Disease classification model construction method and device, electronic equipment and storage medium

Abstract

The application provides a disease classification model construction method and device, electronic equipment and storage medium. The method comprises the following steps: performing wave band screening on a plasma ATR-FTIR spectrum; dividing the spectrum into an original training set and a test set by using a KS algorithm; performing image processing by using a Markov transition field method; converting a two-dimensional matrix into a one-dimensional feature vector; performing dimension reduction processing on the one-dimensional feature vector by using principal component analysis; performing sample augmentation on minority class samples in the dimension reduction training set by using a synthetic minority over-sampling technique; combining the dimension reduction training set and virtual samples to obtain an augmented training set; and constructing a classification model by using a joint sparse boundary Fisher regularization classification algorithm. The application amplifies the weak spectral difference between majority class samples and minority class samples, and can balance the number of majority class samples and minority class samples in the training set, significantly improves the recognition sensitivity of target class patients, and is used for classifying mental diseases or neurodegenerative diseases.

Description

Technical Field

[0001] This application relates to the field of classification model construction technology, and in particular to a method, apparatus, electronic device and storage medium for constructing a disease classification model. Background Technology

[0002] For the clinical diagnosis of mental illnesses and neurodegenerative diseases, current methods mainly rely on three approaches: scale assessment, imaging examinations, and humoral biomarker testing. Scale assessments heavily depend on patient cooperation and the professional experience of the administerer, making it difficult to guarantee consistent results. This is especially true in adolescents, where the volatility and subtlety of emotional expression further increase the risk of misdiagnosis. Imaging equipment is expensive, complex to operate, and some techniques involve radiation exposure, making it difficult to promote in primary healthcare institutions and unsuitable as a large-scale screening tool. While cerebrospinal fluid testing has high diagnostic accuracy, lumbar puncture is an invasive procedure with poor patient compliance and risks of infection, making it unsuitable as a routine follow-up method.

[0003] Among related technologies, attenuated total reflectance-Fourier transform infrared spectroscopy (ATR-FTIR) has shown potential for minimally invasive, rapid, and low-cost plasma testing due to its ability to reflect the vibrational information of biomolecules. However, it faces two common core challenges in clinical translation. First, the disease-related spectral features are extremely weak. Whether it is a mental illness or a neurodegenerative disease, the differences in molecular vibrational information between target category patients and controls in plasma samples are very subtle. These subtle changes are often submerged in complex background signals, and traditional linear analysis methods and shallow learning models are difficult to effectively extract deep features with discriminative value. Furthermore, the inherent imbalance in class distribution in clinical sample collection means that control samples are relatively easy to obtain and are usually plentiful in number. However, the target patient group often has significantly fewer samples than the control group due to the high degree of disease concealment, atypical early symptoms, and low consultation rate. This disparity in sample size can easily lead to prediction bias in the classification model during training, causing it to tend to classify samples as the majority class, thus sacrificing the accuracy of identifying the minority class of target patients. This makes it difficult for the model to balance the two key indicators of sensitivity and specificity in practical applications. Summary of the Invention

[0004] This application provides a method, apparatus, electronic device, and storage medium for constructing a disease classification model, which at least to some extent overcomes the problem that the classification accuracy is limited due to the extreme similarity of the spectra of majority and minority class samples in ATR-FTIR spectral analysis in related technologies, and the imbalance in the number of clinical samples between majority and minority class samples.

[0005] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0006] According to one aspect of this application, a method for constructing a disease classification model is provided, comprising: collecting ATR-FTIR spectra of sample plasma; performing band screening on the plasma ATR-FTIR spectra based on signal-to-noise ratio and biospectral characteristics to obtain an original spectral sample set; dividing the original spectral sample set into an original training set and a test set using the KS algorithm; wherein the samples include majority class samples and minority class samples; performing image processing on the original training set and the test set using the Markov transfer field method to obtain a two-dimensional matrix; converting the two-dimensional matrix into a one-dimensional feature vector; and performing dimensionality reduction processing on the one-dimensional feature vector using principal component analysis to obtain... The process involves: obtaining dimensionality-reduced training and test sets; acquiring minority class samples from the dimensionality-reduced training set and augmenting them using synthetic minority oversampling to obtain a virtual sample set; merging the dimensionality-reduced training set with the virtual sample set to obtain the augmented training set; constructing a classification model based on the augmented training set using a joint sparse boundary Fisher regularization classification algorithm; verifying the performance of the classification model using the test set and qualitatively analyzing its classification effect using classification evaluation metrics; and qualitatively analyzing the classification effect and generalization ability of the classification model by inputting the augmented training set into the classification model for prediction and calculating evaluation metrics based on the confusion matrix.

[0007] In some embodiments, the original training set and the test set are image-processed using the Markov transfer field method to obtain a two-dimensional matrix; converting the two-dimensional matrix into a one-dimensional feature vector includes: dividing the x-domain of a single spectral data point into... Q There are quantile intervals with equal probability, and the spectral data of a single sample is denoted as . x 1, x 2,…, x D Each quantile interval is denoted as Every moment All can be mapped to their respective... Intervals; Construct Markov transition matrix ,

[0008] in, Let be the single-step transition probability, representing The time is located at the The elements of each quantile interval, in t The moment jumps to the first The probability of each quantile interval; Constructing Markov transition fields ,

[0009] Wherein, mapping function Will The index mapped to its corresponding quantile interval, i.e. , express The first mapping quantile intervals Indicates from Corresponding quantile region Transferred to Corresponding quantile region The probability of spectral data; Single sample data in Transform into a Markov transition field Next, Resampling to the specified size Two-dimensional matrix ;Will Flatten along the row into a row vector ;Will n indivual The final image feature matrix is ​​obtained by stacking the rows. .

[0010] In some embodiments, principal component analysis is used to reduce the dimensionality of the one-dimensional feature vector to obtain a dimensionality-reduced training set and a dimensionality-reduced test set, including: for the image-based feature matrix Its covariance matrix is Find a set of orthogonal bases This maximizes the variance of the projected samples and ensures that the principal components are uncorrelated. The resulting new samples after projection... The expression is ,in, It is composed of the covariance matrix The former d A matrix consisting of the eigenvectors corresponding to the largest eigenvalues. d The target dimension is the sample matrix after dimensionality reduction. .

