Disease classification prediction method and device based on prototype transfer learning, and readable medium

The disease classification prediction method based on prototype transfer learning solves the problem of model overfitting in small-sample metabolomics by combining prototype loss and feature distribution difference loss with feature encoder and classifier, and achieves high accuracy and stability in disease classification.

CN122046018APending Publication Date: 2026-05-15XIAMEN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610165548.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-02-05
Publication Date
2026-05-15

AI Technical Summary

Technical Problem

In metabolomics, high sample acquisition costs or scarce sample sources lead to model overfitting under small sample conditions. Traditional methods are difficult to reliably extract metabolic features with discriminative significance, affecting the reliability and generalization ability of the analysis results.

Method used

A disease classification prediction method based on prototype transfer learning is adopted. By constructing a disease classification prediction model, a feature encoder and classifier are used, and prototype loss and feature distribution difference loss are combined for training to achieve cross-domain feature alignment and classification prediction.

Benefits of technology

It significantly improves the accuracy and stability of disease classification under small sample conditions, enhances the sensitivity and specificity of disease classification, avoids complex data amplification or feature selection processes, and maintains the generalization performance of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122046018A_ABST
    Figure CN122046018A_ABST
Patent Text Reader

Abstract

The invention discloses a disease classification prediction method and device based on prototype transfer learning, and a readable medium, and relates to the field of data processing, and the method comprises the steps: obtaining H-NMR metabolic spectrum data of serum samples of a person suffering from a first disease and a person suffering from a second disease, and carrying out the preprocessing, corresponding metabolome features are obtained, and a source domain data set and a target domain data set are constructed; constructing a disease classification prediction model, wherein the disease classification prediction model comprises a feature encoder and a classifier; performing prototype transfer learning-based training on the disease classification prediction model by using the source domain data set and the target domain data set to obtain a trained disease classification prediction model; and acquiring H-NMR metabolic spectrum data of a serum sample of a to-be-predicted person suffering from the second disease, performing preprocessing to obtain corresponding metabolome characteristics, and inputting the metabolome characteristics into the trained disease classification prediction model to obtain a disease classification result of the to-be-predicted person. According to the method, the classification accuracy and the model generalization performance can be effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically to a disease classification and prediction method, apparatus, and readable medium based on prototype transfer learning. Background Technology

[0002] Metabolomics is a research method that systematically analyzes the composition and changes of small-molecule metabolites in biological systems to reflect the influence of physiological states or external disturbances. Nuclear magnetic resonance (NMR) technology, due to its advantages such as simple sample preparation, non-destructive testing, good quantification, and high reproducibility, is widely used in medical research, food quality evaluation, and nutritional analysis. In metabolomics research, by analyzing NMR spectroscopy data, sample types can be identified and metabolic pathways analyzed, providing a basis for differential analysis of complex systems.

[0003] However, in the practical application of metabolomics, problems such as high sample acquisition costs or scarce sample sources often arise, resulting in a limited number of samples available for analysis. Furthermore, nuclear magnetic resonance (NMR) metabolomics data typically exhibit high feature dimensionality and complex correlations between variables, making them prone to model overfitting under small sample conditions. This makes it difficult for traditional statistical analysis and modeling methods to reliably extract discriminative metabolic features, thus affecting the reliability and generalization ability of the analytical results.

[0004] To address the challenges of small-sample metabolomics analysis, existing techniques typically employ improvements such as sample amplification, feature selection, or the construction of complex model structures. However, these methods still have limitations in practical applications. For instance, data amplification methods may introduce noise or disrupt the original metabolic distribution characteristics; while algorithm-based feature selection methods can reduce feature dimensionality to some extent, they often rely on evaluation criteria, are sensitive to parameter settings and data distribution, and exhibit poor stability of selection results under small sample conditions. Therefore, a technical solution is needed that can effectively improve the stability and accuracy of small-sample metabolomics data analysis without relying on large-scale samples and complex feature engineering. Summary of the Invention

[0005] The purpose of this application is to propose a disease classification prediction method, device, and readable medium based on prototype transfer learning to address the aforementioned technical problems.

[0006] In a first aspect, the present invention provides a disease classification prediction method based on prototype transfer learning, comprising the following steps:

[0007] ¹H-NMR metabolomic profile data of serum samples from individuals suffering from the first disease and the second disease were obtained and preprocessed to obtain the corresponding metabolomic features. A source domain dataset was constructed based on the true category labels of individuals suffering from the first disease and their corresponding metabolomic features, and a target domain dataset was constructed based on the corresponding metabolomic features of individuals suffering from the second disease.

[0008] A disease classification prediction model is constructed, which includes a feature encoder and a classifier. The disease classification prediction model is trained using prototype transfer learning based on source domain datasets and target domain datasets to obtain the trained disease classification prediction model.

[0009] ¹H-NMR metabolomic data of serum samples from individuals with a second disease to be predicted are obtained and preprocessed to obtain corresponding metabolomic features. The metabolomic features are then input into a trained disease classification prediction model. The model first passes through a feature encoder to obtain the corresponding embedded feature representation. The embedded feature representation is then passed through a classifier to obtain the disease classification result of the individuals to be predicted.

[0010] Preferably, the total loss function used in the training process of the disease classification prediction model is constructed by weighting the feature distribution difference loss and the prototype loss, as shown in the following formula:

[0011] ;

[0012] in, For loss weighting coefficients, Indicates prototype loss. This represents the loss due to the difference in characteristic distribution. This represents the total loss function.

