Hierarchical explainable classification method, system and device for heterogeneous diseases

Abstract

This invention discloses a hierarchical interpretable classification method, system, and device for heterogeneous diseases, belonging to the field of artificial intelligence-assisted diagnostic technology. This invention maps multi-source heterogeneous data to a unified latent feature space. Through orthogonal constraints and pathological consistency constraints, features are decomposed into independent genotype contributions, phenotypic contributions, and gene-phenotype interaction contributions, achieving decoupling of pathological features. Contrastive learning and hard sample mining mechanisms are employed to strengthen key features, ensuring that sample distribution conforms to the pathological evolution pattern. Hierarchical and accurate discrimination is achieved through two-level cascaded classification, and the results of fine classification are fed back to correct coarse classification features, forming a hierarchical decision-making closed loop. A structured and interpretable diagnostic report is generated based on the orthogonal contribution degree of pathology. This invention achieves multimodal data decoupling and fusion, accurate distinction of fuzzy categories, hierarchical decision-making closed-loop optimization, and fully interpretable output, effectively improving the stability, discrimination accuracy, and clinical interpretability of heterogeneous disease classification.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence-assisted diagnostic technology, and specifically discloses a hierarchical interpretable classification method, system and device for heterogeneous diseases. Background Technology

[0002] Neurodegenerative diseases such as Alzheimer's disease are characterized by high heterogeneity, continuous disease course, and blurred category boundaries. Clinically, mild cognitive impairment (MCI) and cognitively normal (NC) samples exhibit highly overlapping characteristics, making them difficult to distinguish using traditional models. Existing multimodal fusion methods often employ simple splicing or weighted summation, failing to decouple genotype, phenotype, and interaction contributions, leading to information aliasing and poor interpretability. Furthermore, two-level classification models are often unidirectional pipelines, lacking feedback optimization of coarse classification features from finer classification results, making adaptive high-precision classification difficult. In addition, existing diagnostic models cannot provide traceable decision-making evidence at the pathological level, failing to meet clinical reliability and compliance requirements. Summary of the Invention

[0003] To address the aforementioned issues, namely the shortcomings of existing technologies such as information overlap in multi-source data fusion, weak ability to distinguish fuzzy categories, lack of feedback in hierarchical decision-making, and uninterpretable diagnostic results, this invention provides a hierarchical interpretable classification method, system, and device for heterogeneous diseases. Based on a unified framework of pathological orthogonal decoupling, hyperspherical normalization embedding, two-level closed-loop feedback, and contribution path tracing, it achieves high-precision, highly interpretable, and highly robust classification of heterogeneous diseases.

[0004] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides a hierarchical interpretable classification method for heterogeneous diseases, the method comprising: Obtain multi-source heterogeneous data of the object to be classified, and map the data from different sources to a unified latent feature space; apply orthogonal constraints and pathological consistency constraints to the unified latent feature space, decompose the unified latent feature space into mutually orthogonal genotype contribution features, phenotypic contribution features and gene-phenotype interaction contribution features, obtain a unified feature representation of pathological orthogonal decoupling, and output the pathological orthogonal contribution weight matrix. For class pairs with overlapping boundaries in the unified latent feature space, the unified feature representation of pathological orthogonal decoupling is enhanced based on contrast constraints and combined with hard sample mining, and the enhanced features are mapped to the normalized embedding space to form a sample distribution that conforms to the continuity of pathology. A hierarchical classification process is adopted using a two-level cascaded classification process. The first-level classification is based on the unified feature representation of the pathological orthogonal decoupling, and the second-level classification is based on the enhanced features in the normalized embedding space. Feedback information is generated based on the discriminant information of the second-level classification and sent back to the unified feature representation of the pathological orthogonal decoupling to update the first-level classification results. Based on the pathological orthogonal contribution weight matrix, the feature contributions in the classification process are tracked, and the key feature information corresponding to the genotype contribution feature, phenotype contribution feature and gene-phenotype interaction contribution feature is output respectively, generating a structured and interpretable diagnostic report.

[0006] Optionally, obtaining multi-source heterogeneous data of the object to be classified includes: extracting genetic interaction features from the multi-source heterogeneous data as genotype data, extracting body fluid biomarkers as phenotypic data, and performing dimensional alignment and numerical standardization on the genotype data and the phenotypic data.