[0011] In some embodiments, minority class samples are obtained from the dimensionality-reduced training set, and synthetic minority oversampling is used to augment the minority class samples in the dimensionality-reduced training set to obtain a virtual sample set; the dimensionality-reduced training set and the virtual sample set are then combined to obtain the augmented training set, including: the dimensionality-reduced training set being... ,in, n trn The total number of samples in the dimensionality reduction training set. d The feature dimensions after dimensionality reduction for PCA; the dimensionality-reduced training set includes majority class samples. and minority class samples Constructing virtual samples of a minority class , Among them, the minority class of samples are virtual samples. From real samples and its K nearest neighbor samples linear combination, It is a random number in the interval [0, 1]; construct a virtual sample set. ; the original training set and Merging to obtain the augmented training set .

[0012] In some embodiments, the objective function of the classification model is: Where y is the true label of the sample, and w is a weight vector containing a bias term, used to determine the category of the projected sample. Let be the projection matrix. represent Norm, express Norm, It is the experience loss term, used to measure the degree of deviation between the predicted result and the true label; and The MFA module measures intra-class compactness and inter-class separability of samples, respectively. To smooth the regularization term, the L2 norm of w is used to limit the complexity of the model; For sparse regularization terms, through P Norms are used to impose sparsity constraints, preserving key features while eliminating redundant information; These are the weighting coefficients for the corresponding items; L w Let L be the inner Laplace matrix. b The inter-class Laplacian matrix; the discriminant function of the classification model. for Where t is a one-dimensional eigenvector, Let w be the projection matrix and w be the weight vector.

[0013] In some embodiments, the classification model is used to classify mental illnesses; the majority class samples are healthy control samples, and the minority class samples are patient samples with mental illnesses.

[0014] In some embodiments, the classification model is used to classify neurodegenerative diseases; the majority class samples are patient samples of Alzheimer's disease, and the minority class samples are patient samples of mild cognitive impairment.

[0015] According to another aspect of this application, a disease classification model construction apparatus is provided, comprising: a sample data acquisition and preprocessing module: used to acquire plasma ATR-FTIR spectra; to perform band screening on the plasma ATR-FTIR spectra based on signal-to-noise ratio and biospectral characteristics to obtain an original spectral sample set; to divide the original spectral sample set into an original training set and a test set using the KS algorithm; wherein the samples include majority class samples and minority class samples; an image processing module: used to perform image processing on the original training set and the test set using the Markov transfer field method to obtain a two-dimensional matrix; to convert the two-dimensional matrix into a one-dimensional feature vector; and a data dimensionality reduction module: used to reduce the dimensionality of the one-dimensional feature vector using principal component analysis. The system consists of several modules: a dimensionality reduction module (for training and testing sets), a sample augmentation module (for minority class samples in the training set and synthetic minority oversampling to augment these samples to obtain a virtual sample set), a model building module (for constructing a classification model using the Fisher regularization algorithm with joint sparse boundary conditions), a model validation module (for validating the classification model performance on the test set and qualitatively analyzing its classification performance using classification evaluation metrics), and a model verification module (for inputting the augmented training set into the classification model for prediction and calculating evaluation metrics based on the confusion matrix to qualitatively analyze the classification performance and generalization ability of the model).

[0016] According to another aspect of this application, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform the disease classification model construction method by executing the executable instructions.

[0017] According to another aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for constructing a disease classification model.

[0018] The technical solutions provided in the embodiments of this application include at least the following intentional effects: the classification model construction method of this application amplifies the differences in spectral features between majority class samples and minority class samples, and significantly alleviates the prediction bias problem caused by sample class imbalance in clinical data.

[0019] The technical solution provided in the embodiments of this application converts one-dimensional ATR-FTIR spectral data into two-dimensional images through Markov transfer fields, effectively capturing the nonlinear dynamic pathological features hidden in the spectrum, amplifying the weak spectral differences between majority and minority class samples, and improving the sensitivity of target category identification.

[0020] Furthermore, the Synthetic Minority Oversampling Technique (SMOTE) is employed to generate virtual samples of the minority class in the feature space, thereby balancing the number of majority and minority class samples in the training set. This avoids the prediction bias caused by sample imbalance in traditional models and significantly improves the sensitivity of identifying patients of the target class.

[0021] Furthermore, the JSMFRC algorithm jointly optimizes projection learning and classifier training, enabling the projection direction to directly serve the classification task. It introduces a boundary Fisher analysis (MFA) module to maintain the local geometric structure of biological data through intra-class compactness and inter-class separability constraints. By combining smoothing regularization and sparsity regularization, it removes redundant information while retaining key features, thereby enhancing the robustness and generalization ability of the model. Compared with traditional algorithms, JSMFRC achieves an excellent balance between sensitivity and specificity while maintaining the highest accuracy.

[0022] Furthermore, plasma ATR-FTIR spectroscopy detection only requires peripheral blood samples and does not require lumbar puncture or radioactive tracers, offering advantages such as being minimally invasive, radiation-free, and easy to operate. ATR-FTIR has a short detection cycle, making it suitable for promotion in primary healthcare institutions, and providing feasible technical support for large-scale screening and pre-diagnosis of mental illnesses. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0024] Figure 1 This diagram illustrates the structure of a disease classification model construction system in one embodiment of this application.

[0025] Figure 2 This document illustrates a flowchart of a disease classification model construction method in one embodiment of this application.

[0026] Figure 3 The ATR-FTIR spectrum of Embodiment 1 of this application is shown.

[0027] Figure 4 The diagram shows the original training set after PCA dimensionality reduction, the virtual samples generated by SMOTE, and the scatter plot of the test set in Embodiment 1 of this application.

[0028] Figure 5 The ROC curves of different models in Embodiment 1 of this application are shown.

[0029] Figure 6This diagram illustrates a disease classification model construction apparatus according to one embodiment of the present application.

[0030] Figure 7 This diagram illustrates the structure of an electronic device according to one embodiment of the present application. Detailed Implementation

[0031] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0032] Furthermore, the accompanying drawings are merely illustrative of this application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0033] It should be noted that the various types of data obtained in this application embodiment, such as personal identity data, operational data, and behavioral data related to individuals, customers, and groups, have all been authorized.

[0034] The specific implementation methods of the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0035] Figure 1 A schematic diagram of an exemplary application system architecture is shown, illustrating a disease classification model construction method applicable to embodiments of this application. For example... Figure 1 As shown, the system architecture may include a model building device 11, a first server 12, and a second server 13.

[0036] The first server 12 can store plasma ATR-FTIR spectral data of both majority and minority class samples.

[0037] The model building device 11 can acquire plasma ATR-FTIR spectral data of majority class and minority class samples from the first server 12 via the network, and the acquired spectral data can be stored locally on the model building device 11.

