Search

CN-122025071-A - Multi-mode fusion-based chronic ankle instability data analysis method

CN122025071ACN 122025071 ACN122025071 ACN 122025071ACN-122025071-A

Abstract

The invention provides a multi-modal fusion-based chronic ankle joint instability data analysis method which comprises the steps of utilizing structural feature extraction and confidence entropy measurement to quantify uncertainty of output of each mode, obtaining a mode adjustment factor and compensation weight through differential kernel embedding and gating network to achieve dynamic suppression of a high uncertainty mode and enhancement of a low uncertainty mode, utilizing normalized fusion weight to conduct weighted average on each mode decision and outputting fusion prediction results and comprehensive confidence indexes.

Inventors

  • Hou Huige

Assignees

  • 暨南大学附属第一医院（广州华侨医院）

Dates

Publication Date
20260512
Application Date
20251209

Claims (10)

  1. 1. The chronic ankle joint instability data analysis method based on multi-modal fusion is characterized by comprising the following steps of: s1, extracting characteristics of multi-mode ankle joint data to generate original prediction probability distribution and corresponding mode confidence entropy of each single-mode classifier; S2, constructing a modal specificity confidence entropy manifold space based on the original prediction probability distribution and the corresponding modal confidence entropy, and generating a modal adjustment factor through a differentiable kernel embedding function; s3, inputting confidence entropy vectors of all modes into an entropy compensation gating network, and generating compensation weight vectors through activation of two full-connection layers and Sigmoid; s4, performing Softmax normalization operation on the maximum value of the original decision weight to obtain a basic fusion weight, and generating a final fusion weight vector which is dynamically adjusted through Hadamard product operation; S5, performing weighted average operation on the original prediction probability distribution based on the fusion weight vector to generate a fusion prediction result, and calculating a comprehensive confidence index; And S6, constructing a double-target loss function, wherein the main task loss adopts cross entropy of fusion prediction results and real labels, the auxiliary loss calculates the mean square error of each mode confidence entropy and the actual misjudgment rate thereof, and the parameters of the kernel embedding function and the entropy compensation gating network are updated through joint optimization.
  2. 2. The method for analyzing chronic ankle instability data based on multi-modal fusion according to claim 1, wherein the step S1 specifically comprises: preprocessing multi-mode original data acquired from a chronic ankle joint patient, and carrying out fragmentation processing on time sequence signals of all modes based on a sliding window segmentation algorithm to obtain a structured input sample sequence; Performing feature coding on the structured input sample sequence based on a mode specific feature extraction network to generate a feature representation matrix corresponding to each mode; based on the characteristic representation matrix, respectively inputting the characteristic representation matrix into a corresponding single-mode classifier, wherein the single-mode classifier consists of a full-connection layer and a Softmax activation function; Based on elements in the original prediction probability distribution, an information entropy calculation formula is adopted to quantify uncertainty of each mode prediction result, and mode specificity confidence entropy is generated; and constructing a modal discrimination capability evaluation index based on the modal specificity confidence entropy and the original prediction probability distribution.
  3. 3. The method for analyzing chronic ankle instability data based on multi-modal fusion according to claim 2, wherein the multi-modal raw data includes inertial measurement unit data, electromyographic signals, joint angle trajectories and gait image data.
  4. 4. The method for analyzing chronic ankle instability data based on multi-modal fusion according to claim 1, wherein the step S2 specifically comprises: performing confidence entropy calculation on the original prediction probability distribution output by each modal classifier to obtain a modal specificity uncertainty value; Constructing a modal specific feature space map based on the prediction probability distribution and the corresponding confidence entropy, and utilizing a nonlinear activation function to jointly encode the prediction probability distribution and the corresponding confidence entropy into a high-dimensional feature vector to obtain a joint feature vector; Based on the joint feature vector, mapping operation of a differentiable kernel embedding function is executed, embedding representation of a mode in a manifold space is constructed by utilizing a Gaussian kernel function, and a mode-specific confidence manifold vector is obtained; Based on the modal specific confidence manifold vector, performing nonlinear confidence calibration operation, performing nonlinear transformation on the modal specific confidence manifold vector by using a learnable multi-layer perceptron (MLP), and outputting a modal adjustment factor; And executing normalization processing on the modal adjustment factors, executing Softmax normalization operation on the basis of the adjustment factor sets of all the modalities, and generating normalized modal adjustment weight vectors.
  5. 5. The method for analyzing chronic ankle instability data based on multi-modal fusion according to claim 4, wherein step S2 further comprises performing a high-dimensional mapping operation by using a differentiable kernel embedding function, wherein the embedding dimension is 5-128, the kernel function is a gaussian radial basis kernel, and zero mean and unit variance normalization is performed during the embedding process.
  6. 6. The method for analyzing chronic ankle instability data based on multi-modal fusion according to claim 1, wherein the step S3 specifically comprises: Based on the confidence entropy vectors of each mode output after the multi-mode feature extraction, constructing an input tensor of the entropy compensation gating network; Executing a first layer full-connection operation on the input tensor, performing nonlinear transformation on input features by adopting a ReLU activation function, and extracting high-order interaction features of uncertainty distribution among modes; inputting the result output by the first layer full connection into a second layer full connection network, further compressing the characteristic dimension through a trainable parameter matrix, and carrying out normalization processing by adopting Batch Normalization; And applying a Sigmoid activation function at the output end of the second layer to generate a compensation weight vector.
  7. 7. The method for analyzing chronic ankle instability data based on multi-modal fusion according to claim 6, wherein the input of the entropy compensation gating network is a vector formed by splicing a confidence entropy vector and a modal adjustment weight, wherein the first full connection is activated by a ReLU, the second full connection is activated by Batch Normalization and Sigmoid, and the sum of elements of the compensation weight vector is 1 by Softmax normalization.
  8. 8. The method for analyzing chronic ankle instability data based on multi-modal fusion according to claim 1, wherein the step S4 specifically comprises: performing index normalization processing on the maximum value of the original decision weight, and performing weight distribution on each mode based on a Softmax function to generate a basic fusion weight; Performing element-by-element addition operation on the compensation weight vector output by the entropy compensation gating network and the basic fusion weight to generate an enhanced weight factor; Performing Hadamard product operation based on the enhanced weight factors and the basic fusion weights to generate final fusion weight vectors after dynamic adjustment; Performing normalization operation on the final fusion weight vector; and outputting the final fusion weight vector after dynamic adjustment to a subsequent decision fusion module to serve as an input parameter of weighted average operation.
  9. 9. The method for analyzing chronic ankle instability data based on multi-modal fusion according to claim 1, wherein the step S5 specifically comprises: Performing element-by-element weighted summation operation on the dynamic fusion weight vector and the original prediction probability distribution of each mode, and generating a fusion prediction probability vector based on a weighted average mechanism; Performing weighted linear combination operation on the confidence entropy of each mode and the corresponding fusion weight, and calculating a weighted average mode uncertainty index; Calculating a standardized uncertainty compensation factor based on the weighted average modal uncertainty index and the classification category number, and constructing a comprehensive confidence index by using the standardized uncertainty compensation factor; and mapping the fusion prediction probability vector to a normalized probability space through a Softmax function to obtain a final classification prediction result.
  10. 10. The method for analyzing chronic ankle instability data based on multi-modal fusion according to claim 9, wherein in step S5, the fusion confidence level can be used as a decision basis for clinical risk intervention or manual rechecking triggering, and the comprehensive confidence index and the final classification prediction result are jointly output to a system decision interface through interface structured coding.