[0013] As a preferred embodiment, the construction process of the feature distribution difference loss is as follows:

[0014] The metabolomics features from the source and target domain datasets are input into the feature encoder to obtain the embedded feature representations of the source domain samples and the target domain samples, as shown in the following equation:

[0015] ;

[0016] ;

[0017] in, This represents the metabolomic features of the i-th sample in the source domain dataset. This represents the metabolomic feature of the j-th sample in the target domain dataset. This represents the function corresponding to the feature encoder. This represents the embedding feature representation of the i-th sample in the source domain dataset. This represents the embedding feature representation of the j-th sample in the target domain dataset;

[0018] The feature distribution difference loss between the source domain dataset and the target domain dataset is calculated using the maximum mean based on the embedded feature representations of the source domain samples and the target domain samples, as shown in the following formula:

[0019] ;

[0020] in, This represents the loss due to the difference in characteristic distribution. and These represent the number of samples in the source domain dataset and the target domain dataset, respectively. This represents the L2 norm.

[0021] As a preferred approach, the prototype loss is constructed as follows:

[0022] The following formula is used to calculate the number of people suffering from the first disease in the first month. The prototype vector corresponding to each real category label:

[0023] ;

[0024] in, This indicates that the person suffering from the first disease is in the first The prototype vector corresponding to each real category label;

[0025] in, Indicates that the data in the source domain belongs to the first... A sample set of real category labels, Indicates that the data in the source domain belongs to the first... The number of samples in a sample set of real category labels;

[0026] The prototype loss is calculated using the following formula:

[0027] ;

[0028] in, This represents the metabolomics feature of the l-th sample in each batch of training data in the source domain dataset. This represents the total number of samples in each batch of training data in the source domain dataset. This indicates that the person suffering from the first disease is in the first The prototype vector corresponding to each real category label Represents Euclidean distance. This represents the total number of true category labels.

[0029] Preferably, the feature encoder includes a first branch, a second branch, a flattening layer, and a fully connected layer. The first branch includes a first convolutional module, a first max pooling layer, a second convolutional module, a second max pooling layer, a third max pooling layer, and a dropout layer connected in sequence. The second branch includes a first convolutional layer, a fourth max pooling layer, a fifth max pooling layer, and a sixth max pooling layer connected in sequence. The first convolutional module is connected to a second convolutional layer, a first batch normalization layer, and a first ReLU activation function layer in sequence. The second convolutional module is connected to a third convolutional layer, a second batch normalization layer, and a second ReLU activation function layer in sequence. The third convolutional module is connected to a fourth convolutional layer, a third batch normalization layer, and a third ReLU activation function layer in sequence.

[0030] The calculation process of the classifier is as follows:

[0031] ;

[0032] in, This indicates the metabolomic characteristics of serum samples from individuals to be predicted to have a second disease. The metabolomic features corresponding to the serum samples of individuals to be predicted to have a second disease are input into the feature encoder to obtain the embedded feature representation. The disease classification results obtained after inputting into the classifier. Representing embedded feature representation The corresponding disease classification result is the second disease. The conditional probabilities of each category;

[0033] The category with the highest conditional probability is selected as the disease classification result for the person to be predicted.

[0034] As a preferred option, both the first and second diseases are related to children's growth and development, and the clinical manifestations of the first and second diseases are associated with metabolomic characteristics; in contrast, the serum samples of people with the first disease are large samples, while the serum samples of people with the first disease are small samples.

[0035] Secondly, the present invention provides a disease classification prediction device based on prototype transfer learning, comprising:

[0036] The data processing module is configured to acquire ¹H-NMR metabolochromatographic data of serum samples from individuals suffering from the first disease and the second disease, respectively, and preprocess them to obtain the corresponding metabolomic features; construct a source domain dataset based on the true category labels of individuals suffering from the first disease and their corresponding metabolomic features, and construct a target domain dataset based on the metabolomic features of individuals suffering from the second disease.

[0037] The model building module is configured to build a disease classification prediction model, which includes a feature encoder and a classifier. The disease classification prediction model is trained using prototype transfer learning based on the source domain dataset and the target domain dataset to obtain the trained disease classification prediction model.

[0038] The classification module is configured to acquire and preprocess the ¹H-NMR metabolomic profile data of serum samples from individuals with a second disease to be predicted, obtain the corresponding metabolomic features, input the metabolomic features into the trained disease classification prediction model, first pass through the feature encoder to obtain the corresponding embedded feature representation, and then pass through the classifier to obtain the disease classification result of the individual to be predicted.

[0039] Thirdly, the present invention provides an electronic device including one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any implementation of the first aspect.

[0040] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any of the implementations of the first aspect.

[0041] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method as described in any of the implementations in the first aspect.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] (1) The disease classification prediction method based on prototype transfer learning mentioned in this invention constructs a feature encoder based on a convolutional neural network to extract features from metabolomic features and obtain corresponding embedded feature representations; based on the source domain dataset, prototype vectors of each category are constructed in the embedding space, and classification prediction is achieved according to the distance between the embedded feature representation of the sample and the prototype vectors of each category; by performing weighted joint optimization of prototype loss and feature distribution difference loss, classification discrimination and cross-domain feature alignment are achieved, effectively improving the classification accuracy of diseases related to the target domain.

[0044] (2) The disease classification prediction method based on prototype transfer learning mentioned in this invention can effectively classify metabolomics data of dwarfism by combining prototype learning, cross-domain feature alignment and deep convolutional neural networks. Even under small sample conditions, it significantly improves the stability and accuracy of disease classification, while maintaining high sensitivity and high specificity, providing a reliable technical foundation for intelligent diagnosis of diseases.

