CN-121983338-A - Multi-mode intelligent osteoarthritis prediction algorithm based on KAN network, storage medium and electronic equipment

CN121983338ACN 121983338 ACN121983338 ACN 121983338ACN-121983338-A

Abstract

The invention discloses a multi-mode intelligent osteoarthritis prediction algorithm based on a KAN network, a storage medium and electronic equipment, which comprise the following steps of mode data acquisition and preprocessing; the method comprises the steps of multi-modal feature explicit fusion based on KAN, knowledge graph association modeling and multi-modal fusion based osteoarthritis prediction reasoning. And (5) evaluating and optimizing the multi-mode feature fusion model. The invention converts the characteristic combining process of the traditional black box type into a group of learnable explicit mathematical functions, thereby providing reliable explanatory basis for the prediction result of the osteoarthritis. Furthermore, the method introduces a correlation modeling module based on the knowledge graph, realizes the deep fusion of multi-modal data and medical priori knowledge, and enhances the semantic correlation and pathological consistency among different modalities, thereby effectively improving the accuracy and the robustness of early prediction of osteoarthritis.

Inventors

SHAN YIFAN
FU HANG
JIANG SHUAI
Duan Yanran
LU WEI

Assignees

郑州大学第一附属医院

Dates

Publication Date: 20260505
Application Date: 20260128

Claims (8)