[0038] The second server 13 is used to host the comparison model. The model building device 11 can send requests to the second server 13 via the network to apply the comparison model. The specific model to be compared is not limited in the embodiments of this application. For example, the comparison model includes Logistic Regression (LogRegress), Marginal Fisher Analysis (MFA), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), and Linear Graph Embedding (LGE), etc.

[0039] For example, the model building device 11 can send a request to the second server 13 to verify the classification performance of the classification model built in this application using the comparison model.

[0040] The model building device 11 and the first server 12, and the model building device 11 and the second server 13, are connected via a network. This network can be a wired network or a wireless network. Optionally, the aforementioned wireless or wired network uses standard communication technologies and / or protocols. The network is typically the Internet, but can also be any network, including but not limited to a Local Area Network (LAN), Metropolitan Area Network (MAN), Wide Area Network (WAN), mobile, wired or wireless network, private network, or any combination of virtual private network. In some embodiments, technologies and / or formats including HyperText Markup Language (HTML), Extensible Markup Language (XML), etc., are used to represent data exchanged over the network. In addition, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), and Internet Protocol Security (IPSec) can be used to encrypt all or some of the links. In other embodiments, custom and / or dedicated data communication technologies can be used to replace or supplement the aforementioned data communication technologies.

[0041] The model building device 11 can be various electronic devices, including but not limited to smartphones, tablets, laptops, desktop computers, etc.

[0042] Both the first server 12 and the second server 13 can be servers that provide various services. Optionally, the server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0043] Under the above system architecture, this application provides a method for constructing a disease classification model, which can be executed by any electronic device with computing power.

[0044] In some embodiments, the disease classification model construction method provided in this application can be executed by a terminal device of the above-described system architecture; in other embodiments, the disease classification model construction method provided in this application can be executed by a server in the above-described system architecture; in still other embodiments, the disease classification model construction method provided in this application can be implemented by the terminal device and the server in the above-described system architecture through interaction.

[0045] Figure 2 This document illustrates a flowchart of a disease classification model construction method according to an embodiment of this application, as shown below. Figure 2 As shown, the disease classification model construction method provided in this application embodiment includes the following steps S201 to S206: S201. Collect ATR-FTIR spectra of plasma samples; perform band screening on plasma ATR-FTIR spectra based on signal-to-noise ratio and biospectral characteristics to obtain an original spectral sample set; use the KS algorithm to divide the original spectral sample set into an original training set and a test set; wherein, the samples include majority class samples and minority class samples.

[0046] In some embodiments, the instrument used for spectral acquisition is not limited in this application; for example, a Bruker Alpha ATR-FTIR spectrometer. The parameters for spectral acquisition are not limited in this application; for example, the spectral acquisition parameters are set to a wavenumber range of 4000-600 cm⁻¹. -1 4cm resolution -1 A total of 32 scans were performed.

[0047] It should be noted that the purpose of band screening of the plasma ATR-FTIR spectrum based on signal-to-noise ratio and biospectral characteristics is to screen out bands with high signal-to-noise ratio and preserved biospectral characteristics in the plasma ATR-FTIR spectrum of each patient.

[0048] It should be noted that biospectral characteristics reflect the characteristic spectra of changes in the composition and structure of biomolecules under disease states. In other words, the screening process must retain the wavelengths that reflect the characteristics of a certain disease.

[0049] In some embodiments, AD-related biospectral features include amide I / II bands (1700-1500 cm⁻¹) reflecting secondary structural abnormalities of Aβ and p-tau proteins. -1 ), reflecting lipid metabolism disorders, the carbonyl (C=O) stretching vibration band (1800-1700 cm). -1 ) and the stretching vibration band of hydroxyl (OH) related to carbohydrate metabolism (3600-3200cm) -1 The low-frequency region is mainly affected by non-specific interference from residual salts and high-frequency detection noise, resulting in a low signal-to-noise ratio.

[0050] In some embodiments, the biospectral characteristics of adolescent depression include phosphodiester bonds (PO2) reflecting nucleic acid and phospholipid metabolism. - ) Vibration belt (1200-1000cm) -1 Amide I / II bands (1700-1500 cm⁻¹) indicating protein secondary structure abnormalities. -1 ) and the alkyl (CH2 / CH3) stretching vibration band (3000-2800cm) reflecting lipid metabolism disorders. -1 ).

[0051] In some embodiments, regions with a signal-to-noise ratio (SNR) of less than 15000:1 (peak-to-peak value PP) are excluded.

[0052] In some embodiments, the KS algorithm is used to divide the original spectral sample set into an original training set and a test set in a 3:1 ratio.

[0053] S202. The original training set and the test set are imaged using the Markov Transition Field (MTF) method to obtain a two-dimensional matrix; the two-dimensional matrix is ​​then converted into a one-dimensional feature vector (by row).

[0054] It should be noted that, since the spectra of the majority class and minority class samples are extremely similar, MTF is used to convert the one-dimensional ATR-FTIR to two dimensions to amplify the subtle differences in their features, and then the data is straightened out row by row to form one-dimensional data for the convenience of subsequent steps.

[0055] In some embodiments, step S202 includes S2021 to S2026: S2021, quantile interval division, discretizes continuous spectral values ​​into a finite state space, which facilitates the subsequent construction of Markov transition matrices.

[0056] In some embodiments, the value range of a single spectral data x is divided based on the original training and test sets. Q There are quantile intervals with equal probability, and the spectral data of a single sample is denoted as . x 1, x 2,…, x D Each quantile interval is denoted as Every moment All can be mapped to their respective... Interval.

[0057] In some embodiments, the number of equally probable quantile intervals Q This application does not impose any restrictions, for example, Q =5.

[0058] S2022, Constructing the Markov transition matrix It is used to capture the dynamic changes between points in a spectral sequence and reflect the temporal structure information of the spectrum.

[0059] By statistically analyzing the frequency of data points transitioning from one region to another between adjacent time points, and converting this frequency into probabilities, we obtain the Markov transition matrix. : (1) Single-step transition probability ,express The time is located at the The elements of each quantile interval, in t The moment jumps to the first The probability of each quantile interval.

[0060] S2023, Constructing a Markov Transfer Field This is used to expand a one-dimensional spectrum into a two-dimensional image, where each pixel represents the state transition probability between two time points, thereby enhancing the expression of weak features.

[0061] In some embodiments, by The probabilities are recursively derived through matrix exponentiation, and the transition probabilities at any time step are statistically analyzed and concatenated to form the Markov transition field matrix. : (2) Wherein, mapping function Will The index mapped to its corresponding quantile interval, i.e. , express The first mapping quantile intervals Indicates from Corresponding quantile region Transferred to Corresponding quantile region The probability of.

