CN-122025120-A - Multi-mode data fusion type diabetes cardiovascular and cerebrovascular disease risk prediction method and system

CN122025120ACN 122025120 ACN122025120 ACN 122025120ACN-122025120-A

Abstract

The invention discloses a multi-mode data fusion method and a multi-mode data fusion system for predicting risks of cardiovascular and cerebrovascular diseases of diabetes, and relates to the technical field of medical health information processing. According to the method, the accurate identification and modeling of the artifact region are realized by constructing the artifact sensitive factor and multi-scale structure relation diagram, the artifact interference is dynamically restrained by combining the attention regulation and control and weight reconstruction mechanism, and the feature focusing accuracy is improved. The prediction accuracy, stability and clinical reliability of the model under multi-mode data are remarkably enhanced through the self-adaptive closed-loop control formed by reverse comparison and parameter updating.

Inventors

ZHENG CHAO
HE BIHANG
SUN QINGNAN
XIANG PENG
YAO CHANG
ZHANG YIKAI
HE MINGYUAN

Assignees

浙江大学医学院附属第二医院
智联嘉医(杭州)数字技术有限公司
浙江和仁科技股份有限公司

Dates

Publication Date: 20260512
Application Date: 20251230

Claims (8)

1. The multi-mode data fusion diabetes cardiovascular and cerebrovascular disease risk prediction method is characterized by comprising the following steps of: S100, acquiring medical image data, extracting characteristic indexes of suspected interference areas based on image frequency change, edge mutation amplitude and gray level abnormal distribution, and fusing and constructing an artifact sensitive factor set for data support of artifact modeling; s200, constructing a multi-scale context structure relation diagram based on an artifact sensitive factor set, and enhancing the spatial discontinuity expression through local connected residual mapping to enable an artifact region to form a weak correlation edge set in a structural diagram; S300, carrying out graph structure clustering and credibility analysis based on the structure relation graph, and weighting the interference area clustering by combining the space density, the self-similarity index and the edge continuity factor to generate a multi-stage artifact map containing positions and confidence degrees; S400, in the image feature extraction process, taking a multi-level artifact image spectrum as a channel regulation index, adopting an adjustable attention shielding mechanism to dynamically lower and raise the channel weight of a confidence artifact region, and inputting the regulated features into a prediction model to generate a focus attention image and a risk assessment result; S500, performing interference influence inverse comparison based on a model output result, performing focus attention graph and artifact graph spectrum superposition analysis, performing retrospective weight reconstruction if interference superposition exists, and adjusting channel weight and feature activation to correct output errors; s600, based on the comparison result and the reconstruction record, an artifact interference response updating mechanism is constructed, an interference track, a prediction offset and a regulation record are fused, a perception factor, a graph structure parameter and an attention strategy are periodically corrected, and closed-loop self-adaptive control on artifact interference is realized.
2. The method for predicting risk of cardiovascular and cerebrovascular diseases in diabetes based on multimodal data fusion as claimed in claim 1, wherein the step S100 comprises: Acquiring medical image data and carrying out gray scale normalization processing; extracting frequency change characteristics, and identifying a frequency energy concentrated mutation region as a candidate interference region; calculating the edge mutation amplitude of the candidate region, and screening out an image region with abnormal frequency and strong edge mutation; Modeling the gray average value, the gray standard deviation, the gray bias and the gray kurtosis of the screening area to determine a gray abnormal area; And vectorizing the frequency change characteristic, the edge mutation characteristic and the gray distribution characteristic, constructing an artifact sensitive factor set, and merging and storing the artifact sensitive factor set.
3. The method for predicting risk of cardiovascular and cerebrovascular diseases in diabetes based on multimodal data fusion as claimed in claim 1, wherein the step S200 comprises: Extracting the image coordinates, frequency outliers, edge mutation values, gray level discreteness values and confidence scores of each artifact sensitive factor; Dividing the medical image into image blocks of 20 pixels multiplied by 20 pixels, calculating the average characteristic value of the sensitive factors in each block and taking the average characteristic value as a graph structure node; Calculating a structural feature difference value according to adjacent positions of the image blocks, and establishing weak connection edges exceeding a threshold value to generate a structural relation diagram; and executing multi-scale expansion mapping, constructing a connecting edge of a first-level scale, a second-level scale and a third-level scale, and marking a weak connecting area for subsequent modeling processing.
4. The method for predicting risk of cardiovascular and cerebrovascular diseases in diabetes by multimodal data fusion according to claim 1, wherein the step S300 comprises: Extracting image block nodes covered by weak connection edges, and calculating space distribution density based on an image grid; calculating a self-similarity index based on the feature differences of each node and its neighboring nodes; Calculating an edge continuity score based on the amplitude of the image block boundary gray scale variation; And (3) carrying out weighted fusion on the three indexes after normalization to generate comprehensive confidence scores, and dividing the comprehensive confidence scores into different confidence levels to form a multi-stage artifact map.