[0007] Optionally, applying orthogonal constraints to the latent feature space includes: calculating the inner product values ​​between the genotype contribution features, the phenotype contribution features, and the gene-phenotype interaction contribution features, respectively; adding the inner product values ​​to the loss function and performing a minimization process to keep the various contribution features independent of each other.

[0008] Optionally, the output pathological orthogonal contribution weight matrix includes: splitting the pathological orthogonal contribution weight matrix into three sub-weights, respectively weighting and mapping the phenotypic contribution features, the genotype contribution features, and the gene-phenotype interaction contribution features, and adjusting the numerical proportions of each type of feature according to preset pathological rules.

[0009] Optionally, applying pathological consistency constraints to the latent feature space includes: setting corresponding feature proportion rules according to different disease categories, and adjusting the numerical values ​​of various contributing features according to the sample category labels during training to make the feature distribution match the pathological stage.

[0010] Optionally, mapping the enhanced features to the normalized embedding space includes: performing normalized projection processing on the enhanced features, mapping the enhanced features to the same scale space, and adjusting the feature positions of samples with overlapping boundaries, so that samples of the same pathological stage form a continuous distribution and samples of different categories remain separated.

[0011] Optionally, generating feedback information and sending it back to the unified feature representation of pathological orthogonal decoupling includes: extracting classification confidence and feature offset from the secondary classification results, combining the confidence and the offset into a feedback signal, and using the feedback signal to correct the unified feature representation of pathological orthogonal decoupling; The two-level cascaded classification process includes: first, dividing all samples into two major categories through primary classification; then, selecting samples with blurred boundaries from the major categories to enter secondary classification; and finally, sending back the category determination results obtained from secondary classification to update the major category division results of primary classification.

[0012] Optionally, the step of tracking feature contributions during the classification process includes: backtracking the source of features along the pathological orthogonal contribution weight matrix, sequentially locating key genotype contribution features, key phenotype contribution features, and key gene-phenotype interaction contribution features, and associating the source of features with the classification results to generate the explanatory content in the structured interpretable diagnostic report.

[0013] Secondly, the present invention provides a hierarchical and interpretable classification system for heterogeneous diseases, comprising: An interpretable fusion coding module is used to acquire multi-source heterogeneous data of the object to be classified, and map data from different sources to a unified latent feature space; orthogonal constraints and pathological consistency constraints are applied to the unified latent feature space, and the unified latent feature space is decomposed into mutually orthogonal genotype contribution features, phenotypic contribution features and gene-phenotype interaction contribution features, to obtain a unified feature representation of pathological orthogonal decoupling, and output the pathological orthogonal contribution weight matrix. The feature decoupling enhancement module is used to enhance the unified feature representation of pathological orthogonal decoupling based on contrast constraints and combined with hard sample mining for class pairs with overlapping boundaries in the unified latent feature space, and to map the enhanced features to the normalized embedding space to form a sample distribution that conforms to the pathological continuity. A two-level cascaded decision module is used to perform hierarchical classification using a two-level cascaded classification process, wherein first-level classification is performed based on the unified feature representation of pathological orthogonal decoupling, and second-level classification is performed based on the enhanced features in the normalized embedding space; feedback information is generated based on the discriminant information of the second-level classification and sent back to the unified feature representation of pathological orthogonal decoupling to update the first-level classification result; The diagnostic report generation module is used to track the feature contributions in the classification process based on the pathological orthogonal contribution weight matrix, and output the key feature information corresponding to the genotype contribution feature, phenotype contribution feature and gene-phenotype interaction contribution feature respectively, and generate a structured and interpretable diagnostic report.

[0014] Thirdly, the present invention provides an electronic device, the electronic device comprising: At least one processor; and a memory communicatively connected to said at least one processor; The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method described in any one of the first aspects.

[0015] Compared with the closest existing technology, the present invention has the following advantages: This application proposes a hierarchical, interpretable classification method, system, and device for heterogeneous diseases. It forms a closed-loop technology process from data fusion, feature enhancement, hierarchical decision-making to interpretable output, offering several significant advantages over existing technologies. This invention applies orthogonal constraints and pathological consistency constraints to a unified latent feature space, decomposing multi-source heterogeneous data into independent genotype contribution features, phenotypic contribution features, and gene-phenotype interaction contribution features. This fundamentally solves the problems of information aliasing and indistinguishable contributions in traditional multimodal fusion, making feature representations more closely aligned with pathological mechanisms and more discriminative, effectively improving the reliability and stability of subsequent classification.