[0062] S2024, resampling, is used to unify the image size of all samples, which facilitates subsequent processing, reduces computational complexity, and maintains feature consistency.

[0063] In some embodiments, according to step S2023, the spectral data is first... Single sample data in Transform into a Markov transition field Next, Resampling to the specified size Two-dimensional matrix .in, The size is not limited in this application, for example, D img =256.

[0064] S2025, Flattening, is used to transform image features into a vector form suitable for input to machine learning models.

[0065] In some embodiments, Flatten along the row into a row vector .

[0066] S2026, Stacking, is used to construct a complete feature matrix for use in the next step of PCA dimensionality reduction.

[0067] In some embodiments, n indivual The final image feature matrix is ​​obtained by stacking the rows. .

[0068] S203. Principal component analysis is used to reduce the dimensionality of the one-dimensional feature vector to obtain a dimensionality-reduced training set and a dimensionality-reduced test set.

[0069] It should be noted that this step is used to reduce the dimensionality of the high-dimensional image feature matrix Z through principal component analysis, extracting the most important feature components, in preparation for subsequent sample balancing and classification modeling.

[0070] In some embodiments, the result of step S2026 is invoked, at which point the feature dimension is very high, containing a large amount of redundant information and noise.

[0071] For example, , n The total number of samples, This represents the feature dimension of each sample after flattening. The flattened feature dimension is... .

[0072] In some embodiments, for the obtained image matrix Its covariance matrix is Find a set of orthogonal bases This maximizes the variance of the projected samples and ensures that the principal components are uncorrelated. The new sample after projection... The expression is as follows: (3) in It is composed of the covariance matrix The former d A matrix consisting of the eigenvectors corresponding to the largest eigenvalues. d The target dimension. The sample matrix after PCA dimensionality reduction. .

[0073] Perform the above operations on the original training set and the test set respectively to obtain the dimensionality-reduced training set and the dimensionality-reduced test set.

[0074] S204. Obtain minority class patient samples from the dimensionality reduction training set, and use synthetic minority oversampling technology to augment the minority class patient samples in the dimensionality reduction training set to obtain a virtual sample set; combine the dimensionality reduction training set and the virtual sample set to obtain an augmented training set.

[0075] In some embodiments, step S204 includes S2041 to S2046: S2041, Call the original training set after dimensionality reduction.

[0076] The original training set after dimensionality reduction is denoted as the dimensionality-reduced training set. ,in, n trn The total number of samples in the original training set after dimensionality reduction. d The feature dimensions after dimensionality reduction by PCA, for example, d =68.

[0077] It should be noted that the majority class samples are far more numerous than the minority class samples, resulting in severe class imbalance. The purpose of this step is to increase the number of minority class samples through SMOTE (Synthetic Minority Oversampling Technique), thereby alleviating the model prediction bias caused by class imbalance.

[0078] In some embodiments, the dimensionality reduction training set includes majority class samples and minority class samples, wherein the majority class samples are more numerous and are denoted as majority class samples. The minority class samples are fewer in number and are denoted as minority class samples. .

[0079] S2042. Determine the minority class samples and the number generated, clarify the target category to be augmented and the augmentation scale, so that the category ratio of the augmented training set is more balanced.

[0080] In some embodiments, a minority class sample set is separated from the dimensionality-reduced training set. Determine the number of virtual samples to be generated. .

[0081] S2043. For each minority class sample, find K nearest neighbors to ensure that the new sample lies within a local region of the minority class samples. For each sample in the minority class... Calculate the Euclidean distance within the minority class sample set to find The nearest neighbor.

[0082] S2044. Virtual samples are generated by randomly interpolating along the lines connecting minority class samples to their nearest neighbors. In some embodiments, It is by any single sample The generated virtual sample has the following expression: (4) Minority class virtual samples From real samples and nearest neighbor samples linear combination, It is a random number in the interval [0, 1].

[0083] S2045. Construct a virtual sample set to form an augmented dataset that is similar in distribution to the original minority class samples but has a larger number of samples.

[0084] In some embodiments, Representative generated A virtual sample set consisting of a minority class of samples.

[0085] S2046. Merge to obtain an augmented training set, resulting in more balanced training data for subsequent classification model training, avoiding model bias towards the majority class.

[0086] In some embodiments, the original training set and

[0087] Merging to obtain the augmented training set Test set used It means that among them m and s These are the number of samples in the augmented training set and the test set, respectively.

[0088] S205. Based on the augmented training set, a classification model is constructed using the Joint Sparse Boundary Fisher Regularized Classification Algorithm (JSMFRC).

[0089] In some embodiments, the JSMFRC classifier optimization objective function is: Where y is the true label of the sample, and w is a weight vector containing a bias term, used to determine the category of the projected sample. Let be the projection matrix. represent Norm, express Norm. In the loss function, It is the experience loss term, used to measure the degree of deviation between the predicted result and the true label; and The MFA module measures intra-class compactness and inter-class separability of samples, respectively. For example, the intra-class nearest neighbor is 25, and the inter-class nearest neighbor is 250, i.e., k w =25,k b =250; To smooth the regularization term, the L2 norm of w is used to limit the complexity of the model; For sparse regularization terms, through P Norms are used to impose sparsity constraints, preserving key features while eliminating redundant information. These are the weighting coefficients for the corresponding terms, L w Let L be the inner Laplace matrix. b It is the inter-class Laplace matrix.

[0090] Based on the projection matrix P and the weight vector w, the discriminant function for any input t (one-dimensional feature vector) It can be represented as: Where Majority class represents the majority class samples; Minority class represents the minority class samples.

[0091] In some embodiments, the graph construction process of MFA is as follows: For the training set Each sample in ,use and They represent similar k One nearest neighbor sample and the other class k The similarity matrix between two nearest neighbor samples. Separation matrix between heterogeneous samples It can be represented as: (5) (6) Constructing the intraclass Laplace matrix ,in It is a diagonal matrix, and its diagonal elements Similarly, the inter-class Laplace matrix is ​​obtained. Use intra-class compactness and inter-class separation This is used to measure the degree of clustering of similar samples and the degree of dispersion of dissimilar samples after projection. and It is expressed as follows (7) (8) in, Let be the projection matrix, and tr() denotes the trace of the matrix. The ultimate goal of MFA is to determine an optimal projection P such that, after projection, samples of the same class are as close as possible, and samples of different classes are as far apart as possible, i.e. as small as possible As large as possible. (Referring to the MFA algorithm) and Based on this, by adding empirical terms, smoothing regularization terms, and sparsity regularization terms, we obtain the Joint Sparse Marginal Fisher and Regularized Classification (JSMFRC) algorithm, whose objective function is... .

