrees for the (t+1)th iteration and the tth iteration, respectively. For a very small integer, when the error between two iterations is sufficiently smaller than 1 / 2. When the model has converged, it is considered to have converged.

[0040] If the Gaussian model is determined to be convergent, obtain Gaussian models of K classes and output the membership matrix. ,in, .

[0041] Optionally, pre-training is performed on samples from each class to obtain K diagnostic models, including:

[0042] The pre-trained model for each class is ,in Representative subclass model function, For subclass All sample features, This represents all the unknown parameters of the pre-trained model, including the weight matrix and bias vector in the neural network. The model uses FNN layers and Dropout layers, and finally uses... The function outputs the predicted probability, and the loss function of the pre-trained model is... ,in, Features of a certain type of sample The true diagnostic label, For the model The predicted output, the predicted output Compared with actual output The greater the similarity, the better the model fit. During training, the difference between the predicted output and the true label is calculated using the cross-entropy loss function. Then, the backpropagation algorithm is used to adjust the model parameters based on this loss to optimize model performance. At the same time, a regularization term is added to prevent overfitting.

[0043] Optionally, for each diagnostic model, non-class samples are weighted according to the membership matrix, and then the weighted non-class samples are transferred for training to obtain K final models, including:

[0044] Each non-class sample is weighted according to the membership matrix to obtain weighted non-class samples;

[0045] The diagnostic model for this class is trained using weighted non-class samples, and the parameters of the diagnostic model for this class are fine-tuned to obtain the final model for this class.

[0046] Beneficial effects

[0047] The beneficial effects of this invention are as follows: This invention provides a diagnostic method for type 2 diabetes complications based on multi-source heterogeneous feature transfer. By extracting the fusion features of multi-source heterogeneous data and applying clustering algorithms to discover the multi-peak distribution of patients, and by using transfer learning to supplement common domain knowledge, a diagnostic model for diabetes complications is established. This method can take into account data sources with different data structures and can also integrate multi-source information in information processing. It solves the problem of the separation between sub-models of various subgroups, enabling the model to be fully trained and improving the accuracy of the model in diagnosing complications. Attached Figure Description

[0048] Figure 1 This is a schematic diagram illustrating the multi-peaked distribution of patient heterogeneity data.

[0049] Figure 2 A schematic diagram of a complication diagnosis framework based on multi-source heterogeneous feature transfer provided in an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the diagnostic method for type 2 diabetes complications based on multi-source heterogeneous feature transfer provided in an embodiment of the present invention.

[0051] Figure 4 This is a schematic diagram of the multi-source heterogeneous feature extraction and fusion network structure provided in an embodiment of the present invention;

[0052] Figure 5 This is a schematic diagram of the subclass model network structure provided in an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of the subclass feature transfer network structure provided in an embodiment of the present invention. Detailed Implementation

[0054] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] With the use and improvement of various monitoring devices and medical databases, the sources of information for recording patient status have become more extensive and standardized. Different data sources provide different aspects of the target object, i.e., the sample. Therefore, multi-source heterogeneous data can compensate for the incompleteness of data from a single data source, making the target information more comprehensive. In order to find useful information in massive multi-source heterogeneous data, it is necessary to structure and integrate the heterogeneous data, and data fusion technology has become increasingly important. Multi-source heterogeneous data of patients with type 2 diabetes usually includes CGM sequences (continuous glucose monitoring), clinical characteristics and past medication regimens, covering three data structures: time series, continuous univariate, and categorical variables. These multi-source heterogeneous data all contain the information entropy required for patient diagnosis. In the establishment of diagnostic models, it is necessary to design reasonable data fusion methods to improve the comprehensiveness and usability of the data.

[0056] To address the above problems, this invention designs a multi-source heterogeneous feature fusion method to extract latent information from patients. Subsequently, it proposes a soft clustering-based transfer learning method to identify the multimodal distribution in the latent features of patients and utilizes transfer learning to supplement common domain knowledge. The proposed complication diagnosis framework is as follows: Figure 2 As shown.

[0057] Specifically, feature-level information fusion was used to integrate the CGM sequences, clinical characteristics, and past medication regimens of patients with type 2 diabetes. Feature extraction schemes suitable for their respective data structure characteristics were designed. Soft clustering was used to obtain patient groups for the extracted fusion features. Each patient was assigned a membership vector to represent the credibility of belonging to each group. A decision layer network was built for each group for model pre-training. Subsequently, the diagnostic model was improved by transfer learning of patient features based on the membership vectors.

[0058] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0059] Firstly, referring to Figure 3 This embodiment provides a diagnostic method for type 2 diabetes complications based on multi-source heterogeneous feature transfer, including:

[0060] Each data source differs in structure and subject matter, making it difficult to directly apply clustering algorithms to heterogeneous multi-source data, resulting in low similarity and unreliable clustering results. An effective approach is to extract features from heterogeneous datasets, mapping data from different sources to a single feature space for a unified data description, thus improving the accuracy of subsequent clustering operations. After feature extraction, the fundamental information or essential characteristics of the dataset remain unchanged, and the dataset's representation is more suitable for subsequent tasks. Clustering based on the fused features can better capture the latent patterns within the dataset.

[0061] Therefore, in this embodiment, the fusion features of multi-source heterogeneous data are first extracted. The fused features are then used as a new dataset and input into the downstream task. Then, the fused features are used as a new input to apply a clustering algorithm to discover the multi-peak distribution of patients. The feature extraction process is guided by the disease diagnosis target. All patients are divided into training and test sets in a 4:1 ratio. The training set is used to update the model structure, and the test set is used to verify the model performance, thereby updating the hyperparameters. After the model is trained, all heterogeneous data are input into the trained model to obtain the fused features.