[0045] (3) The disease classification prediction method based on prototype transfer learning mentioned in this invention does not require a complex data amplification or feature screening process, and can effectively improve classification accuracy and model generalization performance under the conditions of small sample and high-dimensional nuclear magnetic resonance metabolomics data. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the disease classification prediction method based on prototype transfer learning, which is an embodiment of this application.

[0048] Figure 2 This is a schematic diagram illustrating the training and inference process of the disease classification prediction model in the prototype transfer learning-based disease classification prediction method of this application.

[0049] Figure 3 This is a schematic diagram illustrating the prototype learning principle in the disease classification prediction method based on prototype transfer learning, as described in an embodiment of this application.

[0050] Figure 4 ROC curve of the trained disease classification prediction model in the prototype transfer learning-based disease classification prediction method of this application on the pre-pubertal short stature dataset;

[0051] Figure 5 The confusion matrix of the trained disease classification prediction model in the prototype transfer learning-based disease classification prediction method of this application on the pre-pubertal short stature dataset is shown.

[0052] Figure 6 ROC curve of the trained disease classification prediction model in the prototype transfer learning-based disease classification prediction method of this application on the adolescent short stature dataset;

[0053] Figure 7 The confusion matrix of the trained disease classification prediction model in the prototype transfer learning-based disease classification prediction method of this application on the adolescent short stature dataset is shown.

[0054] Figure 8 This is a schematic diagram of a disease classification prediction device based on prototype transfer learning, which is an embodiment of this application.

[0055] Figure 9This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0057] Figure 1 This application illustrates an embodiment of a disease classification prediction method based on prototype transfer learning, comprising the following steps:

[0058] S1. Obtain ¹H-NMR metabolomic profile data of serum samples from individuals suffering from the first disease and the second disease, respectively, and preprocess them to obtain the corresponding metabolomic features. Construct a source domain dataset based on the true category labels of individuals suffering from the first disease and their corresponding metabolomic features, and construct a target domain dataset based on the corresponding metabolomic features of individuals suffering from the second disease.

[0059] In a specific embodiment, both the first disease and the second disease are diseases related to children's growth and development, and the clinical manifestations of the first disease and the second disease are associated with metabolomic characteristics; in contrast, the serum samples of people with the first disease are large samples, while the serum samples of people with the first disease are small samples.

[0060] Specifically, in one embodiment of this application, childhood malnutrition is used as the first disease, and pre-pubertal dwarfism or pubertal dwarfism is used as the second disease as an example. In other embodiments, other diseases can also be selected. Both the first and second diseases are related to childhood growth and development. The clinical manifestations and metabolomic characteristics of the diseases are related. The first and second diseases are similar but different diseases, and the difference between the types of diseases cannot be too large. The disease that is easier to collect samples for is used as the source domain, and the rarer disease is used as the target domain. The purpose is to use the data from the source domain to assist the target domain in classification.

[0061] In one embodiment of this application, the metabolomic features corresponding to serum samples from individuals suffering from childhood malnutrition are selected as the source domain dataset, and the metabolomic features corresponding to serum samples from individuals suffering from pre-pubertal short stature or pubertal short stature are selected as the target domain dataset. Childhood malnutrition is categorized into mild, moderate, and severe malnutrition; pre-pubertal short stature is categorized into PAC, PAG, and PAI; and pre-pubertal short stature is categorized into ADC, ADG, and ADI. Childhood malnutrition is a relatively common childhood disease with a wide range of sample sources, providing relatively sufficient labeled samples for training the disease classification prediction model. Short stature, on the other hand, is a less common childhood growth and development disorder with a limited number of clinical samples and high data collection costs. By utilizing the metabolic pattern information contained in the metabolic profile data of childhood malnutrition and combining cross-domain feature alignment and prototype transfer learning strategies, this information is effectively transferred to the metabolic profile data analysis task of short stature, thereby significantly improving the accuracy and stability of short stature disease classification under the condition of limited sample size.

[0062] In the process of constructing the source domain dataset and the target domain dataset, serum samples under different disease states are first obtained. 1 H-NMR metabolic spectroscopy data.

[0063] The malnutrition sample included children aged 4 to 84 months who visited the pediatric outpatient department of a hospital in Xiamen between July 2020 and May 2022. Inclusion criteria were children aged 0–7 years, meeting the WHO diagnostic criteria for childhood malnutrition, and without significant comorbidities; exclusion criteria included unresolved infectious diseases, digestive system diseases, cardiovascular diseases, respiratory system diseases, congenital immunodeficiency, congenital malformations, and inherited metabolic diseases.

[0064] Standardized measurement methods were used to calculate age-specific height / length Z-scores (HAZ), age-specific weight Z-scores (WAZ), and height-specific weight Z-scores (WHZ). The minimum of these three values ​​was used to determine the severity of malnutrition: -2 ≤ Z < -1 for mild, -3 ≤ Z < -2 for moderate, and Z < -3 for severe. The control group consisted of participants who fasted overnight before blood collection. Serum was collected according to clinical standards, allowed to stand at room temperature for 30 minutes, and centrifuged at 1000g for 10 minutes at 4°C. The serum supernatant was stored at -80°C for later use. The collected samples included 153 healthy controls, 199 cases of mild malnutrition, 67 cases of moderate malnutrition, and 9 cases of severe malnutrition. Due to the extremely small number of severe malnutrition cases, treating them as a separate class could easily lead to class imbalance, thus affecting the classification performance of the model training. Therefore, in subsequent analyses, moderate and severe malnutrition were combined into the same class. After merging, the three-class classification structure of the source domain children's malnutrition dataset and the target domain short stature dataset is consistent, which is beneficial for cross-domain feature alignment and knowledge transfer under the transfer learning framework.