[0092] S206. Verify the performance of the classification model based on the test set, and qualitatively analyze the classification effect of the model using classification evaluation metrics; based on the augmented training set, make predictions by inputting the augmented training set into the classification model, and calculate evaluation metrics based on the confusion matrix to qualitatively analyze the classification effect and generalization ability of the classification model.

[0093] The qualitative evaluation index system based on the confusion matrix mainly includes the following indicators: accuracy. Sensitivity ), Specificity F1 score (F1-Score) ) and receiver operating characteristic (ROC) curve. Definition These are categorized as True Yang (predicted positive, actually positive), True Yin (predicted negative, actually negative), False Yang (predicted positive, actually negative), and False Yin (predicted negative, actually positive). The formulas for each indicator are as follows: (9); (10); (11); (12); The ROC curve, with the false positive rate on the horizontal axis and the true positive rate on the vertical axis, intuitively evaluates the performance of the classification model. The larger the AUC value (area under the ROC curve), the stronger the classifier performance.

[0094] It should be noted that in this application, the classification model is used to classify mental illnesses or neurodegenerative diseases. The specific type of mental illness is not limited in this application; for example, a mental illness could be adolescent depression. Similarly, the specific type of neurodegenerative disease is not limited in this application; for example, a neurodegenerative disease could be Alzheimer's disease or mild cognitive impairment. The majority class samples refer to the larger number of samples that can be used as controls, while the minority class samples are the target categories that need to be accurately identified. The specific type of control sample is not limited in this application. For example, when determining whether a patient has adolescent depression, the majority class samples are healthy controls, and the minority class samples are adolescent depression patients; when determining whether a patient has Alzheimer's disease or mild cognitive impairment, the majority class samples are Alzheimer's disease patients, and the minority class samples are patients with mild cognitive impairment.

[0095] The following is an example of cases admitted to the neurology department of a hospital in Tianjin: Alzheimer's disease (AD) is a neurodegenerative disease with an insidious onset and progressive development, clinically characterized by comprehensive dementia manifestations such as memory impairment, aphasia, apraxia, and personality changes. With the increasing aging of the global population, the incidence of AD and its precursor stage—mild cognitive impairment (MCI)—is rising year by year. Because the neurological damage in AD is irreversible, accurate identification and intervention at the MCI stage are crucial to slowing disease progression.

[0096] Currently, the main methods for clinical diagnosis of Alzheimer's disease (AD) and Mild Cognitive Impairment (MCI) rely on cerebrospinal fluid (CSF) biomarker detection, positron emission tomography (PET) imaging, and neuropsychological scale assessment. However, CSF extraction requires lumbar puncture, which is highly invasive, leading to poor patient compliance and infection risks. PET scanning equipment is expensive, with high costs per examination, and its availability in primary healthcare institutions is limited by the use of radioactive tracers. Furthermore, cognitive scale tests are highly susceptible to the influence of the subjects' education level and subjective psychology, resulting in significant deficiencies in their sensitivity for early disease identification.

[0097] While attenuated total reflectance-Fourier transform infrared spectroscopy (ATR-FTIR) has shown potential for minimally invasive, rapid, and low-cost plasma testing due to its ability to reflect vibrational information of biomolecules, it still faces key bottlenecks in clinical translation. On the one hand, human serum is complex, and subtle spectral changes caused by Alzheimer's disease (AD) are often masked by complex background signals, making it difficult for traditional linear preprocessing methods to extract deep features. On the other hand, in clinically collected samples, there is often a severe imbalance between the number of healthy controls and patients in the target category, leading to prediction bias in traditional classification algorithms when processing such imbalanced data, significantly limiting diagnostic specificity.

[0098] Example 1: Plasma samples were collected from 275 Alzheimer's disease (AD) patients and 151 patients with mild cognitive impairment (MCI) admitted to the Department of Neurology of a hospital in Tianjin. All subjects fasted for 12-14 hours overnight before collection, and peripheral blood was collected via intravenous puncture between 7:30 and 8:30 am the following morning before breakfast. The collected whole blood samples were immediately transferred to 6 mL anticoagulant tubes containing EDTA, and plasma separation was completed within 2 hours after collection: centrifuged at 3000 rpm for 15 minutes at 4°C. The obtained supernatant plasma was stored at -80°C for subsequent analysis. Spectroscopic data acquisition was performed using a Bruker Alpha ATR-FTIR spectrometer.

[0099] S201. Before analysis, thaw the plasma sample at room temperature, take an appropriate amount and evenly drop it onto the surface of a potassium bromide slide, then allow it to air dry at room temperature for 12 minutes. To ensure the consistency and repeatability of the measurement, an automatic pressurization device is used to apply uniform pressure to the slide to ensure stable sample thickness and contact conditions. Spectral acquisition parameters are set as follows: wavenumber range 4000-600. 4 resolution A total of 32 scans were performed. The diamond crystal was thoroughly cleaned before each sample test, and the background spectrum was collected synchronously after each scan to eliminate systematic errors that may be introduced by changes in environmental humidity and temperature.

[0100] Figure 3 The ATR-FTIR spectrum of an embodiment of this application is shown, such as... Figure 3 As shown, in subsequent data processing, we selected 4000-950. The band is used as the analysis range, and the 950-600 band is discarded. Low-frequency band. This low-frequency region is mainly affected by non-specific interference from residual salts and high-frequency detection noise, resulting in a low signal-to-noise ratio. The selected band covers biospectral characteristics relevant to AD, including the amide I / II band (1700-1500 nm), which reflects abnormal secondary structure of Aβ and p-tau proteins. ), reflecting lipid metabolism disorders, the carbonyl (C=O) stretching vibration band (1800-1700) ) and the hydroxyl (OH) stretching vibration band (3600-3200) related to carbohydrate metabolism. Then, the KS algorithm is used to divide the collected raw spectral samples into the original training set and the test set in a 3:1 ratio.

[0101] S202. Perform MTF image processing on the original training and test set data obtained in S201, and straighten the resulting two-dimensional image into one-dimensional data by row. This is because AD and MCI spectra are extremely similar. Using MTF to convert one-dimensional ATR-FTIR into two dimensions can amplify the subtle differences between the two, and then straighten it into one-dimensional data by row to facilitate subsequent steps.