[0062] S1, acquire multi-source heterogeneous data of multiple samples, the multi-source heterogeneous data including: CGM sequence information of the samples, clinical indicators, past medication information and complication information.

[0063] The clinical indicators include basic information and biochemical test information. Basic information includes the patient's gender, age, disease duration, and weight; biochemical test information includes glycated hemoglobin (HbA1c), glycated albumin (GA), C-peptide, and triglyceride levels; and complication information includes diabetic retinopathy. Each part has a corresponding data warehouse. Biochemical tests and basic information are continuous features, CGM is a one-dimensional time series, and past medication information is a multi-category variable.

[0064] The basic information database contains individual patient information such as gender, age, disease duration, and weight, which doctors record during simple examinations and inquiries. Much of this information has been clinically proven to be closely related to the occurrence of complications. Older patients and those with longer disease durations have a greater risk of microvascular complications. Body mass index (BMI), determined by height and weight, is also a risk factor for complications. Therefore, incorporating this basic information is essential for the diagnosis of complications.

[0065] A retrospective CGM system was used for subcutaneous interstitial glucose monitoring. All participants had sensors implanted on the first day of hospitalization and removed them 3 to 7 days later, recording 288 blood glucose values ​​daily. Because the data from the first day was adaptive, CGM data from the second day onwards were included in the study. The CGM system generates CSV files during monitoring. These files were preprocessed according to the patients, removing those who did not meet the requirements. Since the system generates a blood glucose value every five minutes, patients with missing values ​​exceeding one hour were removed, while those with missing values ​​less than one hour were filled using quadratic curve fitting. CGM is an important indicator of blood glucose fluctuations, and persistent hyperglycemia is a pathological cause of complications. Exploring the potential information from CGM data can aid in the diagnosis of complications.

[0066] Biochemical test information primarily consists of blood tests, including glycated hemoglobin (HbA1c), glycated albumin (GA), C-peptide, and triglycerides. These are important biomarkers for the progression of diabetes. Outliers in these indicators need to be handled, as human error can cause some patients' values ​​to exceed the expected range, requiring exclusion from the panel. Glycated hemoglobin, as a traditional indicator of blood glucose control, remains the gold standard for long-term blood glucose management. Glycated albumin is a measure of short-term blood glucose control, and C-peptide is an important indicator of insulin secretion. These indicators reflect the unique individual circumstances of each patient, providing more information for the diagnosis of complications.

[0067] A patient's past medication regimen significantly impacts their internal glycemic regulation. This typically includes insulin injections and other medications. Insulin injections are suitable for patients with partial or complete loss of pancreatic function, who cannot produce enough insulin for their cells to use. Additionally, insulin-resistant patients may require high doses of insulin to meet their sparse cellular needs. Medication use involves numerous drugs and combinations, including metformin, sulfonylureas, thiazolidinediones, meglitinides, DPP-4 inhibitors, and AGIs. This treatment information is crucial for correcting complications because there are correlations between medications, biochemical indicators, and CGM (cholesterol-glucose levels), and therefore needs to be considered in establishing complication diagnostic models.

[0068] The complication information primarily includes diabetic retinopathy (DR). While the incidence of life-threatening macrovascular complications is decreasing year by year, the age of patients is increasing, leading to a rise in the prevalence of microvascular complications, which also cause long-term suffering for patients. Retinopathy is a leading cause of blindness. Diabetic retinopathy is diagnosed by an ophthalmologist who remains blinded to the subject's characteristics, based on fundus photographs taken using a 45°, 6.3-megapixel digital non-fundus camera (CR6-45NM; Canon, Lake Success, NY).

[0069] Multi-source heterogeneous data fusion is the process of analyzing and integrating multiple information source data for the same target object. The heterogeneity of data sources leads to differences in the feature vectors representing the data, thus creating a gap between the association, crossover, and integration of multi-source heterogeneous data. Only by fully utilizing multi-source data and giving full play to the complementarity of multi-source heterogeneous data can we conduct more thorough data analysis and obtain more valuable analytical results. This requires eliminating the gap between heterogeneous data and integrating various data sources for related analysis. Generally speaking, data-oriented information fusion can be divided into three fusion levels: (1) data-level fusion, (2) feature-level fusion, and (3) decision-level fusion. Among them, feature-level fusion can take into account the advantages of the other two, considering data sources with different data structures and integrating multi-source information in information processing. Multi-source heterogeneous data feature-level fusion.

[0070] S2 extracts features from CGM sequence information, clinical indicators, and past medication information based on a multi-source heterogeneous feature extraction and fusion network. After fusing the extracted features to obtain feature set H, the feature set is divided into training set and test set.

[0071] Feature extraction and fusion network structure for multi-source heterogeneous information, such as Figure 4 As shown, the patient's basic information and biochemical test indicators are both continuous variables with a consistent data structure, therefore they are combined into patient clinical indicators during feature extraction. This framework extracts features from the three parts of information based on their data structure and differences, mainly divided into input data, a feature extraction module, and a feature fusion module. For CGM information mining, knowledge-guided GSP features are supplemented and integrated with other features to jointly represent patient information. The following sections describe the specific implementation of each feature extraction module in three parts.