5. The method for predicting risk of cardiovascular and cerebrovascular diseases in diabetes based on multimodal data fusion as claimed in claim 1, wherein the step S400 comprises: Mapping the two-dimensional space coordinate information of the multi-stage artifact map to each feature map layer according to the downsampling proportion of the feature map; Generating a channel mask map with the same size as the feature map layer on the basis of the mapping result, wherein the corresponding value of the artifact region in the mask map is smaller than 1, and the value of the non-artifact region is 1; multiplying the mask map element by element into the feature map channel response to realize the activation value shielding of the space position; And (3) counting the average activation intensity ratio of the channel in the artifact area and the non-artifact area, and carrying out response value weakening treatment on the hypersensitive channel in the subsequent layer according to the ratio.
6. The method for predicting risk of cardiovascular and cerebrovascular diseases in diabetes by multimodal data fusion according to claim 1, wherein the step S500 comprises: carrying out space coordinate alignment on a focus region attention distribution map and a multi-stage artifact map spectrum, analyzing the difference between an average attention value in an artifact region and an average attention value of a whole map, and judging whether an interference overlapping region exists or not; If an interference overlapping area exists, backtracking channel response of each layer in the interference area in the characteristic extraction path, and identifying a channel with an activation value significantly higher than the average value of the whole graph in the interference area as an artifact sensitive channel; Aiming at the artifact sensitive channel, adjusting the corresponding weight parameter in the upper layer of convolution kernel, applying a weight down-regulating coefficient smaller than 1, and executing normalization processing of the channel output activation intensity; And re-inputting the reconstructed channel characteristics to a model output layer, generating an updated focus region attention distribution map and risk scores, and recording adjustment parameters as a follow-up model self-adaptive optimization basis.
7. The method for predicting risk of cardiovascular and cerebrovascular diseases in diabetes by multimodal data fusion according to claim 1, wherein the step S600 comprises: Acquiring interference recognition tracks, model prediction offset trend and channel adjustment records in interference influence reverse comparison flow and backtracking weight reconstruction flow, and constructing a historical interference data set; carrying out statistical analysis on the historical interference intervention data set in a set reasoning period, and identifying a high-risk artifact region, a frequently activated artifact sensitive channel and a corresponding artifact characteristic type; adjusting an artifact sensitive factor construction threshold value, a multiscale context structure relationship graph edge weight parameter and a channel response intensity control rule of an attention shielding mechanism based on a statistical analysis result; Reloading the corrected artifact sensing factor construction rule, the structural diagram construction parameters and the channel regulation strategy into a model operation flow to realize artifact interference self-adaptive closed-loop updating control.
8. A multi-modal data fusion diabetes cardiovascular and cerebrovascular disease risk prediction system for implementing the multi-modal data fusion diabetes cardiovascular and cerebrovascular disease risk prediction method as claimed in any one of claims 1-7, characterized by comprising an artifact perception construction module, a context construction diagram module, an artifact diagram spectrum generation module, a characteristic channel regulation module, an interference backtracking correction module and an artifact self-adaptive update module: The artifact perception construction module acquires medical image data, extracts characteristic indexes of suspected interference areas based on image frequency change, edge mutation amplitude and gray level abnormal distribution, and fuses and constructs an artifact sensitive factor set for data support of artifact modeling; The context construction diagram module constructs a multi-scale context structure relation diagram based on the artifact sensitive factor set, and enhances the spatial discontinuity expression through local connected residual mapping, so that an artifact region forms a weak correlation edge set in the structure diagram; The artifact map spectrum generation module is used for carrying out map structure clustering and credibility analysis based on the structure relation map, weighting the interference area clustering by combining the space density, the self-similarity index and the edge continuity factor, and generating a multi-stage artifact map containing positions and confidence degrees; the characteristic channel regulation and control module takes the multi-level artifact image spectrum as a channel regulation and control index in the image characteristic extraction process, adopts an adjustable attention shielding mechanism to dynamically regulate down the channel weight of the confidence artifact region, and inputs the regulated and controlled characteristics into the prediction model to generate a focus attention image and a risk assessment result; The interference backtracking correction module is used for executing interference influence reverse comparison based on the model output result, carrying out focus attention graph and artifact graph spectrum superposition analysis, executing backtracking weight reconstruction if interference superposition exists, and adjusting channel weight and feature activation to correct output errors; And the artifact self-adaptive updating module is used for constructing an artifact interference response updating mechanism based on the comparison result and the reconstruction record, fusing the interference track, the prediction offset and the regulation record, and periodically correcting the perception factors, the graph structural parameters and the attention strategy to realize closed-loop self-adaptive control on the artifact interference.