[0102] Specifically: Divide the x-value range of a single spectral data into Q There are three equally probable quantile intervals, each denoted as . Every moment All can be mapped to their respective... Intervals. The frequency of data points transitioning from one region to another between adjacent time points is statistically analyzed and converted into probabilities, thus yielding the Markov transition matrix. :

[0103] Single-step transition probability ,express The time is located at the The elements of each quantile interval, in t The moment jumps to the first The probability of each quantile interval. The probabilities are recursively derived through matrix exponentiation, and the transition probabilities at any time step are statistically analyzed and concatenated to form the Markov transition field matrix. : Wherein, mapping function Will The index mapped to its corresponding quantile interval, i.e. , express The first mapping quantile intervals Indicates from Corresponding quantile region Transferred to Corresponding quantile region The probability of.

[0104] Specifically, first, the spectral data Single sample data in Transform into a Markov transition field Next, Resampling to the specified size Two-dimensional matrix ; then, Flatten along the row into a row vector .Will n indivual The final image feature matrix is ​​obtained by stacking the rows. .

[0105] S203. Process the obtained image data Perform PCA dimensionality reduction.

[0106] Specifically: For the obtained image matrix Its covariance matrix is Find a set of orthogonal bases This maximizes the variance of the projected samples and ensures that the principal components are uncorrelated. The new sample after projection... The expression is as follows: (3) in It is composed of the covariance matrix The former d A matrix consisting of the eigenvectors corresponding to the largest eigenvalues. d The target dimension. The sample matrix after PCA dimensionality reduction. .

[0107] S204. Perform SMOTE sample augmentation on the minority class MCI samples in the original training set after dimensionality reduction, and combine the original training set with the virtual sample set to obtain the augmented training set. This represents the original training set, which contains majority class samples. and minority class samples . It is by any single sample The generated virtual sample has the following expression: (4) Minority class MCI virtual samples From real samples and nearest neighbor samples linear combination, It is a random number in the interval [0, 1]. Representative generated A virtual sample set consisting of minority class samples. The original training set... and Merging to obtain the augmented training set Test set used It means that among them m and s These are the number of samples in the augmented training set and the test set, respectively.

[0108] Figure 4 This embodiment of the application shows the original training set after PCA dimensionality reduction, the virtual samples generated by SMOTE, and the scatter plot of the test set, as shown in the example. Figure 4 As shown, PC1 (dimension 1) and PC2 (dimension 2) are the two principal component dimensions with the largest eigenvalues ​​after PCA dimensionality reduction. From a distribution perspective, the virtual sample points are tightly clustered within the feature neighborhood of the real MCI samples, enriching the feature diversity of the MCI samples. Simultaneously, the test set samples are completely surrounded by the overall feature distribution of the training set, indicating that the partitioning of the training and test sets in S201 demonstrates good representativeness and consistency, providing crucial assurance for the reliability of subsequent model evaluation.

[0109] S205. Construct the JSMFRC classification model using the augmented training set U obtained in S204.

[0110] Specifically: The objective function for optimizing the JSMFRC classifier is: Where y is the true label of the sample, and w is a weight vector containing a bias term, used to determine the category of the projected sample. represent Norm, express Norm. In the loss function, It is the experience loss term, used to measure the degree of deviation between the predicted result and the true label; and That is, the MFA module, which measures the intra-class compactness and inter-class separability of samples respectively; To smooth the regularization term, the L2 norm of w is used to limit the complexity of the model; For sparse regularization terms, through P Norms are used to impose sparsity constraints, preserving key features while eliminating redundant information. These are the weighting coefficients for the corresponding items.

[0111] Based on the projection matrix P and the weight vector w, the discriminant function for any input t (one-dimensional feature vector) It can be represented as: .

[0112] To verify the superiority of the intelligent diagnostic process of this application, five other comparative models were added: Logistic Regression (LogRegress), Marginal Fisher Analysis (MFA), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), and Linear Graph Embedding (LGE).

[0113] Except for JSMFRC, the other five contrastive classifiers were selected for optimal hyperparameters through five-fold cross-validation: For the MFA classifier, the search ranges for intra-class nearest neighbors and inter-class nearest neighbors were set to [1, 20] and [1, 30], respectively. The optimal values ​​were determined to be 5 and 20 through cross-validation. Specifically, the intra-class value was 20 and the inter-class value was 200 (set as needed). Regularization was enabled with a regularization coefficient of 0.01, and the others were set to default. The KNN classifier uses a nearest neighbor search range of [1, 9], with an optimal value of 5. The nearest neighbor search method is exhaustive search, the distance metric is Minkowski distance, data standardization is enabled, and other settings are default. The SVM classifier uses a Gaussian kernel function, with its kernel parameter search range set to [value missing]. Standardization is enabled, the function type is Gaussian kernel, and the kernel scale parameter is automatically optimized; the dimensionality reduction search range of the LGE classifier is [1, 50], the optimal dimensionality reduction dimension is 30, regularization is enabled, and the regularization coefficient is 0.001.

[0114] In this embodiment of the application, the parameters for the AD / MCI classification process are set as follows: MTF quantile interval number. Q =5. Image size Feature dimensions after PCA dimensionality reduction Number of virtual samples generated by SMOTE The JSMFRC classifier involves four key parameters, among which... The search range is set to , The search range is set to Given the characteristics of virtual samples augmenting the training set, using cross-validation to select the optimal parameters can easily lead to a deterioration in the model's generalization performance on real samples. To minimize the differences in sensitivity and specificity between the original training set (real samples) and the augmented training set (real samples and virtual samples), the following formula is used as the criterion for parameter selection.

[0115] in, and These represent the sensitivity of the original training set and the augmented training set, respectively. and These correspond to the specificities of the two respectively. The optimal formula is obtained from the above equation. The values ​​are 0.1, 0.1, 1, and 10, respectively.

[0116] S206. Input the test set into the classification model established in S205 to obtain the classification performance of the JSMFRC classification model and the comparison model in this embodiment of the application. The results are shown in Table 1. F1 score (F1-Score, ) and subject operating characteristic curves.

[0117] Table 1: Modeling results of different models

[0118] Figure 5 The ROC curves of different models in the embodiments of this application are shown, such as... Figure 5 As shown, comparing the performance of different models, traditional algorithms generally suffer from prediction imbalance. LogRegress, MFA, and KNN models have low accuracy and insufficient specificity; while SVM and LGE models, although possessing extremely high sensitivity (both above 90%), exhibit a sharp drop in specificity (only 5.26% and 21.05%, respectively). This indicates that traditional models are essentially ineffective in identifying negative samples, resulting in severe prediction bias and an inability to balance various performance metrics. In contrast, the JSMFRC algorithm significantly overcomes this performance imbalance. While maintaining the highest accuracy (72.90%), this model achieves an excellent balance between sensitivity (72.46%) and specificity (73.68%). This demonstrates that JSMFRC achieves equally efficient identification of both positive and negative samples.