[0065] The samples of children with short stature came from a hospital in Xiamen and a hospital in Bengbu. Before the study began, the purpose of the study was fully explained to the families of all the children involved, and informed consent forms were signed.

[0066] Inclusion criteria include:

[0067] (1) The height is more than 1.88 standard deviations below the average height of children of the same region, age and gender;

[0068] (2) The growth rate has been consistently below the normal level for no less than 6 months;

[0069] (3) Undergo growth hormone stimulation test;

[0070] (4) No abnormalities were found on pituitary MRI.

[0071] Exclusion criteria include:

[0072] (1) Previously or currently receiving growth hormone therapy;

[0073] (2) The presence of intracranial abnormalities, such as malignant tumors, hypothalamic hamartomas, or Arnold-Chiari malformation;

[0074] (3) Comorbid chronic liver and kidney diseases, bone diseases, congenital heart disease, thyroid dysfunction or chromosomal abnormalities;

[0075] (4) The MRI image has obvious artifacts or the sagittal plane cannot be clearly delineated.

[0076] The growth hormone stimulation test requires the participating children to fast after 8 pm the previous night. Arginine (500 mg / kg) is administered intravenously within 30 minutes of administration the following morning. If the initial GH peak value is <10 ng / mL, a second stimulation test is performed with oral levodopa (10 mg / kg). Blood samples are collected at 30, 60, 90, and 120 minutes after administration to measure serum GH levels. Children with GH peak values ​​<10 ng / mL in both stimulation tests are diagnosed with growth hormone deficiency (GHD), while those with GH peak values ​​≥10 ng / mL are diagnosed with idiopathic short stature (ISS).

[0077] Based on the Tanner stage of puberty, the study subjects were divided into pre-pubertal and pubertal groups. Ultimately, a total of 217 children with short stature were included, further divided into four groups according to etiology and developmental stage: 94 pre-pubertal GHD (PAG), 61 pre-pubertal ISS (PAI), 32 pubertal GHD (ADG), and 10 pubertal ISS (ADI). In addition, 134 healthy children were recruited as a control group, including 94 healthy pre-pubertal children (PAC) and 40 healthy pubertal children (ADC).

[0078] Serum sample pretreatment: Take 200 μL of serum and add 400 μL of phosphate buffer containing 0.9% physiological saline (60 mM, K2HPO4:NaH2PO4=4:1, 99.9% D2O, pH 7.4). Mix well and centrifuge at 13000 g for 10 minutes at 4°C. Take 550 μL of the supernatant and place it in a 5 mm NMR tube for analysis.

[0079] Metabolic spectrum acquisition and preprocessing: ¹H-NMR metabolic spectra were acquired using a CPMG pulse sequence on a Bruker AV-600 NMR spectrometer (ultra-low temperature probe, 298 K) (spectral width 12 kHz, sampling time 1.36 s, relaxation delay 4.0 s, number of sampling points 32 K, cumulative 128 times). The free induction decay (FID) signal in the ¹H-NMR metabolic spectrum was zero-charged to 64 K, multiplied by a 1.0 Hz exponential function, Fourier transformed, and phase and baseline corrected using MestReNova software. Calibration was performed using the methyl peak of lactate (CH3, δ 1.33). The water peak (δ 4.67–5.20) and the urea peak (δ 5.50–6.08) were removed. The region δ 0.60–8.60 was segmented, integrated at 0.002 ppm, and normalized. 1 After the above piecewise integration process, the H-NMR metabolomics data yields a dataset containing several metabolic features, which is the metabolomics feature. Each metabolic feature represents... 1 The peak signal corresponding to each chemical shift on H-NMR is shown, and the chemical shift corresponds to the metabolite.

[0080] The above process yields the metabolomic features of serum samples from children with malnutrition and those with short stature. The source domain dataset is constructed by combining the metabolomic features of serum samples from children with malnutrition with their corresponding ground truth class labels. The target domain dataset, which does not contain labels, is then used as the metabolomic features of serum samples from those with short stature.

[0081] S2. Construct a disease classification prediction model, which includes a feature encoder and a classifier. Train the disease classification prediction model using prototype transfer learning with source domain dataset and target domain dataset to obtain the trained disease classification prediction model.

[0082] For details, please refer to Figure 2 First, a disease classification prediction model is constructed, which consists of a feature encoder and a classifier. The feature encoder uses a deep convolutional neural network to extract features from high-dimensional metabolomics features. In the embedded feature representation space, a prototype vector for each disease category is constructed. During training, prototype loss and feature distribution difference loss are combined for joint optimization, thereby achieving high accuracy and stable classification of metabolomics data in the target domain under small sample conditions.

[0083] In a specific embodiment, the feature encoder includes a first branch, a second branch, a flattening layer, and a fully connected layer. The first branch includes a first convolutional module, a first max pooling layer, a second convolutional module, a second max pooling layer, a third max pooling layer, and a dropout layer connected in sequence. The second branch includes a first convolutional layer, a fourth max pooling layer, a fifth max pooling layer, and a sixth max pooling layer connected in sequence. The first convolutional module is connected to a second convolutional layer, a first batch normalization layer, and a first ReLU activation function layer in sequence. The second convolutional module is connected to a third convolutional layer, a second batch normalization layer, and a second ReLU activation function layer in sequence. The third convolutional module is connected to a fourth convolutional layer, a third batch normalization layer, and a third ReLU activation function layer in sequence.