[0072] Optionally, extracting CGM sequence information features includes:

[0073] Define the CGM sequence dataset C={ There are N sample data points in total. Let i be the CGM sequence of the i-th patient sample, i∈[1,N], for dimensional vector, Determined based on the duration the patient wears the CGM device;

[0074] Construct an encoder with a one-dimensional CNN and LSTM combined structure, and the decoder structure is symmetrical to the encoder;

[0075] Deep features of dataset C are extracted based on the encoder. Q∈[1,N], where, , Represents the coding layer network. This is the weight parameter vector of the coding layer network. This represents the bias vector of the coding layer network;

[0076] Deep features The reconstructed data is obtained through the decoder. { }, ,in, Indicates the decoding layer network. This is the weight parameter vector of the decoding layer network. This represents the bias vector of the decoding layer network;

[0077] Based on error function Combining deep features Obtain the reconstructed data with the minimum error compared to dataset C. , ;

[0078] deep features The diagnostic output is obtained through a feedforward neural network. { Diagnostic error Added to network training, among which, , , This is the weight parameter vector of the feedforward neural network. This represents the bias vector of the feedforward neural network. The function will output the diagnostic results with probabilities. The loss function uses cross-entropy loss, and the final loss function, Loss, is a linear combination of the reconstruction error and the diagnostic error. .

[0079] The design of the autoencoder module extracts features. The autoencoder can be used for representation learning and consists of two parts: an encoder and a decoder. Usually, the two parts adopt a symmetrical structure to complete the extraction of deep features. It is an effective data dimensionality reduction method. The encoding and decoding network can be designed according to the actual application.

[0080] The encoder and decoder network structures primarily consist of Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). CNNs extract regional features from data based on the size and number of convolutional kernels and are widely used in image processing. Based on the kernel dimension, they can be categorized into one-dimensional and multi-dimensional CNNs. One-dimensional CNNs are often used for processing data streams, while two-dimensional CNNs can traverse image features. Recurrent Neural Networks (RNNs) are capable of handling time-dependent features of the input and are frequently used for feature mining of time-series data. For long-term dependency problems, Long Short-Term Memory (LSTM) networks are employed, adding the ability to mine features in the time dimension. The encoder and decoder layers use a combination of one-dimensional CNNs and LSTMs to extract the temporal and spatial information of the CGM sequence.

[0081] First, the encoding layer uses a one-dimensional convolutional network to extract spatial features from the CGM data, with a kernel size of 3. Features are extracted for 15-minute blood glucose status. Then, an LSTM network is used to extract the temporal dependency features of the spatial status. At this point, each sample data is a two-dimensional tensor. Features are extracted through two layers of CNN network, and important features are selected and retained after pooling operation. Then, the ReLU activation function is used to ensure the non-linear feature mining capability of the network. Finally, global average pooling is used to obtain deep features. The decoding layer structure also uses a combination of CNN and LSTM.

[0082] For dataset C, deep features can be obtained based on the encoder module. After obtaining the deep features, the decoding layer network... Obtain reconstructed data The model objective is to minimize the input data. With reconstructing data The better the dimensionality reduction, the better the reconstructed data obtained from the deep features h has with smaller errors compared to the input data. In the backpropagation of the algorithm, the error function of the reconstructed part is... .

[0083] Due to the unsupervised feature extraction property of autoencoders, in order to incorporate task guidance information, the deep features h are processed through feedforward neural networks (FNNs) to obtain diagnostic output. Diagnostic error Add it to network training.

[0084] The deep features of the CGM sequence were obtained using the method described above. The spatial features of CGM are extracted by a CNN network, and the temporal dependency features of the sequence are extracted by an LSTM network. The selection of convolution kernels takes into account a fixed time interval of 15 minutes. An encoder-decoder structure is used to reduce the dimensionality of the data, and the effectiveness of the features is increased by considering reconstruction error and diagnostic error.

[0085] Optionally, extracting clinical indicator features includes:

[0086] Define clinical indicator feature dataset }, corresponding to the features of N samples, the i-th sample Dimensions Depending on the number of clinical indicators included, the dataset will be... Weight vectors are generated through a feedforward neural network. For each sample, and Having the same dimensions , and Multiplication yields weighted clinical indicator characteristics ;

[0087] in, , and This is the weight matrix for the first and second layers of the feedforward neural network. and The bias vectors of the first and second layers of the feedforward neural network, the number of neurons in the FNN network, and the sample dimension. This ensures that each sample has its own importance metric weight. , Let be the weight of the j-th feature in the i-th sample. .

[0088] Due to the black-box nature of machine learning, doctors cannot clearly understand the source of diagnostic results during use, and they cannot fully comprehend the processing logic of machine learning. This has led to many models encountering obstacles in clinical practice. Clinical features, as indicators with clear significance, are the basis for clinical judgments of patients' physical condition. In the deep information extraction of clinical features, it is even more important to clarify their contribution. Therefore, attention weights are used to compare the importance of each clinical feature when designing the extraction module.

[0089] use The function guarantees that the output is within the range of 0 to 1, and then standardization is performed to ensure that any sample has... This facilitates comparison and ranking of importance, ultimately leading to deeper clinical features. .

[0090] Optionally, extracting features from past medication information includes:

[0091] For the rule text description, an encoding method is used to transform past medication information into a treatment description vector. Each position in the treatment description vector represents the medication usage, with each position sequentially corresponding to the use of metformin, sulfonylureas, thiazolidinediones, meglitinides, DPP-4 inhibitors, AGIs, and insulin. Positions where medication is used are assigned a value of 1, and positions where medication is not used are assigned a value of 0. For example... This indicates that the patient used metformin and AGI oral medication, and also received insulin injections;

[0092] Define a dataset of past medication information },in This section describes the medication use of the i-th patient, primarily consisting of descriptions of the administration of various hypoglycemic drugs and insulin injections.