[0016] This invention targets difficult-to-distinguish samples with overlapping boundaries. It employs contrastive constraints combined with difficult sample mining for feature enhancement and maps the features to a normalized embedding space to form a sample distribution that conforms to pathological continuity. This significantly amplifies the subtle differences between ambiguous categories, greatly improving the accuracy of distinguishing between mild cognitive impairment and transitional states such as normal cognition, and effectively overcoming the shortcomings of existing technologies in identifying intermediate disease stages.

[0017] This invention adopts a two-level cascaded classification mode and introduces a feedback correction mechanism from the second-level classification to the first-level classification. It can dynamically optimize the features used in the first-level classification based on the discriminant information of the sub-classification, so that the feature representation and the hierarchical task are adaptively matched to form a closed-loop decision structure, which significantly improves the overall classification robustness. It can still maintain stable output in scenarios with small samples and imbalanced data, and effectively reduce the false positive rate and false negative rate.

[0018] Meanwhile, this invention tracks feature contributions throughout the entire process based on a pathological orthogonal contribution weight matrix, which can directly locate key genetic features, phenotypic features, and interaction features, automatically generate structured and interpretable diagnostic reports, clearly quantify the classification criteria, make the diagnostic process transparent and traceable, and have clinical credibility, thus breaking through the bottleneck of traditional black box models that are uninterpretable and difficult to implement.

[0019] The overall solution features coordinated and progressive support across all stages, achieving a balance between high-precision classification and high interpretability. It possesses strong versatility and practicality, making it widely applicable to various intelligent assisted diagnostic scenarios for heterogeneous diseases. It effectively improves diagnostic efficiency and standardization, demonstrating significant technological advancements and practical application value. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0021] Figure 1 This is a flowchart of a hierarchical and interpretable classification method for heterogeneous diseases provided by the present invention; Figure 2 This is a schematic diagram of dynamic hard sample mining and hyperspherical constraint provided by the present invention; Figure 3 This is a schematic diagram of the two-level cascaded decision-making process provided by the present invention; Figure 4 This is a schematic diagram of a hierarchical interpretable classification system for heterogeneous diseases provided by the present invention; Figure 5 This is the overall architecture diagram of the hierarchical interpretable classification system provided by the present invention; Figure 6 This is a diagram of the internal structure of the electronic device provided by the present invention. Detailed Implementation

[0022] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention.

[0023] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0024] This invention provides a hierarchical and interpretable classification method, system, and device for heterogeneous diseases. The embodiments of this invention are described below with reference to the accompanying drawings.

[0025] Example 1: As Figure 1 As shown, Embodiment 1 of the present invention provides a hierarchical interpretable classification method for heterogeneous diseases, which specifically includes the following steps: S101 acquires multi-source heterogeneous data of the object to be classified, maps data from different sources to a unified latent feature space; applies orthogonal constraints and pathological consistency constraints to the unified latent feature space, decomposes the unified latent feature space into mutually orthogonal genotype contribution features, phenotypic contribution features and gene-phenotype interaction contribution features, obtains a unified feature representation of pathological orthogonal decoupling, and outputs a pathological orthogonal contribution weight matrix. S102 For class pairs with overlapping boundaries in the unified latent feature space, enhance the unified feature representation of the pathological orthogonal decoupling based on contrast constraints and combined with hard sample mining, and map the enhanced features to the normalized embedding space to form a sample distribution that conforms to the pathological continuity. S103 employs a two-level cascaded classification process for hierarchical classification, wherein primary classification is performed based on the unified feature representation of pathological orthogonal decoupling, and secondary classification is performed based on the enhanced features in the normalized embedding space; feedback information is generated based on the discriminant information of the secondary classification and sent back to the unified feature representation of pathological orthogonal decoupling to update the primary classification result; S104 tracks the feature contributions during the classification process based on the pathological orthogonal contribution weight matrix, and outputs the key feature information corresponding to genotype contribution features, phenotype contribution features, and gene-phenotype interaction contribution features, respectively, to generate a structured and interpretable diagnostic report.

[0026] In the above steps, obtaining multi-source heterogeneous data of the object to be classified includes: extracting genetic interaction features from genetic data as genotype data, and extracting biomarkers from body fluid detection data as phenotypic data; and performing sample ID matching, missing value imputation, dimension alignment, and numerical standardization on the genotype data and the phenotypic data.