1. The multi-mode intelligent osteoarthritis prediction algorithm based on the KAN network is characterized by comprising the following steps of: Firstly, acquiring multi-mode data related to osteoarthritis, respectively performing data cleaning, standardization and feature initial extraction operations aiming at the characteristics of different mode data, and simultaneously realizing time alignment of different mode data with individual dimensions based on a unique patient identifier to ensure data consistency; step 2, based on KAN multimode feature explicit fusion, constructing a multimode feature fusion model by taking a KAN network as a core, mapping each preprocessed mode feature into KAN input variables respectively, converting a traditional black box type feature merging process into a group of learnable explicit mathematical functions through training, realizing effective fusion of mode features by optimizing function parameters in training according to specific components of functions of each mode feature, and providing a mathematical level basis for the interpretability of a subsequent prediction result; Step 3, knowledge graph association modeling, namely building a knowledge graph of the osteoarthritis field, wherein the knowledge graph comprises pathological mechanism, symptom mode association, medical priori knowledge, clinical diagnosis and treatment guide, drug action mechanism and rehabilitation exercise standard core entity and relationship; Step 4, multi-modal fusion osteoarthritis prediction reasoning, namely taking the KAN fused characteristics and the knowledge graph correlation result as input to construct an osteoarthritis prediction reasoning module, wherein in the reasoning process, the characteristic weight is corrected by combining the pathology correlation logic of the knowledge graph, the modal information strongly related to the early pathological signals of osteoarthritis is preferentially reserved, and the early osteoarthritis prediction result is finally output; And 5, evaluating and optimizing the multi-modal feature fusion model, namely synchronously carrying out interpretability analysis in the training and testing stage of the multi-modal feature fusion model, quantifying the contribution degree of each modal feature to a prediction result by analyzing an explicit mathematical function of KAN, and combining with pathological logic associated with the tracing feature of a knowledge graph to clearly predict the basis, thereby improving the prediction robustness and clinical applicability of the model.
2. The multi-modal intelligent osteoarthritis prediction algorithm based on the KAN network of claim 1, wherein the step 2 comprises the following steps: step 2-1, unifying the unimodal feature dimensions, namely linearly mapping preprocessing features of different modalities into unified dimensions, and eliminating input differences; (1) Wherein, the Represents the first The number of modes of operation is one, In order to unify the dimensional characteristics of the web, In order to pre-process the features, And Representing the mapping parameters; Step 2-2, single-mode KAN explicit modeling, namely converting the unified dimension characteristic into explicit mathematical expression through KAN: (2) Wherein, the For a single-mode KAN output, As a function of the basis function, And Is the core parameter of the basis function, Index for output dimension; Step 2-3, multi-mode explicit fusion, namely fusing KAN output of each mode by using a learnable weight, and keeping the explicit; (3) Wherein, the In order to be a post-fusion feature, And is also provided with Is the modal weight; Step 2-4, parameter training optimization, namely firstly obtaining a prediction result by using a linear layer: (4) Wherein, the In order to predict the outcome of the result, And As a function of the linear layer parameters, Is a sigmoid function; all parameters are updated with the goal of minimizing predictive loss. Optimizing by using binary cross entropy loss: (5) Wherein the method comprises the steps of Data total.
3. The multi-modal intelligent osteoarthritis prediction algorithm based on the KAN network as claimed in claim 1, wherein the specific steps of knowledge-graph-based correlation modeling in the step 3 are as follows: step 3-1, OA-KG entity-embedding learning for entities and relationships in osteoarthritis maps, such as "cartilage degradation", "joint pain" and "lead", discrete entities are transformed into low-dimensional value vectors by optimizing the rationality of the triplet "entity-relationship-entity": (6) Wherein the method comprises the steps of As the vector of the header entity, As a final entity vector of the vector, As a vector of the relationship(s), As a result of the marginal parameter(s), Taking a positive value; and 3-2, calculating entity association strength, namely substituting the multi-modal characteristics and the map entity vector into a formula, and calculating the similarity of the multi-modal characteristics and the map entity vector: (7) Wherein the method comprises the steps of Representing the association strength, whose value ranges are [ -1,1], the closer the value is to 1, the more closely the description feature is associated with the entity, Representing a map entity vector; Step 3-3, calculating the overall loss of the whole training framework, namely the multi-modal feature fusion model, wherein the model parameters are adjusted by combining the predicted loss of osteoarthritis, the embedded loss of map entities and the pathological consistency loss and optimizing the overall loss: (8) Wherein the method comprises the steps of Indicating the predicted loss of osteoarthritis, Representing the loss of the embedding of the map entity, Indicating a loss of pathological consistency, 、 The weight is lost.
4. The multi-modal intelligent osteoarthritis prediction algorithm based on the KAN network as claimed in claim 1, wherein the specific steps of the multi-modal fusion osteoarthritis prediction reasoning of the step 4 are as follows: step 4-1, integrating a weighted integration formula, namely integrating enhancement features of the mode fusion features and the map association of the KAN network: (9) Wherein, the , Representing a transpose operation; And 4-2, outputting the osteoarthritis prediction, namely obtaining a prediction result through a linear layer and a sigmoid activation function forming an osteoarthritis prediction reasoning module to obtain the prediction result: (10)。
5. The multi-modal intelligent osteoarthritis prediction algorithm based on the KAN network according to claim 1, wherein the step 5 specifically comprises the steps of synchronously carrying out interpretability analysis in a model training and testing stage, specifically comprising the steps of extracting a leavable activation function corresponding to each modal feature in the KAN network, drawing a function curve to analyze a nonlinear mapping relation between the feature and a prediction result, calculating an amplitude norm of the activation function, quantifying contribution degree weight of each modal feature to the osteoarthritis prediction result, mapping the high-contribution degree feature to an osteoarthritis field knowledge graph, searching a graph path between a feature entity and a disease entity, and generating a prediction interpretation report containing pathological logic.
6. The multi-modal intelligent osteoarthritis prediction algorithm based on the KAN network as claimed in claim 1, wherein said step 5 comprises the steps of 5-1, KAN activates the function by calculation The average amplitude of (a) measures the importance of a feature, and the formula is as follows: ; Wherein, the In order to verify the number of samples of the set, In order to input the characteristics of the feature, Activating a function for a trained B spline in the KAN network; The larger the value, the larger the influence of the feature on the osteoarthritis prediction result; 5-2, checking the pathological consistency based on the knowledge graph, namely calculating a pathological consistency score by combining the knowledge graph in order to verify whether the key features found by the model accord with medical common knowledge Comparing the vector representation of the high contribution feature with the vector of the "osteoarthritis" pathological entity, the formula is as follows: ; Wherein, the Is an embedded vector of the feature in the knowledge graph, Is an osteoarthritis entity vector.
7. A computer readable storage medium having stored thereon a computer program, wherein the computer program, when executed by a processor, causes a device in which the computer readable storage medium is located to perform the KAN network based multimodal intelligent osteoarthritis prediction algorithm of any of claims 1-6.
8. An electronic device comprising a memory and a processor, wherein the memory has stored thereon a program executable on the processor, and wherein the processor implements the KAN network-based multimodal intelligent osteoarthritis prediction algorithm of any of claims 1-6 when the program is executed.