[0093] Dataset based on a one-dimensional CNN network Different convolutional kernel sizes were used to extract features from combinations of different treatment methods. Variations in the receptive field were employed to extract as many complete features as possible. Then, a feedforward neural network was used to extract deep features. The convolutional kernels were used in different ways. , and The size of the feature extraction is determined by the size of the feature extraction process, and finally, the feature extraction is performed through two layers of FNN network to obtain the deep feature. .

[0094] Optionally, the extracted features are fused to obtain a feature set H, including:

[0095] The feature extraction for each part is not done in isolation, but rather by concatenating the features obtained from each layer. This involves concatenating the extracted CGM sequence information, clinical indicators, past medication information, and GSP features. The GSP features are derived from the CGM time series based on blood glucose levels and fluctuations. A joint training approach is used to optimize the multi-source heterogeneous feature extraction and fusion network, enabling it to collectively leverage the strengths of each feature component. The concatenated features are then forward-trained using an FNN, with a Dropout layer used to enhance the model's generalization ability. After training iterations, a feature set is obtained. , The loss function Loss is a linear combination of reconstruction error and diagnostic error.

[0096] GSP features are a type of feature mined from CGM time series data. The first dimension: based on blood glucose levels, blood glucose can be divided into three categories: hypoglycemia, normal blood glucose, and hyperglycemia, represented by the symbols B, I, and A. The second dimension: based on blood glucose fluctuation information, blood glucose decreases, increases, and stable patterns can be represented by the symbols D, I, and S. Combining these two dimensions, blood glucose patterns can be divided into nine groups: BS: stable hypoglycemia fluctuation pattern; BI: increasing hypoglycemia fluctuation pattern; BD: decreasing hypoglycemia fluctuation pattern; IS: stable normal blood glucose fluctuation pattern; II: increasing normal blood glucose fluctuation pattern; ID: decreasing normal blood glucose fluctuation pattern; AS: stable hyperglycemia fluctuation pattern; AI: increasing hyperglycemia fluctuation pattern; AD: decreasing hyperglycemia fluctuation pattern.

[0097] In data-driven methods, clustering is an effective way to discover different subsets in the full sample set. In Chapter 3, we analyzed the patient subtypes from the perspective of CGM. In disease diagnosis, when using machine learning models, we need to consider their general assumptions, namely independent and identically distributed and sufficient samples. This leads to the question of how to make full use of common domain and subdomain information when making a diagnosis. Here, common domain information refers to the common information that patients with type 2 diabetes share. For example, high blood glucose concentration is a common physical manifestation for diabetic patients. Subdomain refers to the characteristics that the subclasses have but are not reflected in other categories. However, clustering still cannot completely divide the entire sample set, and there is overlap between them.

[0098] For a single sample, it contains more or less information from other categories. Therefore, the weights obtained by soft clustering can represent the amount of information contained. For each class of samples, a separate pre-trained model is prepared. During the pre-training stage, the private domain information of each subclass can be fully learned to obtain the pre-trained model parameters for each subclass. Then, based on the membership matrix obtained by soft clustering, the features of other samples besides the samples in this class are transferred to the pre-trained model of this class according to the membership weights. This completes the final fine-tuning of the pre-trained model, and finally, a subclass model with complete information and sufficient training is obtained.

[0099] In the overall training process, the extracted multi-source fusion feature dataset is divided into training and test sets in an 8:2 ratio. The training set is used for optimizing model parameters and evaluating model structure, while the test set is used to evaluate and select the final training model. First, a Gaussian Mixture Model (GMM) clustering algorithm is applied to the training set. The GMM algorithm fits a Gaussian model to K classes and a membership matrix for each sample based on the input K value. Then, a diagnostic model is pre-trained for each class of samples, with each sample belonging to the Gaussian model with the highest membership value. A total of K diagnostic models can be pre-trained. Finally, samples from other classes are used to train the final model for each class, with these samples weighted according to their membership values ​​in that class. After training, the probability of the test set in each trained Gaussian model is calculated, the test set is assigned to each class, and the assigned samples are input into the corresponding final model according to their class. The output results are then obtained, and the model performance is evaluated.

[0100] S3 uses the Gaussian mixture clustering algorithm on the training set to obtain the Gaussian model of K classes and the membership matrix of each sample based on the preset number of clusters K.

[0101] Optionally, a Gaussian mixture clustering algorithm is used on the training set to obtain a Gaussian model of K classes and a membership matrix of each sample based on a preset number of clusters K, including:

[0102] Initialize the initial parameters of K Gaussian models based on the number of clusters K. and ;

[0103] The membership degree of each sample to each Gaussian model is determined according to the following formula. : , Let X be the i-th sample data in the training set, and let X = { }, i∈[1,N];

[0104] Based on membership degree The unknown parameters of each Gaussian model are determined using the following formula. and :

[0105] , , ;

[0106] Determine whether the Gaussian model converges using the following formula:

[0107] ,in, and Let be the membership degrees for the (t+1)th iteration and the tth iteration, respectively. For a very small integer, when the error between two iterations is sufficiently smaller than 1 / 2. When the model has converged, it is considered to have converged.

[0108] If the Gaussian model is determined to be convergent, obtain Gaussian models of K classes and output the membership matrix. ,in, .

[0109] The Gaussian distribution is one of the most commonly used distribution models in mathematical statistics, also known as the normal distribution. It is one of the most common and frequently used probability distributions in nature. In reality, the distribution of height and weight, manufacturing errors in components, and weather phenomena such as temperature and precipitation all approximate Gaussian distributions. The probability density function of a Gaussian model of a D-dimensional random variable is as follows: Where X is a D-dimensional data sample, The mean is D-dimensional. for The covariance matrix.