[0027] Furthermore, the multi-source heterogeneous data includes at least: Genotype data: Extracting genetic interaction features from genetic data; Phenotypic data: Extracting body fluid biomarkers from body fluid test data.

[0028] Perform dimensional alignment and numerical standardization on genotype and phenotypic data. For example: Dimension alignment: Use a unified sample ID for matching different modalities, and use a missing mask or mean / model imputation for missing modalities; Standardization: Z-score standardization or min-max standardization is used for continuous variables, and one-hot encoding / frequency encoding is used for discrete variables.

[0029] Mapping data from different sources to a unified latent feature space includes: constructing a unified feature representation as follows: Z=Z p ⊙W p +(Z g Z c )⊙W c ; In the formula, Z p Z represents the phenotypic contribution feature. g Z represents the genotype contribution characteristic. cW represents the interaction contribution feature. p W represents the phenotypic subweight. c Indicates the interaction sub-weight, and ⊙ indicates the pathological perception weighting. This indicates interactive projection operation.

[0030] To ensure that the three are independent of each other, orthogonal constraints are applied: Genotype contribution features are independent of phenotypic contribution features, genotype contribution features are independent of gene-phenotype interaction contribution features, and phenotypic contribution features are independent of gene-phenotype interaction contribution features. Constructing the pathological orthogonal loss: L op = ||Zg T ·Zp||+||Zg T ·Zc||+||Zp T ·Zc|| The superscript T indicates the transpose operation of a vector; The training objective is to minimize L op This fully decouples the three types of contribution characteristics.

[0031] Applying orthogonal constraints and pathological consistency constraints to the space, it is decomposed into mutually orthogonal genotype contribution features Z. g Phenotypic contribution characteristics Z p Gene-phenotype interaction contribution feature Z c This yields a unified feature representation of pathological orthogonal decoupling and outputs a pathological orthogonal contribution weight matrix.

[0032] Apply pathological proportion constraints to different categories: AD Sample: Z c >Z g +Z p ; MCI Sample: Z g >Z c ; NC Sample: Z p maximum; The final output is a unified feature representation of pathological orthogonal decoupling and a weight matrix of pathological orthogonal contribution.

[0033] In step S101, applying orthogonal constraints to the latent feature space includes: calculating the inner product values ​​between the genotype contribution features, the phenotype contribution features, and the gene-phenotype interaction contribution features, respectively; adding the inner product values ​​to the loss function and performing minimization processing to keep the various contribution features independent of each other.

[0034] Imposing pathological consistency constraints on the latent feature space includes: setting corresponding feature proportion rules according to different disease categories, and adjusting the numerical values ​​of various contributing features according to the sample category labels, so that the feature distribution matches the pathological stage.

[0035] The output pathological orthogonal contribution weight matrix includes: splitting the pathological orthogonal contribution weight matrix into three sub-weights, respectively weighting and mapping the phenotypic contribution features, the genotype contribution features, and the gene-phenotype interaction contribution features, and adjusting the numerical proportion of each type of feature according to preset pathological rules.

[0036] In step S102, mapping the enhanced features to the normalized embedding space includes: performing normalized projection processing on the enhanced features, mapping the enhanced features to the same scale space, and adjusting the feature positions of samples with overlapping boundaries so that samples of the same pathological stage form a continuous distribution and samples of different categories remain separated.

[0037] The contrast constraint employs InfoNCE loss based on cosine similarity, defined as: ; The hyperspherical constraint terms are: ; in, sim The cosine similarity is represented by τ; τ is the temperature coefficient. Embed the vector for the anchor point sample; z is the embedding vector of positive samples of the same type; j This is the embedding vector for samples within the batch; The cosine similarity function between feature vectors u and v is: ; τ >0 represents the temperature coefficient; λ 3>0 is the balance coefficient; It is an L2 norm.

[0038] Joint topological loss: L total =L nce +λ 1 L inter +λ 2 L ortho ; L inter To calculate the inter-class separability loss, the average distance between the centers of different classes is calculated. L ortho The loss is due to orthogonal constraints. λ 1∈[0.1,10]、 λ2∈[0.01,1] is the balancing hyperparameter, which is set to 1.0 and 0.1 respectively in this embodiment.

[0039] In step S103, generating feedback information and sending it back to the unified feature representation of pathological orthogonal decoupling includes: extracting classification confidence and sample-to-class center feature offset from the secondary classification results, weighting the confidence and offset to form a feedback signal, and using the feedback signal to numerically correct the unified feature representation of pathological orthogonal decoupling.