[0084] The calculation process of the classifier is as follows:

[0085] ;

[0086] in, This indicates the metabolomic characteristics of serum samples from individuals to be predicted to have a second disease. The metabolomic features corresponding to the serum samples of individuals to be predicted to have a second disease are input into the feature encoder to obtain the embedded feature representation. The disease classification results obtained after inputting into the classifier. Representing embedded feature representation The corresponding disease classification result is the second disease. The conditional probabilities of each category;

[0087] The category with the highest conditional probability is selected as the disease classification result for the person to be predicted.

[0088] Specifically, the feature encoder mentioned in the embodiments of this application is used to learn discriminative representations from high-dimensional one-dimensional nuclear magnetic resonance spectra, mapping metabolomics features containing the original spectra to a low-dimensional embedding space, providing a unified feature representation for subsequent prototype vector construction and distance measurement.

[0089] In one example, such as Figure 2 As shown, the feature encoder uses a one-dimensional convolutional neural network with residual connections. The overall structure is as follows: The feature encoder includes a first branch and a second branch. The first branch consists of three concatenated one-dimensional convolutional modules and a max pooling operation with a stride of 2 after each layer. The three concatenated one-dimensional convolutional modules are the first convolutional module, the second convolutional module, and the third convolutional module. Each one-dimensional convolutional module includes a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function layer. That is, the first convolutional module is connected to the second convolutional layer, the first batch normalization layer, and the first ReLU activation function layer in sequence. The second convolutional module is connected to the third convolutional layer, the second batch normalization layer, and the second ReLU activation function layer in sequence. The third convolutional module is connected to the fourth convolutional layer, the third batch normalization layer, and the third ReLU activation function layer in sequence. Specifically, the first 1D convolutional module's 1D convolutional layer (i.e., the second convolutional layer in the first convolutional module) uses 32 convolutional kernels with a kernel size of 11 and padding of 5; the second 1D convolutional module's 1D convolutional layer (i.e., the third convolutional layer in the second convolutional module) uses 64 convolutional kernels with a kernel size of 5 and padding of 2; the third 1D convolutional module's 1D convolutional layer (i.e., the fourth convolutional layer in the third convolutional module) further uses 128 convolutional kernels with a kernel size of 3 and padding of 1, progressively downsampling the feature length to reduce the feature dimensionality. To further suppress overfitting, a Dropout layer (with a ratio of 0.4) is introduced at the end of the first branch.

[0090] To mitigate the feature degradation problem that may occur in deep convolutional neural networks under small sample conditions, a second branch is introduced in addition to the first branch, and the first branch and the second branch are residually connected. The second branch performs channel mapping on the metabolomics features of the original input through a first convolutional layer with a kernel size of 1 (128 channels), and performs the same number of max pooling operations as in the first branch to match the feature dimension. Finally, it is added element-wise with the output features of the first branch.

[0091] After the convolutional features are extracted, the output features are flattened and mapped to a 256-dimensional embedding space through a fully connected layer to obtain an embedded feature representation for prototype construction and cross-domain alignment.

[0092] During the classifier usage phase, the ¹H-NMR metabolometric data of serum samples from individuals with a second disease to be predicted are used to distinguish between the Euclidean distance between their embedded feature representations and the prototype vectors of each class. The class corresponding to the prototype vector with the smaller Euclidean distance is considered to be more similar to the sample.

[0093] Specifically, each source domain sample in the source domain dataset is input into the feature encoder to obtain the embedded feature representation of each source domain sample, which is the source domain feature; each target domain sample in the target domain dataset is input into the feature encoder to obtain the embedded feature representation of each target domain sample, which is the target domain feature.

[0094] In a specific embodiment, the total loss function used in the training process of the disease classification prediction model is constructed by weighting the feature distribution difference loss and the prototype loss, as shown in the following formula:

[0095] ;

[0096] in, For loss weighting coefficients, Indicates prototype loss. This represents the loss due to the difference in characteristic distribution. This represents the total loss function.

[0097] In a specific embodiment, the process of constructing the feature distribution difference loss is as follows:

[0098] The metabolomics features from the source and target domain datasets are input into the feature encoder to obtain the embedded feature representations of the source domain samples and the target domain samples, as shown in the following equation:

[0099] ;

[0100] ;

[0101] in, This represents the metabolomic features of the i-th sample in the source domain dataset. This represents the metabolomic feature of the j-th sample in the target domain dataset. This represents the function corresponding to the feature encoder. This represents the embedding feature representation of the i-th sample in the source domain dataset. This represents the embedding feature representation of the j-th sample in the target domain dataset;

[0102] The feature distribution difference loss between the source domain dataset and the target domain dataset is calculated using the maximum mean based on the embedded feature representations of the source domain samples and the target domain samples, as shown in the following formula:

[0103] ;

[0104] in, This represents the loss due to the difference in characteristic distribution. and These represent the number of samples in the source domain dataset and the target domain dataset, respectively. This represents the L2 norm.

[0105] Specifically, to address the potential differences in feature distribution between the embedded feature representations of source domain samples and target domain samples due to factors such as sample source and disease type, a feature distribution difference loss is introduced into the embedding space for cross-domain feature alignment, thereby improving the generalization performance of the disease classification prediction model under the condition of small sample size in the target domain.