[0110] Gaussian Mixture Model (GMM) treats a set of data samples as a combination of multiple Gaussian models. As a mixture of probability models, GMM can, in principle, approximate any continuous probability distribution model when the number of combinations is sufficient. GMM clustering uses algorithmic optimization to fit each Gaussian distribution, obtaining the mean and variance of each distribution. Each sample has a probability value in that distribution, which serves as its classification into that category, thus completing the clustering.

[0111] For dataset X={ Given N samples, assume their distribution is a linear combination of K Gaussian models with unknown parameters. For the i-th sample... Its probability density function is as follows: ,in Let the probability density function of the k-th Gaussian model be denoted as . Let be the mean of k Gaussian models. Let be the covariance matrix of k Gaussian models. Let represent the probability that the k-th Gaussian model selects this sample, and satisfy the condition. , A higher value indicates that the sample conforms more closely to the Gaussian distribution.

[0112] Clustering a dataset using a Gaussian mixture model involves fitting K corresponding Gaussian models and the membership degree of each sample to a class based on the existing observable sample data. A common method for solving unknown parameters is to maximize the likelihood function. To make the solution easier, the problem is often transformed into maximizing the log-likelihood function. The log-likelihood function of a Gaussian mixture model is as follows: Maximizing the log-likelihood function to complete parameter estimation typically uses the Expectation Maximization (EM) algorithm. This is an iterative optimization algorithm where each iteration improves the estimated parameters. Each iteration consists of two steps: calculating the expectation and maximizing the expectation. The overall algorithm steps for GMM clustering are as follows:

[0113] (1) Given a mixture number k, randomly initialize the initial parameters of k Gaussian models. and .

[0114] (2) Estimate the expected membership degree of each sample to each model. The calculation formula is as follows: .

[0115] (3) Estimate the unknown parameters of each Gaussian model and The calculation formula is as follows: , , .

[0116] (4) Repeat steps two and three until convergence.

[0117] The membership matrix is ​​generated when the log-likelihood function is maximized until convergence. Each row of values ​​corresponds to the probability of the Gaussian model to which the sample belongs, and the maximum value is the class of the sample. This completes the clustering process based on GMM. Compared with the K-means clustering method, this method can find multiple Gaussian distributions in the dataset and can give the probability of belonging to each class rather than just one class. Therefore, it belongs to the soft clustering algorithm, and the multi-class probability attribute of each sample contains richer information.

[0118] Specifically, regarding the feature set of multi-source heterogeneous data samples in this embodiment... Each feature is a combination of the aforementioned four deep features; the goal here is to discover the feature set. The subset of features, the assumptions of Gaussian Mixture Clustering (GMM) are similar to the heterogeneity of diseases, and feature sets are discovered through GMM. Beyond the subset, we need to obtain the inter-group differences and intra-group similarities for each sample in the feature space. This is represented by a membership matrix, which is the probability that each sample's features belong to each class. The clustering endpoint is determined based on convergence criteria, and no specific number of clustering iterations is set. The specific steps of Gaussian mixture clustering are as follows:

[0119] (1) Input sample feature set Set the number of clusters K and initialize the parameters of K Gaussian models.

[0120] (2) Estimate the expected membership degree of each sample to each model. .

[0121] (3) Based on membership degree calculate and .

[0122] (4) Based on membership degree and Calculate the covariance matrix .

[0123] (5) Determine whether the convergence criteria are met. If not, return to step (2).

[0124] (6) Output the membership matrix And then it ended.

[0125] This gives us the probability that each sample belongs to each class, and we have .

[0126] S4. The class with the highest probability in the membership matrix of each sample is taken as the membership class of that sample.

[0127] Choose the class with the highest probability for each sample, ensuring that samples within each class are unique. Sample features Category For the k-th category, there are The number of samples is ,satisfy Each sample is assigned to its maximum probability class, while other class information is hidden in the membership matrix. middle.

[0128] S5 involves pre-training samples for each class to obtain K diagnostic models.

[0129] Optionally, pre-training is performed on samples from each class to obtain K diagnostic models, including:

[0130] The pre-trained model for each class is ,in Representative subclass model function, For subclass All sample features, This represents all the unknown parameters of the pre-trained model, including the weight matrix and bias vector in the neural network. The model uses FNN layers and Dropout layers, and finally uses... The function outputs the predicted probability, and the loss function of the pre-trained model is... ,in, Features of a certain type of sample The true diagnostic label, For the model The predicted output, the predicted output Compared with actual output The greater the similarity, the better the model fit. During training, the difference between the predicted output and the true label is calculated using the cross-entropy loss function. Then, the backpropagation algorithm is used to adjust the model parameters based on this loss to optimize model performance. At the same time, a regularization term is added to prevent overfitting.

[0131] For each class of samples, the model is pre-trained. The pre-trained model structure for each class is as follows: Figure 5 As shown, the class with the highest probability is selected. Pre-training within a class simplifies model complexity and reduces computation, essentially performing preliminary model tuning to bring the model closer to its optimal state. It also reduces the likelihood of overfitting or local optima during subsequent training, improving model stability and robustness. With better initialized parameters, the model can converge to the global optimum more quickly during later training phases. Each subclass diagnostic model uses the same network structure, primarily a feedforward neural network, with Dropout layers used after some network layers to prevent overfitting.