[0040] The feedback correction vector is calculated as follows: ; Among them, W correction This represents the feedback correction weight matrix. This represents the feature confidence vector output by the secondary classification. This represents the probability value that the current sample is classified as MCI (mild cognitive impairment) by the model output. The probability value for the current sample being classified as NC (Cognitive Normal) by the model output; both are posterior probabilities output by the secondary classification (MCI vs. NC).

[0041] The corrected feature is represented as: Z new =Z+ΔF; Furthermore, such as Figure 2 The two-level cascaded classification process includes: firstly, dividing all samples into two major categories, AD and non-AD, through primary classification; then, selecting MCI / NC samples with ambiguous boundaries from the non-AD category to enter secondary classification; and finally, transmitting the secondary classification results and confidence scores back to update the primary classification results.

[0042] In step S104, the tracking of feature contributions during the classification process includes: backtracking the source of features along the pathological orthogonal contribution weight matrix, locating key genotype contribution features, key phenotype contribution features, and key gene-phenotype interaction contribution features in descending order of absolute contribution value, and associating the feature source, contribution value, and classification result to generate the explanatory content in the structured interpretable diagnostic report.

[0043] Example 2: Based on the above inventive concept, Example 2 proposes an application for early diagnosis of Alzheimer's disease based on a public disease database, using a hierarchical interpretable classification method for heterogeneous diseases. This embodiment demonstrates how the hierarchical interpretable decision system of the present invention can be applied to the early diagnosis and state classification of Alzheimer's disease (AD). The various stages will be described in detail below.

[0044] I. Data Preparation and Preprocessing Data source: This example uses subject data from multiple phases in the Alzheimer's Disease Neuroimaging Project (ADNI) public dataset.

[0045] 1. Genotype data processing: Genotyping data is obtained through mainstream genotyping arrays, covering SNP information across the entire genome. Based on the quality control principles of genotyping data, eligible SNP loci and subjects with complete genotyping data and cerebrospinal fluid (CSF) test data are selected. Their clinical diagnoses include three categories: Alzheimer's disease (AD), mild cognitive impairment (MCI), and normal cognition (NC).

[0046] Genome-wide interaction search (GWIS) was used to detect significant SNP-SNP interactions across multiple CSF quantitative phenotypes. A statistical significance threshold of p ≤ 10 was set. 5 Finally, SNP-SNP pairs with significant interaction effects were selected and used as genotype feature inputs X. G .

[0047] 2. Tabular data processing: 1) Extract Aβ from CSF 42 The concentration values ​​of the three core biomarkers, T-tau and P-tau, were used as the original phenotypic data.

[0048] 2) First, perform Min-Max normalization on the data to compress all values ​​to the range [0, 100].

[0049] 3) Secondly, due to Aβ 42 The pathological trend of Aβ (decreased in AD patients) is opposite to that of T-tau and P-tau (increased in AD patients). To maintain pathological consistency, the normalized Aβ... 42 Data is reverse-calibrated: Aβ 42 ′ =100 Aβ 42 .

[0050] 4) Finally, the standardized phenotypic data matrix X is obtained. P .

[0051] II. Characteristic Fusion of Pathological Orthogonal Decoupling Map the features to a unified latent feature space, and construct: Z=Z p ⊙W p +(Z g Z c )⊙W c ; And minimize pathological orthogonal loss: Constructing the pathological orthogonal loss: L op = ||Zg T ·Zp||+||Zg T ·Zc||+||Zp T ·Zc||.

[0052] Loss function: To optimize the weight matrix W, the following joint loss function is designed. L total : L total =L intra +λ 1 L inter +λ 2 L ortho ; L intra Intra-class compactness loss: Calculates the average Euclidean distance between features of the same class of samples.

[0053] L inter Inter-class separability loss: Calculates the average distance between different class centers.

[0054] L ortho Orthogonal constraint loss penalizes the correlation between row vectors of W, ensuring the independence of different SNPs in representing information; λ 1∈[0.1,10]、 λ 2∈[0.01,1] are the balancing hyperparameters, which are set to 1.0 and 0.1 respectively in the example. The L-BFGS optimizer is used to optimize the parameters. L total Perform a minimization solution.