Description

Multi-mode intelligent osteoarthritis prediction algorithm based on KAN network, storage medium and electronic equipment Technical Field The invention relates to the technical field of deep learning, in particular to a multi-mode intelligent osteoarthritis prediction algorithm based on a KAN network. Background Currently, osteoarthritis is a degenerative joint disease that is centered on progressive degeneration and wear of articular cartilage. Under the influence of various factors such as obesity, strain and the like, the metabolic imbalance of articular cartilage, and the repair capability can not counteract the damage progress, so that the cartilage is denatured, broken and even damaged in the whole layer. With the progress of the disease, there are often compensatory structural changes such as formation of bone tag at the joint margin and subchondral bone sclerosis. Clinically, osteoarthritis usually has hidden disease and slow progression, and typical manifestations include limited joint movement, swelling, pain, stiffness, and the like. The harm of the disease is far more than joint pain itself, and is a systemic health problem which seriously affects the life quality of patients. Persistent pain and increasingly severe joint dysfunction can make patients feel difficult in everyday activities (e.g., walking, squatting, going up and down stairs), and serious persons even gradually lose independent lifestyle, leading to disability. Because of high disability rate, great medical expenses and care burden are brought to families and society of patients. Since osteoarthritis manifests itself only as intermittent mild pain and stiffness in the early stages, it is difficult to distinguish between soft tissue strain, age-related normal changes, and the like. Therefore, it is very difficult to achieve accurate prediction of osteoarthritis. The method is very important in early prevention and determination of the illness state and the establishment of a treatment scheme. However, traditional prediction methods rely mainly on imaging examinations, with significant hysteresis and insensitivity—when structural changes can be observed with X-ray films, the pathological course of the joint often has progressed to an irreversible stage. Despite advances in this area of technology, existing models still face significant challenges in efficiently integrating multimodal data, fusing heterogeneous features, and capturing complex nonlinear relationships. The limitations limit the clinical application effect and popularization potential of the osteoarthritis prediction model. The existing multi-mode osteoarthritis prediction model can integrate image data, clinical data, biomarker data and the like, but has obvious limitation in key links. The fundamental problem is the fusion mechanism itself-the existing method relies on end-to-end black-box joint training and lacks explicit modeling capability for intra-modal correlations. This results in the model often being in the dilemma of modality competition in training, i.e., the most informative, most feature-extractable modalities (e.g., high quality images) dominate the overall network optimization direction, while those weak but critical signals (e.g., specific biomarkers or functional scores) in early predictions are systematically ignored. The method is not only difficult to realize complementary enhancement among modes, but also has generally insufficient credibility and usability of the prediction result in clinical practice due to opaque decision process. Disclosure of Invention The invention aims to provide a multi-mode intelligent osteoarthritis prediction algorithm based on a KAN network, which can effectively relieve the problem of multi-mode data competition in osteoarthritis prediction, and improves the interpretability of a prediction model by explicitly modeling each mode. According to the invention, the association modeling module based on the knowledge graph is adopted, and medical priori knowledge is introduced into the multi-modal data, so that semantic association and pathological consistency among different modalities are enhanced, more accurate and reliable expert knowledge is provided for osteoarthritis prediction, and the prediction performance is improved. The invention adopts the technical scheme that: A multi-modal intelligent osteoarthritis prediction algorithm based on a KAN network comprises the following steps: Firstly, acquiring multi-mode data related to osteoarthritis, respectively performing data cleaning, standardization and feature initial extraction operations aiming at the characteristics of different mode data, and simultaneously realizing time alignment of different mode data with individual dimensions based on a unique patient identifier to ensure data consistency; step 2, based on KAN multimode feature explicit fusion, constructing a multimode feature fusion model by taking a KAN network as a core, mapping each preprocessed mode featu