[0132] S6. For each diagnostic model, weight the non-class samples according to the membership matrix and then transfer the weighted non-class samples for training to obtain K final models.

[0133] Optionally, for each diagnostic model, non-class samples are weighted according to the membership matrix, and then the weighted non-class samples are transferred for training to obtain K final models, including:

[0134] Each non-class sample is weighted according to the membership matrix to obtain weighted non-class samples;

[0135] The diagnostic model for this class is trained using weighted non-class samples, and the parameters of the diagnostic model for this class are fine-tuned to obtain the final model for this class.

[0136] Transfer learning has the following definitions: source domain (Non-class samples) and the target domain (Samples of this class), source domain feature space Target domain feature library space Source task With the target task Classification tasks based on transfer learning are divided into many research directions based on factors such as the presence or absence of sample labels, domain matching degree, and task similarity, but their purpose remains the same: to learn from the source domain. To the source task In the process of learning, knowledge is applied to the target task. Based on the source of the learned knowledge, transfer learning is generally divided into three categories: instance-based transfer learning, feature-based transfer learning, and shared parameter-based transfer learning.

[0137] Transfer learning, as a solution when traditional classification assumptions are not met, has provided insights into many real-world problems and has been widely applied in fields such as image recognition, bearing diagnosis, energy and power prediction, and emotion recognition. The idea behind transfer learning can provide effective information-sharing strategies for patients with multi-modal distributions; however, finding suitable transfer subjects from each latent class and designing transfer strategies are the key challenges, while avoiding negative transfer phenomena.

[0138] Using other samples for transfer training on an already pre-trained model allows for better guidance of model training by leveraging existing information, thereby improving model performance and generalization ability. Only subclasses were used Sample feature set In the rest Each subclass still contains elements belonging to the subclass. Information, this similarity information is obtained using a membership matrix. Perform measurement, for any subclass Define the target domain as an example ,remove Sample features in the class In addition, other sample features are source domain features. Number of samples in the target domain Number of source domain samples Target domain input sample for Source domain input sample , and They share the same feature and label spaces, but their marginal probability density distributions differ, and the complete information for each category is supplemented using the source domain. This is because the target domain model has already been pre-trained. ,exist Based on the weighted instances in the source domain as input, the instance information is transferred to complete the model. Fine-tuning is performed to obtain a complete category diagnostic model. A schematic diagram of the soft subclass transfer model is shown below. Figure 6 As shown.

[0139] Source domain samples Fine-tuning model based on values ​​in the membership matrix loss function Predict output for source domain Compared with the actual output of the source domain The cross-entropy error means that important samples contribute more to the loss function, and the importance of each sample is... From the membership matrix The sample-weighted loss function is obtained from this. The calculation formula is as follows: The maximum mean discrepancy (MMD) is used to measure the difference between the source domain distribution and the target domain distribution, and is denoted as the distribution loss. MMD maps data to a regenerated Hilbert space. (Reproducing Kernel Hilbert Space, RKHS) performs distance measurement and maps kernel functions. Gaussian functions and distributed loss are often used. The calculation is as follows: Total loss function To predict losses With distribution loss The weighted combination of prediction loss and distribution loss allows the model to learn the mapping from the original input to the diagnostic labels, improving diagnostic accuracy. Distribution loss is used to reduce negative transfer caused by distributional differences during the transfer process. The model is updated through the backpropagation algorithm. Parameters, total loss function The calculation formula is as follows: For the j-th soft subclass Update the pre-trained model using samples from other subclasses as source domain samples. Each class utilizes the similarity distribution information of other samples to complete the final fine-tuning of the model, generating a total of K final models. Each model is a class-specific model, which not only contains the distribution information of its own class, but also utilizes the knowledge transferred from the source domain samples as an expansion of information.

[0140] S7. Determine the membership matrix of the test set to classify the test set. Input the divided test set into the corresponding class of the final model according to the class to obtain the diagnostic results and evaluate the model performance.

[0141] S8 uses the final model to diagnose complications of type 2 diabetes.

[0142] The following specific embodiment further illustrates a diagnostic method for type 2 diabetes complications based on multi-source heterogeneous feature transfer:

[0143] A total of 1,636 patients with type 2 diabetes were enrolled, of whom 1,308 participated in training and 328 were used to test the final model's effectiveness. Type 2 diabetes was diagnosed according to the American Diabetes Association (ADA) diagnostic criteria. Patients had to be at least 18 years old and have complete and valid clinical records (CGM) and clinical documentation. All patients underwent islet autoantibody testing, including glutamate decarboxylase antibody and islet tyrosine phosphatase 2 antibody, to exclude patients with type 1 diabetes and other specific types of diabetes according to the ADA diagnostic criteria. Furthermore, patients with diabetic ketoacidosis, a history of malignancy or mental illness, or who had not had a stable treatment regimen in the past three months were excluded.

[0144] The CGM information for each patient was extracted from data from the early morning of the second day to the third day, with 576 blood glucose values ​​per patient. Basic information included sex, age, disease duration, and body mass index (BMI). Biochemical information included glycated hemoglobin (HbA1c), glycated albumin (GA), insulin resistance (HOMA-IR), insulin sensitivity (HOMA-IS), β-cell function (HOMA-B), fasting blood glucose (0.0 pG), fasting insulin (0.0 INS), fasting C-peptide (0.0 CP), 2-hour C-peptide CP (2h), 120-minute C-peptide (0.1 CP), triglycerides (TG), total cholesterol (TC), high-density lipoprotein (HDL), and low-density lipoprotein (LDL). Medication information included whether treatment was initiated, metformin, sulfonylureas, thiazolidinediones, meglitinides, DPP-4 inhibitors, AGI, and whether insulin injection was administered.