[0055] L-BFGS can effectively approximate the Hessian matrix and converges faster and more stably than first-order optimizers when dealing with medium-sized parameter optimization problems. The iteration stopping condition is: the loss change is less than 10. -6 Or the number of iterations reaches 100.

[0056] Pathological consistency constraint: Set corresponding feature proportion rules according to different disease categories (AD, MCI, NC). During the training process, adjust the numerical values ​​of genotype, phenotype, and interaction contribution features according to the sample category label so that the feature distribution matches the pathological stage (e.g., the phenotype contribution weight is higher in the AD stage).

[0057] After training, the pathological orthogonal contribution weight matrix W is output and split into three sub-weights, which are used to perform weighted mapping on the phenotypic contribution features, the genotype contribution features, and the gene-phenotype interaction contribution features, respectively.

[0058] Interpretable output: After training, the pathological orthogonal contribution weight matrix W is fixed. By analyzing the elements with larger absolute values ​​in W, the contribution to specific pathological phenotypes (such as P-tau / Aβ) can be identified. 42 The increased ratio of SNP-SNP interactions provides clues for biological mechanism research. In the feature space before and after fusion, the inter-class distance increases and the intra-class distance becomes more compact.

[0059] III. Feature Enhancement Based on Contrast Constraints and Hard Sample Mining For the category pair MCI and NC with overlapping boundaries, the sample features labeled MCI and NC from the fusion feature Z obtained in the above steps are input.

[0060] Using a 3-layer projection network f θ The fused feature Z is mapped to a 128-dimensional embedding space for contrastive learning. In each training batch, positive samples and difficult negative samples (hard sample mining) are dynamically selected for each "anchor" sample. Loss function... The definition is as follows: + ; in, Indicates the first i Each sample (referred to as an "anchor sample") is output through dynamic hard sample mining and hyperspherical constraints. d dimensional latent feature vectors, such as Figure 3 . Indicates the anchor point sample i The feature vector of "positive samples" belonging to the same disease category (e.g., all are mild cognitive impairment, MCI). This represents the th element in a training batch, excluding the anchor point itself. j The feature vector of each sample contains positive samples and all other "negative samples" belonging to different categories.

[0061] For feature vectors u and v The cosine similarity function between them is used to measure the degree of alignment between them in terms of direction. τ =0.1∈(0,1]: Temperature coefficient, used to control the sharpness of similarity distribution; λ 3=0.5∈[0.1,5]: Balance coefficient, used to adjust the weight of the hyperspherical constraint term in the total loss.

[0062] Let L2 be the L2 norm (Euclidean length) of the vector.

[0063] The first term is the contrast loss, which brings similar samples closer together and dissimilar samples further apart; the second term is the hyperspherical constraint, which projects the feature vectors onto a unit sphere.

[0064] After processing, the enhanced features are mapped to the normalized embedding space, and samples of the same pathological stage (such as MCI) form a continuous distribution, while different categories (MCI and NC) remain separated, and the decision boundary is clear.

[0065] IV. Two-level cascaded classification and feedback correction First-level classification: Based on the pathological orthogonal decoupled unified feature representation Z obtained from the above steps, a multilayer perceptron (MLP) is used for binary classification: AD vs. non-AD (including MCI and NC). The MLP structure is as follows: Input layer 256 128 The output layer uses ReLU as the activation function and is trained with DropPath regularization and Nesterov momentum optimizer. If PAD ≥ 0.5, it is considered AD and the process terminates; otherwise, it is considered non-AD and proceeds to the second level of fine classification.

[0066] Second-level classification: Based on the enhanced features in the normalized embedding space obtained in the above steps, a contrastive learning network (reusing the projection network) is used for classification. f θ (and linear classifiers) perform fine-grained classification of non-AD samples using MCI and NC.

[0067] Feedback correction: Extract classification confidence and feature offset (such as the distance difference between the current sample and the MCI / NC class center) from the secondary classification results, combine the confidence and offset into a feedback signal, and feed it back to the unified feature representation Z of the pathological orthogonal decoupling, perform numerical correction on Z (such as weighted adjustment), and update the first-level major category classification results.

[0068] V. Generation of Interpretable Diagnostic Reports Based on the pathological orthogonal contribution weight matrix W, the feature sources are traced back along the matrix: the SNP-SNP interaction pairs (key genetic features) corresponding to the weight element with the largest absolute value are located, and the phenotypic indicators that contribute the most to classification (such as P-tau / Aβ) are identified. 42 The ratio (key phenotypic features), and their interaction term.