[0106] After the source and target domain samples are mapped by a shared feature encoder, embedded feature representations of the source and target domain samples are obtained, respectively. To measure the difference in feature distributions between the two domains without relying on target domain label information, the maximum mean difference is used as a distribution alignment constraint. This feature distribution difference loss guides the feature encoder to learn metabolomics features with domain invariance by constraining the difference in the mean level between the embedded feature representations of the source and target domain samples, thereby reducing the distribution shift between different domains.

[0107] In a specific embodiment, the process of constructing the prototype loss is as follows:

[0108] The following formula is used to calculate the number of people suffering from the first disease in the first month. The prototype vector corresponding to each real category label:

[0109] ;

[0110] in, This indicates that the person suffering from the first disease is in the first The prototype vector corresponding to each real category label;

[0111] in, Indicates that the data in the source domain belongs to the first... A sample set of real category labels, Indicates that the data in the source domain belongs to the first... The number of samples in a sample set of real category labels;

[0112] The prototype loss is calculated using the following formula:

[0113] ;

[0114] in, This represents the metabolomics feature of the l-th sample in each batch of training data in the source domain dataset. This represents the total number of samples in each batch of training data in the source domain dataset. This indicates that the person suffering from the first disease is in the first The prototype vector corresponding to each real category label Represents Euclidean distance. This represents the total number of true category labels.

[0115] Specifically, to address the issues of limited sample size and overfitting issues in small-sample one-dimensional NMR spectroscopy data, traditional parametric classifiers are prone to overfitting. This application's embodiments introduce a prototypical learning-based metric classification strategy to achieve stable and interpretable disease discrimination. For example... Figure 3 As shown, the core idea of ​​prototype learning is to represent the class by the center of each class sample in the embedding space, thus transforming the classification problem into a distance metric problem between the sample and the prototype vector of each class. This method does not require learning independent classification weights for each class, which can effectively reduce model complexity, and is therefore particularly suitable for small-sample, high-dimensional metabolomics data scenarios.

[0116] To ensure good discriminative power in the embedding space, the optimization objective of prototype learning is to make the features of each source domain sample as close as possible to the prototype vector of its own class, while distancing them from prototype vectors of other classes. This transforms the distance metric into a probability distribution, where smaller distances between prototype vectors correspond to higher classification confidence. Based on this, the prototype loss is defined as the negative log-likelihood of the true class probability. By minimizing this prototype loss, the disease classification prediction model is constrained to form a compact within-class distribution in the embedding space, while increasing the margin between prototype vectors of different classes, thereby improving class discriminative power. The source domain dataset, containing true class labels, is used to calculate the prototype loss. Therefore, the Euclidean distance between the prototype vectors corresponding to the true class labels and their corresponding metabolomics features, input into the feature encoder, is used to construct the prototype loss, which can then be further transferred to the target domain.

[0117] Therefore, the overall optimization objective of the disease classification prediction model is composed of a weighted average of the prototype loss and the feature distribution difference loss. The loss weight coefficients are used to balance classification discriminative power and cross-domain alignment. In the embodiments of this application, the following is adopted: Through this joint optimization strategy, the disease classification prediction model, during training, utilizes the source domain dataset to construct stable prototype vectors for each category and enhance disease discrimination capabilities. Simultaneously, it aligns the feature distributions of the source and target domain datasets under unsupervised conditions. This enables effective classification of small-sample target domain NMR metabolomics data without the need for complex data augmentation or target domain labeling.

[0118] The disease classification prediction model mentioned in the embodiments of this application is implemented using Python version 3.9.21 and the PyTorch 2.7.0 framework. During training, the Adam optimizer (i.e., the adaptive moment estimator optimizer) is used. This optimizer combines momentum and adaptive learning rate when handling gradient descent, which can accelerate convergence and improve training stability. The initial learning rate of this optimizer is set to 0.0001, and the learning rate is gradually reduced through learning rate decay to avoid oscillations in the later stages of training.

[0119] In the pre-pubertal metabolic profile data of short stature, the trained disease classification prediction model achieved the following performance indicators: macro-mean sensitivity (SN) of 97.09%, macro-mean specificity (SP) of 98.39%, overall accuracy (ACC) of 96.76%, and F1-score of 96.73%. These results indicate that the trained disease classification prediction model exhibits high classification ability on the training data and maintains good generalization performance on the test data.

[0120] like Figure 4As shown, the area under the curve (AUC) of the trained disease classification prediction model for different types of prepubertal short stature reached 0.998 (PAC), 0.996 (PAG), and 0.998 (PAI), respectively. This result indicates that the trained disease classification prediction model has high discriminative ability in distinguishing between healthy individuals and different types of prepubertal short stature.

[0121] like Figure 5 As shown, the confusion matrix intuitively reflects the model's classification performance on three categories of samples: healthy, pre-pubescent GHD, and pre-pubescent ISS. The results indicate that virtually all samples can be correctly classified. This result demonstrates that the trained disease classification prediction model possesses strong discriminative ability in the overall classification task.

[0122] In the metabolic profile data of adolescent short stature, the trained disease classification prediction model achieved the following performance indicators: macro average sensitivity (SN) of 93.54%, macro average specificity (SP) of 97.62%, overall accuracy (ACC) of 95.12%, and F1-score of 92.68%. These results indicate that the trained disease classification prediction model can maintain high classification performance even with smaller sample conditions, balancing sensitivity and specificity, and demonstrating good generalization ability.

