CN-122000066-A - IgA nephropathy risk assessment method based on multi-modal data

CN122000066ACN 122000066 ACN122000066 ACN 122000066ACN-122000066-A

Abstract

The invention relates to the technical field of medical clinic, and discloses a multi-modal data-based IgA nephropathy risk assessment method, which comprises the steps of obtaining clinical index data, kidney biopsy pathological image and genome data of a patient, respectively extracting features of the preprocessed clinical index data, kidney biopsy pathological image and genome data through a first multi-layer perceptron network, a visual transducer and a second multi-layer perceptron network, adding the extracted features and embedding vectors of corresponding modal types to obtain a plurality of modal feature vectors, and splicing the enhanced feature sequences along the sequence dimension, performing multi-head self-attention calculation, performing weight modulation on feature vectors of all modes through a gating network to obtain gating modulation features, determining attention weights of the learnable query vectors and the gating modulation features, performing weighted fusion on the gating modulation features based on the learning query vectors to obtain fusion feature vectors, and obtaining a risk assessment result according to the fusion feature vectors through a two-class logistic regression model to improve the risk assessment precision.

Inventors

JIANG JIAWANG
LI HAILUN
ZHONG LILI
GAO XUEPING
ZHOU RONG
CHEN YAN
WANG JIAQING
LI QING

Assignees

淮安市第二人民医院

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (9)

1. A multi-modal data-based IgA nephropathy risk assessment method, comprising: Acquiring clinical index data, a kidney biopsy pathological image and genome data of a patient, and respectively preprocessing the clinical index data, the kidney biopsy pathological image and the genome data; The method comprises the steps of carrying out feature extraction on preprocessed clinical index data through a first multi-layer perceptron network to obtain a first feature vector, carrying out feature extraction on preprocessed kidney biopsy pathological images through a vision transducer to obtain a second feature vector, carrying out feature extraction on preprocessed genome data through a second multi-layer perceptron network to obtain a third feature vector; Adding the first feature vector, the second feature vector and the third feature vector with corresponding modal type embedded vectors respectively to obtain a plurality of modal feature vectors, and splicing the modal feature vectors along a sequence dimension to obtain an enhanced feature sequence; Performing multi-head self-attention calculation on the enhanced feature sequence to obtain a self-attention feature sequence, and performing weight modulation on feature vectors of all modes in the self-attention feature sequence through a gating network to obtain gating modulation features of all modes; determining attention weights between the learnable query vector and the gating modulation features of each mode, and carrying out weighted fusion on the gating modulation features of each mode based on the attention weights to obtain a fusion feature vector; And obtaining a risk assessment result based on the fusion feature vector through the two-classification logistic regression model.
2. The method of claim 1, wherein preprocessing clinical index data, kidney biopsy pathology image, and genomic data comprises: Carrying out normalization processing on continuous variable in the initial clinical index data, and carrying out coding processing on two kinds of variable in the initial clinical index data to obtain preprocessed clinical index data; Performing background correction and intensity normalization treatment on the kidney biopsy pathological image to obtain a preprocessed kidney biopsy pathological image; and carrying out deletion value processing, standardized coding and feature screening on the genome data to obtain the genome data after pretreatment.
3. The method of claim 2, wherein performing missing value filling and outlier rejection on the clinical index data to obtain initial clinical index data comprises: determining the numerical value deletion rate of each variable to be processed in the clinical index data; when the value deletion rate of the variable to be processed is smaller than a preset deletion value threshold, filling the variable to be processed by using the average value of the variable to be processed, and removing abnormal values of the filled variable to be processed; When the value deletion rate of the variable to be processed is greater than or equal to a preset deletion value threshold, filling the variable to be processed by utilizing the median of the variable to be processed, and removing abnormal values of the filled variable to be processed; And obtaining initial clinical index data based on the rejected variables to be processed.
4. The method of claim 1, wherein the modality type embedding vector comprises a clinical modality, a pathology image modality, and a genomic modality; Adding the first feature vector, the second feature vector and the third feature vector with a preset modal type embedding vector respectively to obtain a plurality of modal feature vectors, wherein the method comprises the following steps: adding the first feature vector to the clinical modality to obtain a first modality feature vector; adding the second feature vector to the pathological image mode to obtain a second mode feature vector; And adding the third eigenvector and the genome mode to obtain the third mode eigenvector.
5. The method of claim 1, wherein weight modulating feature vectors of each mode in the self-attention feature sequence by the gating network to obtain gating modulation features of each mode comprises: Determining gating weights of feature vectors of all modes in the self-attention feature sequence based on the self-attention feature sequence and the enhancement feature sequence through a gating network; and modulating the feature vector of each mode based on the gating weight to obtain the gating modulation feature of each mode.
6. The method of claim 1, wherein obtaining the risk assessment result based on the fused feature vector by a two-classification logistic regression model comprises: determining a linear weighted sum of the fusion feature vectors based on the linear combination by a two-classification logistic regression model; probability mapping is carried out on the linear weighted sum based on the Sigmoid function through a two-class logistic regression model, so that a risk assessment probability value is obtained; under the condition that the risk assessment probability value is larger than or equal to a preset risk probability threshold value, determining that the risk assessment result is high risk; and under the condition that the risk assessment probability value is smaller than a preset risk probability threshold value, determining that the risk assessment result is low risk.
7. The method according to claim 1, wherein the method further comprises: Determining at least one key fusion feature dimension with the highest contribution to the risk assessment result based on the weight vector of the classification logistic regression model; determining the correlation between each key fusion feature dimension and the gating modulation feature of each mode in the fusion feature vector, and determining the source mode to which each key fusion feature dimension belongs based on the correlation; correlating each key fusion feature dimension with corresponding original input data based on a source mode, wherein the original input data comprises clinical index data, a kidney biopsy pathological image and genome data; and integrating the associated results of the key fusion feature dimensions to generate a structured explanatory evidence report.
8. The method of claim 7, wherein determining at least one key fusion feature dimension that contributes most to a risk assessment result based on a weight vector of a two-classification logistic regression model comprises: determining the absolute value of each weight component in the weight vector, and determining the fusion feature vector dimension corresponding to the weight component with the absolute value larger than the preset absolute value threshold as the key fusion feature dimension.
9. The method of claim 7, wherein the interpretive evidence report includes at least one of an attention thermodynamic diagram, a target clinical index, and a target genomic locus; integrating the associated results of each key fusion feature dimension to generate a structured interpretive evidence report, including: When the association result indicates that the key fusion feature dimension is associated with the kidney biopsy pathological image, determining an attention weight matrix generated when the vision transducer processes the kidney biopsy pathological image, determining a first contribution degree of each image block in the kidney biopsy pathological image to the key fusion feature dimension based on the attention weight matrix, and generating an attention thermodynamic diagram corresponding to the kidney biopsy pathological image according to the first contribution degree; When the association result indicates that the key fusion feature dimension is associated with the clinical index data, a first logistic regression analysis model is constructed based on the corresponding relation between the clinical index data and the key fusion feature dimension; And determining a genome locus with the largest contribution to the key fusion feature dimension in the genome data as a target genome locus based on a weight vector of the second logistic regression analysis model.