[0069] By associating feature sources with classification results, a structured diagnostic report is generated. For example: "1) rs123-rs456 interaction pairs against P-tau / Aβ"42 The increased ratio contributed the most; 2) the CSF characteristics of this sample were close to those of the AD cluster in the 'tau pathology' dimension; 3) in the contrast space, this sample was far from the center of the NC cluster. Overall, it was determined to be MCI. Performance Validation: Stratified combined cross-validation was employed, with samples randomly grouped according to diagnosis and repeated multiple times. Results showed that the first-level AD vs. non-AD classification accuracy reached 96.2%; the second-level MCI vs. NC classification accuracy in the non-AD population reached 84.7%; and the end-to-end three-class (AD, MCI, NC) average accuracy reached 89.3%, significantly outperforming the baseline model of direct three-class classification. Literature review revealed that most of the key SNP-SNP pairs identified in the weight matrix W were located near known AD risk genes, validating the biological rationale for the explanation.

[0070] Example 3: Based on the same technical concept, Example 3 of the present invention also provides a hierarchical interpretable classification system for heterogeneous diseases, such as... Figure 4 As shown, it includes: an interpretable fusion coding module 210, a feature decoupling enhancement module 220, a two-level cascaded decision module 230, and a diagnostic report generation module 240. Figure 5 ,in: The interpretable fusion coding module 210 is used to acquire multi-source heterogeneous data of the object to be classified, map data from different sources to a unified latent feature space, apply orthogonal constraints and pathological consistency constraints to the unified latent feature space, decompose the unified latent feature space into mutually orthogonal genotype contribution features, phenotypic contribution features and gene-phenotype interaction contribution features, obtain a unified feature representation of pathological orthogonal decoupling, and output the pathological orthogonal contribution weight matrix. The feature decoupling enhancement module 220 is used to enhance the unified feature representation of the pathological orthogonal decoupling based on contrast constraints and combined with hard sample mining for class pairs with overlapping boundaries in the unified latent feature space, and to map the enhanced features to the normalized embedding space to form a sample distribution that conforms to the pathological continuity. The two-level cascaded decision module 230 is used to perform hierarchical classification using a two-level cascaded classification process, wherein first-level classification is performed based on the unified feature representation of pathological orthogonal decoupling, and second-level classification is performed based on the enhanced features in the normalized embedding space; feedback information is generated based on the discrimination information of the second-level classification and sent back to the unified feature representation of pathological orthogonal decoupling to update the first-level classification result; The diagnostic report generation module 240 is used to track the feature contributions in the classification process based on the pathological orthogonal contribution weight matrix, and output the key feature information corresponding to the genotype contribution feature, phenotype contribution feature and gene-phenotype interaction contribution feature respectively, and generate a structured and interpretable diagnostic report.

[0071] Example 4: In one embodiment, Example 4 of the present invention also provides an electronic device; the electronic device may be a terminal, and its internal structure diagram may be as follows. Figure 6 As shown. The electronic device includes a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements the hierarchical interpretable classification method for heterogeneous diseases as described in any one of steps S101 to S104. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0072] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0073] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0074] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0075] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0076] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0077] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A hierarchical and interpretable classification method for heterogeneous diseases, characterized in that, The method includes: Obtain multi-source heterogeneous data of the object to be classified, and map the data from different sources to a unified latent feature space; apply orthogonal constraints and pathological consistency constraints to the unified latent feature space, decompose the unified latent feature space into mutually orthogonal genotype contribution features, phenotypic contribution features and gene-phenotype interaction contribution features, obtain a unified feature representation of pathological orthogonal decoupling, and output the pathological orthogonal contribution weight matrix. For class pairs with overlapping boundaries in the unified latent feature space, the unified feature representation of pathological orthogonal decoupling is enhanced based on contrast constraints and combined with hard sample mining, and the enhanced features are mapped to the normalized embedding space to form a sample distribution that conforms to the continuity of pathology. A hierarchical classification process is adopted using a two-level cascaded classification process. The first-level classification is based on the unified feature representation of the pathological orthogonal decoupling, and the second-level classification is based on the enhanced features in the normalized embedding space. Feedback information is generated based on the discriminant information of the second-level classification and sent back to the unified feature representation of the pathological orthogonal decoupling to update the first-level classification results. Based on the pathological orthogonal contribution weight matrix, the feature contributions in the classification process are tracked, and the key feature information corresponding to the genotype contribution feature, phenotype contribution feature and gene-phenotype interaction contribution feature is output respectively, generating a structured and interpretable diagnostic report.