[0145] The model's evaluation metrics primarily utilize accuracy, precision, recall, and F1-score to diagnose patients with diabetic renal syndrome (DR). The experiment consists of three parts. First, features are extracted from each part of the multi-source heterogeneous data, termed the Multi-source Heterogeneous Features (MSHF) part. Then, a soft subclassing strategy is applied to these features to obtain samples from different clusters. A model is pre-trained for each sample group (Subclass pre-trained networks, SPTN). Finally, knowledge transfer is performed using features from other classes, termed the Soft Subclass Transfer Network (SSTN).

[0146] In the experimental part of the SPTN network, SPTN performs soft subclassing of samples based on MSHF and selects samples from the class with the highest probability for model pre-training. The sample class K is specified as 4 classes. First, the performance of subclassing and not subclassing is compared. The results are shown in Table 1.

[0147] Table 1

[0148]

[0149] The feature set H obtained from MSHF was divided into four subclasses and the model was pre-trained separately for each subclass. The results are shown in Table 1. SPTN-Class represents the sample feature set of a certain class. Using the proposed model Pre-training showed that, except for class B, which performed worse than the MSHF method, other classes benefited from individual subclass training. This indicates that splitting the multimodal data in the sample set effectively learns the private domain information of each class. Pre-training for each class can better capture its unique patterns and reduce interference from other classes. Although class D had slightly lower precision than the MSHF model, it had significantly higher recall, demonstrating strong recall capabilities and minimizing misclassification of positive samples. Class C showed the best discriminative performance in diagnostic segmentation, and SPTN still outperformed the MSHF model in terms of average performance across all samples. The number of samples and positive cases for each class is shown in Table 2. Class B had fewer samples than other classes, which may be due to overfitting caused by fewer samples. It performed well in the training set but declined in the test set, indicating a lack of generalization ability.

[0150] Table 2

[0151]

[0152] Based on the pre-training of subclasses, further consideration is given to the cross information between subclasses. The membership matrix of soft subclasses is used to complete the fine-tuning of the pre-training. SSTN is based on the multi-source heterogeneous features extracted by MSHF and the pre-trained model of SPTN subclasses. The information transfer between subclasses is carried out. In the experiment, the performance of SSTN on each class is compared with that of SPTN. In the training, the training set and the test set are still divided. The experimental results are shown in Table 3.

[0153] Table 3

[0154]

[0155] As shown in Table 3, the SSTN-based model significantly outperforms the SPTN model among the four classes. Adding knowledge transfer to the SPTN model compensates for the lack of shared domain information in individual classes and also greatly increases the sample size. This is most evident in class B, where the performance improvement is the most significant, with an accuracy increase of 0.081. In the SPTN model, the limited sample size in class B leads to reduced generalization ability, while this problem is significantly improved in the SSTN model. Transferring sample knowledge from source domain samples other than class B greatly enhances model performance, and improvements are also seen in other classes. Class C still exhibits the highest discriminative ability among the four classes, achieving an accuracy of 0.945, an increase of 0.052 compared to the SPTN model, with an F1-score increase of 0.052. Class D shows the smallest increase in accuracy, which may be partly due to the abundance of samples, but the lack of shared domain information still improves its discriminative performance within the SSTN model.

[0156] All the above experiments were based on the complication of diabetic retinopathy. In summary, the performance of various methods was compared with that of the proposed method in the experimental section. At the same time, ablation experiments were used to compare the performance of each module, which verified the effectiveness of feature extraction and fusion network and soft subclass transfer.

[0157] In a second aspect, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed, implements the diagnostic method for type 2 diabetes complications based on multi-source heterogeneous feature transfer as described in any of the first aspects above.

[0158] Thirdly, embodiments of the present invention provide a storage device, including a storage medium and a processor, wherein the storage medium stores a computer program, and when the program is executed by the processor, it implements the diagnostic method for type 2 diabetes complications based on multi-source heterogeneous feature migration as described in any of the first aspects above.

[0159] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

[0161] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A diagnostic method for complications of type 2 diabetes based on multi-source heterogeneous feature transfer, characterized in that, include: Acquire multi-source heterogeneous data from multiple samples, including: CGM sequence information, clinical indicators, past medication information, and complication information of the samples; Based on the multi-source heterogeneous feature extraction and fusion network, features of CGM sequence information, clinical indicators and past medication information are extracted, and the extracted features are fused to obtain feature set H. The feature set is then divided into training set and test set. Gaussian mixture clustering algorithm is used on the training set to obtain the Gaussian model of K classes and the membership matrix of each sample based on the preset number of clusters K; The class with the highest probability in the membership matrix of each sample is taken as the membership class of that sample; For each class of samples, pre-training is performed to obtain K diagnostic models; For each diagnostic model, non-class samples are weighted according to the membership matrix, and then the non-class weighted samples are transferred for training to obtain K final models; Determine the membership matrix of the test set to classify the test set, input the divided test set into the corresponding class of the final model according to the class, obtain the diagnostic results and evaluate the model performance; The final model was used to diagnose complications of type 2 diabetes.