[0123] like Figure 6 As shown, the trained disease classification prediction model achieved AUCs of 1.000 (ADC), 0.993 (ADG), and 0.996 (ADI) for different malnutrition grades. This result indicates that the trained disease classification prediction model has high discriminative ability in predicting the stage of adolescent short stature.

[0124] like Figure 7 As shown, the confusion matrix intuitively reflects the classification performance of the trained disease classification prediction model on three categories of samples: healthy, adolescent GHD, and adolescent ISS. The results show that virtually all samples can be correctly classified. This demonstrates that the trained disease classification prediction model possesses strong discriminative ability in the overall classification task and exhibits good discrimination on the adolescent short stature metabolic profile dataset.

[0125] Experimental results show that the trained disease classification prediction model proposed in the embodiments of this application still maintains high sensitivity and high specificity on a dwarfism dataset with a limited number of samples. It can accurately identify the metabolomics differences between diseased samples and normal control samples, demonstrating its robustness and reliability in small-sample disease classification tasks.

[0126] S3. Obtain the ¹H-NMR metabolomic profile data of serum samples from individuals with a second disease to be predicted and preprocess them to obtain the corresponding metabolomic features. Input the metabolomic features into the trained disease classification prediction model. First, it passes through the feature encoder to obtain the corresponding embedded feature representation. The embedded feature representation passes through the classifier to obtain the disease classification result of the individual to be predicted.

[0127] Specifically, the trained disease classification prediction model is deployed. The ¹H-NMR metabolomic data of serum samples from individuals suspected of having a second disease are preprocessed to obtain corresponding metabolomic features; the preprocessing procedure is the same as that used during training. These metabolomic features are then input into the trained disease classification prediction model to obtain the disease classification result for the individuals to be predicted.

[0128] Further reference Figure 8 As an implementation of the methods shown in the above figures, this application provides an embodiment of a disease classification prediction device based on prototype transfer learning. This device embodiment is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0129] This application provides a disease classification prediction device based on prototype transfer learning, including:

[0130] Data processing module 1 is configured to acquire ¹H-NMR metabolochromatographic data of serum samples from individuals suffering from the first disease and the second disease, respectively, and preprocess them to obtain the corresponding metabolomic features; construct a source domain dataset based on the true category labels of individuals suffering from the first disease and their corresponding metabolomic features, and construct a target domain dataset based on the metabolomic features of individuals suffering from the second disease.

[0131] Model building module 2 is configured to build a disease classification prediction model, which includes a feature encoder and a classifier; the disease classification prediction model is trained using prototype transfer learning based on the source domain dataset and the target domain dataset to obtain the trained disease classification prediction model.

[0132] Classification module 3 is configured to acquire ¹H-NMR metabolomic profile data of serum samples from individuals with a second disease to be predicted and preprocess them to obtain corresponding metabolomic features. The metabolomic features are then input into a trained disease classification prediction model. The model first passes through a feature encoder to obtain the corresponding embedded feature representation. The embedded feature representation is then passed through a classifier to obtain the disease classification result of the individual to be predicted.

[0133] Figure 9 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. For example... Figure 9As shown, the electronic device in this embodiment includes a processor 901 and a memory 902; wherein the memory 902 is used to store computer execution instructions; and the processor 901 is used to execute the computer execution instructions stored in the memory to implement the various steps performed by the electronic device in the above embodiment. For details, please refer to the relevant descriptions in the foregoing method embodiments.

[0134] Alternatively, the memory 902 can be either standalone or integrated with the processor 901.

[0135] When the memory 902 is set up independently, the electronic device also includes a bus 903 for connecting the memory 902 and the processor 901.

[0136] This invention also provides a computer storage medium storing computer execution instructions, which, when executed by processor 901, implement the above method.

[0137] This invention also provides a computer program product, including a computer program that, when executed by a processor 901, implements the above-described method.

[0138] In the embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0139] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0140] Furthermore, the functional modules in the various embodiments of this invention can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit formed by the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0141] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor 901 to execute some steps of the methods of the various embodiments of this application.

[0142] It should be understood that the processor 901 described above can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor, or the processor 901 can be any conventional processor 901. The steps of the method disclosed in this invention can be directly manifested as execution by the hardware processor 901, or execution by a combination of hardware and software modules within the processor 901.

[0143] The memory 902 may include high-speed RAM memory, and may also include non-volatile memory NVM, such as at least one disk storage device, and may also be a USB flash drive, portable hard drive, read-only memory, disk or optical disc, etc.

[0144] Bus 903 can be an Industry Standard Architecture (ISA), a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Bus 903 can be divided into address bus, data bus, control bus, etc. For ease of illustration, the bus 903 in the accompanying drawings of this application is not limited to only one bus 903 or one type of bus 903.

[0145] The aforementioned storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium accessible to general-purpose or special-purpose computers.

[0146] An exemplary storage medium is coupled to a processor 901, enabling the processor 901 to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor 901. The processor 901 and the storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor 901 and the storage medium can exist as discrete components in an electronic device or a host device.