[0119] In summary, JSMFRC effectively balances various prediction and classification tasks. Its comprehensive metrics, such as AUC (0.73) and F1-Score (0.78), are superior to other models, demonstrating that this application exhibits higher stability and prediction accuracy in AD / MCI classification tasks.

[0120] Example 2: A classification model was constructed using samples of adolescents with depression and healthy controls, and its effectiveness was validated. The method and steps were exactly the same as in Example 1, except that the model training parameters were adjusted according to the characteristics of samples from different diseases.

[0121] Adolescent depression primarily detects small molecule spectral signals such as neurotransmitter metabolites. These signals have relatively concentrated characteristic peaks but exhibit significant individual variability, requiring shallow tree depth and high learning rates to capture common patterns and suppress individual noise. In contrast, Alzheimer's disease mainly detects macromolecular structural signals such as Aβ protein conformational changes. These signals have high dimensionality, are weak, and exhibit long-term nonlinear correlations with disease progression, necessitating increased tree depth and number to fit complex protein aggregation processes. The basic principle for parameter selection is "molecular features determine model strategy." For small molecule metabolites, the focus is on rapid convergence and generalization; for protein conformational changes, the focus is on deep fitting and stability. Finally, hierarchical cross-validation is used to achieve the optimal parameter configuration specific to the disease. The parameters are shown in Table 2.

[0122] Table 2. Model parameters for adolescent depression example

[0123] The test results of the embodiment of the present invention for adolescent depression are shown in Table 3: Table 3. Results of Implementation Tests for Adolescent Depression

[0124] Based on the same inventive concept, this application also provides a disease classification model construction device, as described in the following embodiments. Since the principle of this device embodiment in solving the problem is similar to that of the above method embodiments, the implementation of this device embodiment can refer to the implementation of the above method embodiments, and repeated details will not be repeated.

[0125] Figure 6 This illustration shows a schematic diagram of a disease classification model construction device according to an embodiment of this application, such as... Figure 6 As shown, the device includes: Sample data acquisition and preprocessing module 601: used to acquire plasma ATR-FTIR spectra; to perform band screening on plasma ATR-FTIR spectra based on signal-to-noise ratio and biospectral characteristics to obtain an original spectral sample set; to divide the original spectral sample set into an original training set and a test set using the KS algorithm; wherein, the samples include majority class samples and minority class samples; Image processing module 602: used to perform image processing on the original training set and the test set using the Markov transfer field method to obtain a two-dimensional matrix; and to convert the two-dimensional matrix into a one-dimensional feature vector; Data dimensionality reduction module 603: used to perform dimensionality reduction processing on the one-dimensional feature vector using principal component analysis to obtain a dimensionality reduction training set and a dimensionality reduction test set; Sample augmentation module 604: Obtain minority class samples from the dimensionality reduction training set, use synthetic minority oversampling technology to augment the minority class samples in the dimensionality reduction training set to obtain a virtual sample set; combine the dimensionality reduction training set and the virtual sample set to obtain the augmented training set; Model building module 605: Based on the augmented training set, a classification model is constructed using the joint sparse boundary Fisher regularization classification algorithm; and Model Validation Module 606: Used to validate the performance of the classification model based on the test set, and to qualitatively analyze the classification effect of the model using classification evaluation metrics; based on the augmented training set, the model is input into the augmented training set for prediction, and evaluation metrics are calculated based on the confusion matrix to qualitatively analyze the classification effect and generalization ability of the classification model.

[0126] It should be noted that the above-described sample data acquisition and preprocessing module 601, image processing module 602, data dimensionality reduction module 603, sample augmentation module 604, model building module 605, and model validation module 606 correspond to S201-S206 in the method embodiment. These modules and their corresponding steps implement the same examples and application scenarios, but are not limited to the content disclosed in the above method embodiment. It should also be noted that these modules, as part of the apparatus, can be executed in a computer system such as a set of computer-executable instructions.

[0127] Those skilled in the art will understand that various aspects of this application can be implemented as a system, method, or program product. Therefore, various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, collectively referred to herein as a "circuit," "module," or "system."

[0128] The following reference Figure 7 To describe an electronic device 700 according to this embodiment of the present application. Figure 7 The electronic device 700 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0129] like Figure 7 As shown, the electronic device 700 is manifested in the form of a general-purpose computing device. The components of the electronic device 700 may include, but are not limited to: at least one processing unit 710, at least one storage unit 720, and a bus 730 connecting different system components (including storage unit 720 and processing unit 710).

[0130] The storage unit stores program code that can be executed by the processing unit 710, causing the processing unit 710 to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this application. For example, the processing unit 710 can perform steps S201 to S206 of the above method embodiments.

[0131] Storage unit 720 may include readable media in the form of volatile storage units, such as random access memory (RAM) 7201 and / or cache memory 7202, and may further include read-only memory (ROM) 7203.

[0132] The storage unit 720 may also include a program / utility 7204 having a set (at least one) program module 7205, such program module 7205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0133] Bus 730 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0134] Electronic device 700 can also communicate with one or more external devices 740 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 700, and / or with any device that enables electronic device 700 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 750. Furthermore, electronic device 700 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 760. As shown, network adapter 760 communicates with other modules of electronic device 700 via bus 730. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0135] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, terminal device, or network device, etc.) to execute the methods according to the embodiments of this application.

[0136] In particular, according to embodiments of this application, the process described in the above-mentioned flowchart can be implemented as a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for constructing a disease classification model.

[0137] In an exemplary embodiment of this application, a computer-readable storage medium is also provided, which may be a readable signal medium or a readable storage medium. The computer-readable storage medium stores a program product capable of implementing the methods described above in this application.

[0138] In some possible implementations, various aspects of this application may also be implemented as a program product comprising program code that, when the program product is run on a terminal device, causes the terminal device to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this application.

[0139] More specific examples of computer-readable storage media in this application may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0140] In this application, a computer-readable storage medium may include a data signal propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The readable signal medium may also be any readable medium other than a readable storage medium, capable of transmitting, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0141] Optionally, the program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0142] In practical implementation, program code for performing the operations of this application can be written using any combination of one or more programming languages. These programming languages ​​include object-oriented programming languages—such as Java and C++—and conventional procedural programming languages—such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0143] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0144] Furthermore, although the steps of the method in this application are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or a step may be broken down into multiple steps.

[0145] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, mobile terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0146] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the appended claims.