2. The diagnostic method for type 2 diabetes complications based on multi-source heterogeneous feature transfer according to claim 1, characterized in that, The features extracted from CGM sequence information include: Define the CGM sequence dataset C={ There are N sample data points in total. Let i be the CGM sequence of the i-th patient sample, i∈[1,N], for dimensional vector, Determined based on the duration the patient wears the CGM device; Construct an encoder with a one-dimensional CNN and LSTM combined structure, and the decoder structure is symmetrical to the encoder; Deep features of dataset C are extracted based on the encoder. Q∈[1,N], where, , Represents the coding layer network. This is the weight parameter vector of the coding layer network. This represents the bias vector of the coding layer network; Deep features The reconstructed data is obtained through the decoder. { }, ,in, Indicates the decoding layer network. This is the weight parameter vector of the decoding layer network. This represents the bias vector of the decoding layer network; Based on error function Combining deep features Obtain the reconstructed data with the minimum error compared to dataset C. , ; deep features The diagnostic output is obtained through a feedforward neural network. { Diagnostic error Added to network training, among which, , , This is the weight parameter vector of the feedforward neural network. This represents the bias vector of the feedforward neural network. The function will output the diagnostic results with probabilities. The loss function uses cross-entropy loss, and the final loss function, Loss, is a linear combination of the reconstruction error and the diagnostic error. .

3. The method for diagnosing type 2 diabetes complications based on multi-source heterogeneous feature transfer according to claim 2, characterized in that, Extracted clinical indicator characteristics include: Define clinical indicator feature dataset }, corresponding to the features of N samples, the i-th sample Dimensions Depending on the number of clinical indicators included, the dataset will be... Weight vectors are generated through a feedforward neural network. For each sample, and Having the same dimensions , and Multiplication yields weighted clinical indicator characteristics ; in, , and This is the weight matrix for the first and second layers of the feedforward neural network. and The bias vectors of the first and second layers of the feedforward neural network, the number of neurons in the FNN network, and the sample dimension. same, , Let be the weight of the j-th feature in the i-th sample. .

4. The method for diagnosing type 2 diabetes complications based on multi-source heterogeneous feature transfer according to claim 3, characterized in that, Features extracted from past medication information include: The past medication information is converted into a treatment description vector using an encoding method. Each position in the treatment description vector represents the use of a drug. The positions are, in order, metformin, sulfonylureas, thiazolidinediones, meglitinides, DPP-4 inhibitors, AGI and insulin. The position is assigned a value of 1 for the drug used and 0 for the drug not used. Define a dataset of past medication information },in The medication status of the i-th patient; Dataset based on a one-dimensional CNN network Different convolutional kernel sizes were used to extract features from combinations of different treatment methods. Then, a feedforward neural network was used to extract deep features. The convolutional kernels were used to extract features from combinations of different treatment methods. , and The size of the feature extraction is determined by the size of the feature extraction process, and finally, the feature extraction is performed through two layers of FNN network to obtain the deep feature. .

5. The method for diagnosing type 2 diabetes complications based on multi-source heterogeneous feature transfer according to claim 4, characterized in that, The extracted features are fused to obtain a feature set H, which includes: The extracted CGM sequence information, clinical indicators, past medication information, and GSP features were concatenated. The GSP features were obtained from the CGM time series based on blood glucose levels and fluctuations. A joint training approach was used to optimize the multi-source heterogeneous feature extraction and fusion network. The concatenated features were then trained using a Free Network Neural Network (FNN), and a Dropout layer was used to enhance the model's generalization ability. After training iterations, the feature set was obtained. , The loss function Loss is a linear combination of reconstruction error and diagnostic error.

6. The method for diagnosing type 2 diabetes complications based on multi-source heterogeneous feature transfer according to claim 5, characterized in that, Gaussian mixture clustering algorithm is used on the training set to obtain the Gaussian model of K classes and the membership matrix of each sample based on the preset number of clusters K, including: Initialize the initial parameters of K Gaussian models based on the number of clusters K. and ; The membership degree of each sample to each Gaussian model is determined according to the following formula. : , Let X be the i-th sample data in the training set, and let X = { }, i∈[1,N]; Based on membership degree The unknown parameters of each Gaussian model are determined using the following formula. and : 、 、 ； Determine whether the Gaussian model converges using the following formula: ,in, and Let be the membership degrees for the (t+1)th iteration and the tth iteration, respectively. For a very small integer, when the error between two iterations is sufficiently smaller than 1 / 2. When the model has converged, it is considered to have converged. If the Gaussian model is determined to be convergent, obtain Gaussian models of K classes and output the membership matrix. ,in, .

7. The method for diagnosing type 2 diabetes complications based on multi-source heterogeneous feature transfer according to claim 6, characterized in that, For each class of samples, K diagnostic models are pre-trained to obtain: The pre-trained model for each class is ,in Represents the model function for each subclass. For subclass All sample features, This represents all the unknown parameters of the pre-trained model, including the weight matrix and bias vector in the neural network. The model uses FNN layers and Dropout layers, and finally uses... The function outputs the predicted probability, and the loss function of the pre-trained model is... ,in, Features of a certain type of sample The true diagnostic label, For the model The predicted output, the predicted output Compared with actual output The greater the similarity, the better the model fit. During training, the difference between the predicted output and the true label is calculated using the cross-entropy loss function. Then, the backpropagation algorithm is used to adjust the model parameters based on this loss to optimize model performance. At the same time, a regularization term is added to prevent overfitting.

8. The method for diagnosing type 2 diabetes complications based on multi-source heterogeneous feature transfer according to claim 7, characterized in that, For each diagnostic model, non-class samples are weighted according to the membership matrix, and then the training is performed using the weighted non-class samples to obtain K final models, including: Each non-class sample is weighted according to the membership matrix to obtain weighted non-class samples; The diagnostic model for this class is trained using weighted non-class samples, and the parameters of the diagnostic model for this class are fine-tuned to obtain the final model for this class.

Images