[0147] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A disease classification prediction method based on prototype transfer learning, characterized in that, Includes the following steps: ¹H-NMR metabolomic profile data of serum samples from individuals suffering from the first and second diseases were obtained and preprocessed to obtain the corresponding metabolomic characteristics. A source domain dataset is constructed based on the true category labels of people with the first disease and their corresponding metabolomics features, and a target domain dataset is constructed based on the metabolomics features of people with the second disease. A disease classification prediction model is constructed, which includes a feature encoder and a classifier; the disease classification prediction model is trained using the source domain dataset and the target domain dataset based on prototype transfer learning to obtain a trained disease classification prediction model. The ¹H-NMR metabolomic profile data of serum samples from individuals with a second disease to be predicted are obtained and preprocessed to obtain corresponding metabolomic features. The metabolomic features are then input into a trained disease classification prediction model. The model first passes through the feature encoder to obtain the corresponding embedded feature representation. The embedded feature representation is then passed through the classifier to obtain the disease classification result of the individual to be predicted.

2. The disease classification prediction method based on prototype transfer learning according to claim 1, characterized in that, The total loss function used in the training process of the disease classification prediction model is constructed by weighting the feature distribution difference loss and the prototype loss, as shown in the following formula: ; in, For loss weighting coefficients, Indicates prototype loss. This represents the loss due to the difference in characteristic distribution. This represents the total loss function.

3. The disease classification prediction method based on prototype transfer learning according to claim 2, characterized in that, The construction process of the feature distribution difference loss is as follows: The metabolomics features from the source domain dataset and the target domain dataset are input into the feature encoder to obtain the embedded feature representations of the source domain samples and the target domain samples, as shown in the following equation: ; ; in, This represents the metabolomic features of the i-th sample in the source domain dataset. This represents the metabolomic features of the j-th sample in the target domain dataset. This represents the function corresponding to the feature encoder. This represents the embedding feature representation of the i-th sample in the source domain dataset. This represents the embedding feature representation of the j-th sample in the target domain dataset; The feature distribution difference loss between the source domain dataset and the target domain dataset is calculated using the maximum mean based on the embedded feature representations of the source domain samples and the target domain samples, as shown in the following formula: ; in, This represents the loss due to the difference in characteristic distribution. and These represent the number of samples in the source domain dataset and the target domain dataset, respectively. This represents the L2 norm.

4. The disease classification prediction method based on prototype transfer learning according to claim 3, characterized in that, The process of constructing the prototype loss is as follows: The following formula is used to calculate the number of people suffering from the first disease in the first month. The prototype vector corresponding to each real category label: ; in, This indicates that the person suffering from the first disease is in the first The prototype vector corresponding to each real category label; in, This indicates that the data in the source domain belongs to the first... A sample set of real category labels, This indicates that the data in the source domain belongs to the first... The number of samples in a sample set of real category labels; The prototype loss is calculated using the following formula: ; in, This represents the metabolomics feature of the l-th sample in each batch of training data in the source domain dataset. This represents the total number of samples in each batch of training data in the source domain dataset. This indicates that the person suffering from the first disease is in the first The prototype vector corresponding to each real category label Represents Euclidean distance. This represents the total number of true category labels.

5. The disease classification prediction method based on prototype transfer learning according to claim 4, characterized in that, The feature encoder includes a first branch, a second branch, a flattening layer, and a fully connected layer. The first branch includes a first convolutional module, a first max pooling layer, a second convolutional module, a second max pooling layer, a third max pooling layer, and a Dropout layer connected in sequence. The second branch includes a first convolutional layer, a fourth max pooling layer, a fifth max pooling layer, and a sixth max pooling layer connected in sequence. The first convolutional module is connected to a second convolutional layer, a first batch normalization layer, and a first ReLU activation function layer in sequence. The second convolutional module is connected to a third convolutional layer, a second batch normalization layer, and a second ReLU activation function layer in sequence. The third convolutional module is connected to a fourth convolutional layer, a third batch normalization layer, and a third ReLU activation function layer in sequence. The calculation process of the classifier is as follows: ; in, This indicates the metabolomic characteristics of serum samples from individuals to be predicted to have a second disease. The embedded feature representation is obtained by inputting the metabolomic features corresponding to the serum sample of the person to be predicted to have a second disease into the feature encoder. The disease classification results obtained after inputting into the classifier. Representing embedded feature representation The corresponding disease classification result is the second disease. The conditional probabilities of each category; The category corresponding to the highest conditional probability is selected as the disease classification result for the person to be predicted.

6. The disease classification prediction method based on prototype transfer learning according to claim 1, characterized in that, Both the first and second diseases are related to children's growth and development, and their clinical manifestations are associated with metabolomic characteristics. In contrast, the serum samples of people with the first disease are large samples, while the serum samples of people with the first disease are small samples.

7. A disease classification prediction device based on prototype transfer learning, characterized in that, include: The data processing module is configured to acquire ¹H-NMR metabolospectral data of serum samples from individuals suffering from the first and second diseases, respectively, and preprocess them to obtain the corresponding metabolomic characteristics. A source domain dataset is constructed based on the true category labels of people with the first disease and their corresponding metabolomics features, and a target domain dataset is constructed based on the metabolomics features of people with the second disease. The model building module is configured to build a disease classification prediction model, which includes a feature encoder and a classifier; the disease classification prediction model is trained using the source domain dataset and the target domain dataset based on prototype transfer learning to obtain a trained disease classification prediction model. The classification module is configured to acquire ¹H-NMR metabolomic profile data of serum samples from individuals to be predicted who have a second disease and preprocess them to obtain corresponding metabolomic features. The metabolomic features are then input into a trained disease classification prediction model, first passing through the feature encoder to obtain the corresponding embedded feature representation. The embedded feature representation is then passed through the classifier to obtain the disease classification result of the individual to be predicted.

8. An electronic device, comprising: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-6.