2. The method according to claim 1, characterized in that, The process of obtaining multi-source heterogeneous data of the object to be classified includes: extracting genetic interaction features from the multi-source heterogeneous data as genotype data, extracting body fluid biomarkers as phenotypic data, and performing dimensional alignment and numerical standardization on the genotype data and the phenotypic data.

3. The method according to claim 1, characterized in that, Applying orthogonal constraints to the latent feature space includes: calculating the inner product values ​​between the genotype contribution features, the phenotype contribution features, and the gene-phenotype interaction contribution features, respectively; adding the inner product values ​​to the loss function and performing a minimization process to keep the various contribution features independent of each other.

4. The method according to claim 1, characterized in that, The output pathological orthogonal contribution weight matrix includes: splitting the pathological orthogonal contribution weight matrix into three sub-weights, respectively weighting and mapping the phenotypic contribution features, the genotype contribution features, and the gene-phenotype interaction contribution features, and adjusting the numerical proportion of each type of feature according to preset pathological rules.

5. The method according to claim 1, characterized in that, Applying pathological consistency constraints to the latent feature space includes: setting corresponding feature proportion rules according to different disease categories, and adjusting the numerical values ​​of various contributing features according to the sample category labels during training to make the feature distribution match the pathological stage.

6. The method according to claim 1, characterized in that, The step of mapping the enhanced features to the normalized embedding space includes: performing normalized projection processing on the enhanced features, mapping the enhanced features to the same scale space, and adjusting the feature positions of samples with overlapping boundaries so that samples of the same pathological stage form a continuous distribution and samples of different categories remain separated.

7. The method according to claim 6, characterized in that, The process of generating feedback information and sending it back to the unified feature representation of pathological orthogonal decoupling includes: extracting classification confidence and feature offset from the secondary classification results, combining the confidence and the offset into a feedback signal, and using the feedback signal to correct the unified feature representation of pathological orthogonal decoupling. The two-level cascaded classification process includes: first, dividing all samples into two major categories through primary classification; then, selecting samples with blurred boundaries from the major categories to enter secondary classification; and finally, sending back the category determination results obtained from secondary classification to update the major category division results of primary classification.

8. The method according to claim 1, characterized in that, The process of tracking feature contributions during classification includes: backtracking the source of features along the pathological orthogonal contribution weight matrix, sequentially locating key genotype contribution features, key phenotype contribution features, and key gene-phenotype interaction contribution features, and associating the source of features with the classification results to generate the explanatory content in the structured interpretable diagnostic report.

9. A hierarchical and interpretable classification system for heterogeneous diseases, characterized in that, include: The interpretable fusion coding module is used to acquire multi-source heterogeneous data of the object to be classified and map data from different sources to a unified latent feature space. Orthogonality constraints and pathological consistency constraints are applied to the unified latent feature space, and the unified latent feature space is decomposed into mutually orthogonal genotype contribution features, phenotype contribution features and gene-phenotype interaction contribution features to obtain a unified feature representation of pathological orthogonal decoupling, and outputs the pathological orthogonal contribution weight matrix. The feature decoupling enhancement module is used to enhance the unified feature representation of pathological orthogonal decoupling based on contrast constraints and combined with hard sample mining for class pairs with overlapping boundaries in the unified latent feature space, and to map the enhanced features to the normalized embedding space to form a sample distribution that conforms to the pathological continuity. A two-level cascaded decision module is used to perform hierarchical classification using a two-level cascaded classification process, wherein first-level classification is performed based on the unified feature representation of the pathological orthogonal decoupling, and second-level classification is performed based on the enhanced features in the normalized embedding space; Feedback information is generated based on the discrimination information of the secondary classification and sent back to the unified feature representation of the pathological orthogonal decoupling to update the primary classification result; The diagnostic report generation module is used to track the feature contributions in the classification process based on the pathological orthogonal contribution weight matrix, and output the key feature information corresponding to the genotype contribution feature, phenotype contribution feature and gene-phenotype interaction contribution feature respectively, and generate a structured and interpretable diagnostic report.

10. An electronic device, characterized in that, The electronic device includes: At least one processor; and a memory communicatively connected to said at least one processor; The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1-8.

Images