Description

Multi-mode data fusion type diabetes cardiovascular and cerebrovascular disease risk prediction method and system Technical Field The invention relates to the technical field of medical health information processing, in particular to a diabetes cardiovascular and cerebrovascular disease risk prediction method and system based on multi-mode data fusion. Background The intelligent prediction of the risk of the cardiovascular and cerebrovascular diseases of the diabetes based on multi-modal data fusion is to construct a unified expression model by utilizing a data fusion technology through collecting and integrating various types of data generated by the diabetes patient in the diagnosis and treatment process, wherein the data comprises clinical diagnosis indexes (such as blood sugar, blood pressure and blood fat), medical images (such as electrocardiogram and brain CT/MRI), physiological signals recorded by a wearable device (such as heart rate, activity and sleeping conditions), life behavior data (such as diet and exercise), historical medical history and other multi-modal information, and the inherent association and characteristic complementarity among various data are captured. And then, establishing a risk prediction model capable of dynamic learning and evolution by artificial intelligent methods such as machine learning or deep learning, and intelligently evaluating the probability of complications of cardiovascular and cerebrovascular diseases (such as myocardial infarction, apoplexy and the like) of an individual in a future period of time, thereby providing scientific basis for personalized intervention, early warning and accurate treatment. The method improves the comprehensiveness and accuracy of prediction, and overcomes the limitation that the traditional prediction only depends on a single index. The prior art has the following defects: In the cardiovascular and cerebrovascular disease risk prediction process based on medical image data (such as CT and MRI), a deep learning model such as a convolutional neural network is often relied on to automatically identify and extract features of potential focus areas in images. However, in the actual acquisition process, uncontrollable artifacts such as high reflection signals generated by metal implants, motion blurring caused by micro body movements of patients, tissue edge dislocation and the like are easily generated in medical images due to factors such as patient states, scanning equipment, environmental conditions and the like, and the artifacts are not true pathological structures, but have higher local contrast and texture complexity in visual characteristics. Because the existing model lacks an explicit recognition and filtering mechanism for the artifacts, the deep neural network may activate the artifact errors into high-weight characteristic areas in the training and reasoning process, so that the focusing capability of the model on the real focus is affected. Once the artifact area is judged as a key lesion area by mistake, or the real lesion characteristics are covered up by artifact interference, the model is easy to produce high-confidence erroneous judgment, so that the risk assessment of cardiovascular and cerebrovascular events is distorted, and the clinical intervention decision and health management safety of an individual are influenced when serious. The above information disclosed in the background section is only for enhancement of understanding of the background of the disclosure and therefore it may include information that does not form the prior art that is already known to a person of ordinary skill in the art. Disclosure of Invention The invention aims to provide a multi-mode data fusion method and system for predicting risks of cardiovascular and cerebrovascular diseases of diabetes, so as to solve the problems in the background technology. In order to achieve the aim, the invention provides the following technical scheme that the diabetes cardiovascular and cerebrovascular disease risk prediction method based on multi-mode data fusion comprises the following steps: S100, acquiring medical image data, extracting characteristic indexes of suspected interference areas based on image frequency change, edge mutation amplitude and gray level abnormal distribution, and fusing and constructing an artifact sensitive factor set for data support of artifact modeling; s200, constructing a multi-scale context structure relation diagram based on an artifact sensitive factor set, and enhancing the spatial discontinuity expression through local connected residual mapping to enable an artifact region to form a weak correlation edge set in a structural diagram; S300, carrying out graph structure clustering and credibility analysis based on the structure relation graph, and weighting the interference area clustering by combining the space density, the self-similarity index and the edge continuity factor to generate a multi-stage artifact m