Claims

1. A method for constructing a disease classification model, characterized in that, include: Collect plasma ATR-FTIR spectra; Based on signal-to-noise ratio and biospectral characteristics, band selection was performed on plasma ATR-FTIR spectra to obtain an original spectral sample set; the KS algorithm was used to divide the original spectral sample set into an original training set and a test set; wherein, the samples include majority class samples and minority class samples; The original training set and the test set are visualized using the Markov transfer field method to obtain a two-dimensional matrix; the two-dimensional matrix is ​​then converted into a one-dimensional feature vector. Principal component analysis is used to reduce the dimensionality of the one-dimensional feature vector to obtain a dimensionality-reduced training set and a dimensionality-reduced test set. Obtain minority class samples from the dimensionality-reduced training set, and use synthetic minority oversampling technology to augment the minority class samples in the dimensionality-reduced training set to obtain a virtual sample set; combine the dimensionality-reduced training set and the virtual sample set to obtain the augmented training set; Based on the augmented training set, a classification model is constructed using the joint sparse boundary Fisher regularization classification algorithm. The performance of the classification model is verified based on the test set, and the classification effect of the model is qualitatively analyzed using classification evaluation metrics. Based on the augmented training set, the classification model is predicted by inputting the augmented training set into the classification model, and the evaluation metrics are calculated based on the confusion matrix to qualitatively analyze the classification effect and generalization ability of the classification model.

2. The method for constructing a disease classification model according to claim 1, characterized in that, The original training set and the test set are visualized using the Markov transfer field method to obtain a two-dimensional matrix; the two-dimensional matrix is ​​then converted into a one-dimensional feature vector, including: Based on the original training set and the test set, the value range of a single spectral data x is divided into... Q There are quantile intervals with equal probability, and the spectral data of a single sample is denoted as . x 1, x 2,…, x D Each quantile interval is denoted as Every moment All can be mapped to their respective... interval; Construct the Markov transition matrix , ， in, Let be the single-step transition probability, representing The time is located at the The elements of each quantile interval, in t The moment jumps to the first The probability of each quantile interval; Constructing Markov transition fields , ; Wherein, mapping function Will The index mapped to its corresponding quantile interval, i.e. , express The first mapping quantile intervals Indicates from Corresponding quantile region Transferred to Corresponding quantile region The probability of; spectral data Single sample data in Transform into a Markov transition field Next, Resampling to the specified size Two-dimensional matrix ; Will Flatten along the row into a row vector ; Will n indivual The final image feature matrix is ​​obtained by stacking the rows. .

3. The method for constructing a disease classification model according to claim 2, characterized in that, Principal component analysis is used to reduce the dimensionality of the one-dimensional feature vector, resulting in a dimensionality-reduced training set and a dimensionality-reduced test set, including: for the image-based feature matrix Its covariance matrix is Find a set of orthogonal bases This maximizes the variance of the projected samples and ensures that the principal components are uncorrelated. The resulting new samples after projection... The expression is ,in, It is composed of the covariance matrix The former d A matrix consisting of the eigenvectors corresponding to the largest eigenvalues. d The target dimension is the sample matrix after dimensionality reduction. .

4. The method for constructing a disease classification model according to claim 3, characterized in that, Minority class samples are obtained from the dimensionality-reduced training set, and synthetic minority oversampling technique is used to augment the minority class samples in the dimensionality-reduced training set to obtain a virtual sample set. The augmented training set is obtained by combining the dimensionality-reduced training set with the virtual sample set, including: Dimensionality reduction training set is ,in, n trn The total number of samples in the dimensionality reduction training set. d The feature dimensions after dimensionality reduction for PCA; the dimensionality-reduced training set includes majority class samples. and minority class samples ; Constructing minority class virtual samples , Among them, a minority of virtual samples From real samples and K nearest neighbor samples linear combination, It is a random number in the interval [0, 1]; construct a virtual sample set. ; the original training set and Merging to obtain the augmented training set .

5. The method for constructing a disease classification model according to claim 4, characterized in that, The objective function for the classification model is: Where y is the true label of the sample, and w is a weight vector containing a bias term, used to determine the category of the projected sample. Let be the projection matrix. represent Norm, express Norm, It is the experience loss term, used to measure the degree of deviation between the predicted result and the true label; and The MFA module measures intra-class compactness and inter-class separability of samples, respectively. To smooth the regularization term, the L2 norm of w is used to limit the complexity of the model; For sparse regularization terms, through P Norms are used to impose sparsity constraints, preserving key features while eliminating redundant information; These are the weighting coefficients for the corresponding items; L w Let L be the inner Laplace matrix. b The inter-class Laplace matrix; The discriminant function of the classification model for , where t is a one-dimensional eigenvector, P is a projection matrix, and w is a weight vector.

6. The method for constructing a disease classification model according to claim 1, characterized in that, The classification model is used to classify mental illnesses; the majority class samples are healthy control samples, and the minority class samples are patient samples of mental illnesses.

7. The method for constructing a disease classification model according to claim 1, characterized in that, The classification model is used to classify neurodegenerative diseases; the majority class samples are samples of patients with Alzheimer's disease, and the minority class samples are samples of patients with mild cognitive impairment.

8. A disease classification model construction device, characterized in that, include: Sample data acquisition and preprocessing module: used to acquire ATR-FTIR spectra of plasma samples; Based on signal-to-noise ratio and biospectral characteristics, band selection was performed on plasma ATR-FTIR spectra to obtain an original spectral sample set; the KS algorithm was used to divide the original spectral sample set into an original training set and a test set; wherein, the samples include majority class samples and minority class samples; Image processing module: used to image the original training set and the test set using the Markov transfer field method to obtain a two-dimensional matrix; and to convert the two-dimensional matrix into a one-dimensional feature vector; Data dimensionality reduction module: used to perform dimensionality reduction processing on the one-dimensional feature vector using principal component analysis to obtain a dimensionality reduction training set and a dimensionality reduction test set; Sample augmentation module: Obtain minority class samples from the dimensionality reduction training set, use synthetic minority oversampling technology to augment the minority class samples in the dimensionality reduction training set to obtain a virtual sample set; combine the dimensionality reduction training set and the virtual sample set to obtain the augmented training set; Model building module: Based on the augmented training set, a classification model is built using the joint sparse boundary Fisher regularization classification algorithm; and Model Validation Module: Used to validate the performance of the classification model based on the test set, and to qualitatively analyze the classification effect of the model using classification evaluation metrics; based on the augmented training set, the model is input into the augmented training set for prediction, and evaluation metrics are calculated based on the confusion matrix to qualitatively analyze the classification effect and generalization ability of the classification model.

9. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the disease classification model construction method according to any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a method for constructing a disease classification model as described in any one of claims 1 to 7.

Images