Description

IgA nephropathy risk assessment method based on multi-modal data Technical Field The invention relates to the technical field of medical clinic, in particular to an IgA nephropathy risk assessment method based on multi-modal data. Background Immunoglobulin A (IgA) kidney disease is the most common primary glomerular disease, accounting for 45% -50% of primary glomerular disease in China, and about 20% -40% of patients progress to end stage renal disease within 10-20 years. At present, renal tissue samples can be obtained through puncture under the guidance of ultrasound, then pathological sections are manufactured, and a main doctor combines a plurality of technical means such as immunofluorescence, a light mirror, an electron microscope and the like to mainly observe whether immune complex deposition mainly comprising IgA exists in a glomerular membrane region. However, the above method relies on the experience of the doctor, resulting in poor evaluation accuracy. Disclosure of Invention Accordingly, the embodiment of the invention provides an IgA nephropathy risk assessment method based on multi-modal data, which can improve the assessment accuracy of IgA nephropathy risk. The technical scheme of the embodiment of the invention is as follows: In a first aspect of the present application, there is provided a method for IgA nephropathy risk assessment based on multimodal data, comprising: Acquiring clinical index data, a kidney biopsy pathological image and genome data of a patient, and respectively preprocessing the clinical index data, the kidney biopsy pathological image and the genome data; The method comprises the steps of carrying out feature extraction on preprocessed clinical index data through a first multi-layer perceptron network to obtain a first feature vector, carrying out feature extraction on preprocessed kidney biopsy pathological images through a vision transducer to obtain a second feature vector, carrying out feature extraction on preprocessed genome data through a second multi-layer perceptron network to obtain a third feature vector; Adding the first feature vector, the second feature vector and the third feature vector with corresponding modal type embedded vectors respectively to obtain a plurality of modal feature vectors, and splicing the modal feature vectors along a sequence dimension to obtain an enhanced feature sequence; Performing multi-head self-attention calculation on the enhanced feature sequence to obtain a self-attention feature sequence, and performing weight modulation on feature vectors of all modes in the self-attention feature sequence through a gating network to obtain gating modulation features of all modes; determining attention weights between the learnable query vector and the gating modulation features of each mode, and carrying out weighted fusion on the gating modulation features of each mode based on the attention weights to obtain a fusion feature vector; And obtaining a risk assessment result based on the fusion feature vector through the two-classification logistic regression model. In some embodiments, preprocessing the clinical index data, the renal biopsy pathology image, and the genomic data includes: Carrying out normalization processing on continuous variable in the initial clinical index data, and carrying out coding processing on two kinds of variable in the initial clinical index data to obtain preprocessed clinical index data; Performing background correction and intensity normalization treatment on the kidney biopsy pathological image to obtain a preprocessed kidney biopsy pathological image; and carrying out deletion value processing, standardized coding and feature screening on the genome data to obtain the genome data after pretreatment. In some embodiments, performing missing value filling and outlier rejection on clinical index data to obtain initial clinical index data includes: determining the numerical value deletion rate of each variable to be processed in the clinical index data; when the value deletion rate of the variable to be processed is smaller than a preset deletion value threshold, filling the variable to be processed by using the average value of the variable to be processed, and removing abnormal values of the filled variable to be processed; When the value deletion rate of the variable to be processed is greater than or equal to a preset deletion value threshold, filling the variable to be processed by utilizing the median of the variable to be processed, and removing abnormal values of the filled variable to be processed; And obtaining initial clinical index data based on the rejected variables to be processed. In some embodiments, the modality type embedding vector includes a clinical modality, a pathology image modality, and a genomic modality; Adding the first feature vector, the second feature vector and the third feature vector with a preset modal type embedding vector respectively to obtain a plurality of modal fea