Description

Multi-mode fusion-based chronic ankle instability data analysis method Technical Field The invention relates to the technical field of intelligent medical data analysis and information fusion, in particular to a chronic ankle instability data analysis method based on multi-mode fusion. Background At present, the field of intelligent medical data analysis, in particular to a multi-mode fusion technology oriented to chronic disease management and dyskinesia assessment, has become a core development trend of a data-driven decision support system. In real-world clinical or rehabilitation scenarios, joint analysis is often performed on multiple heterogeneous data sources (such as inertial sensor signals, surface myoelectricity, joint angles or gait images, etc.) to mine multi-angle information of pathological states, functional levels or risk events. The prior art has the following limitations and technical blank that firstly, the prior multi-mode fusion system is mainly focused on improving the average classification performance, and the influence of uncertainty in each mode prediction result on the final fusion reliability is ignored. In medical science and rehabilitation application scenes, the confidence capability of prediction is easy to generate severe fluctuation due to the problems of noise, shielding or abnormal acquisition and the like of each mode of data. The mainstream weighted fusion method usually adopts the maximum value of the output probability of each mode or the experience to set weight, and fails to dynamically sense the confidence change under the input data, so that the mode discrimination result with high uncertainty causes misleading to the whole evaluation; Secondly, aiming at uncertainty quantification, the traditional modes based on output probability distribution information entropy, temperature scaling and the like can roughly reflect the discreteness of model output, but the actual influence of data noise and modal specificity errors on discrimination reliability is difficult to map accurately. Some theoretical models attempt to construct a link between uncertainty and confidence based on distribution assumptions (such as dirichlet, polynomials, etc.) or evidence accumulation mechanisms, but these methods have high requirements on data distribution, large calculation amount, limited interpretability in clinical deployment, and difficulty in realizing end-to-end dynamic weight adjustment; in addition, most of the existing methods do not design a method model capable of adaptively sensing uncertainty states under multi-mode input combination, and the difficult problems of real-time inhibition of high uncertainty modes and transparent expression of fusion result confidence level are not solved, wherein the problem is how to ensure the overall discrimination accuracy. Disclosure of Invention The invention aims to solve the technical problems and provides a method for analyzing chronic ankle instability data based on multi-mode fusion. The technical scheme of the invention is realized by a method for analyzing chronic ankle instability data based on multi-mode fusion, which comprises the following steps: s1, extracting characteristics of multi-mode ankle joint data to generate original prediction probability distribution and corresponding mode confidence entropy of each single-mode classifier; S2, constructing a modal specificity confidence entropy manifold space based on the original prediction probability distribution and the corresponding modal confidence entropy, and generating a modal adjustment factor through a differentiable kernel embedding function; s3, inputting confidence entropy vectors of all modes into an entropy compensation gating network, and generating compensation weight vectors through activation of two full-connection layers and Sigmoid; s4, performing Softmax normalization operation on the maximum value of the original decision weight to obtain a basic fusion weight, and generating a final fusion weight vector which is dynamically adjusted through Hadamard product operation; S5, performing weighted average operation on the original prediction probability distribution based on the fusion weight vector to generate a fusion prediction result, and calculating a comprehensive confidence index; And S6, constructing a double-target loss function, wherein the main task loss adopts cross entropy of fusion prediction results and real labels, the auxiliary loss calculates the mean square error of each mode confidence entropy and the actual misjudgment rate thereof, and the parameters of the kernel embedding function and the entropy compensation gating network are updated through joint optimization. The method for analyzing the chronic ankle instability data based on multi-mode fusion has the following beneficial effects: (1) According to the invention, a differential kernel is introduced to embed and generate a mode-specific confidence adjustment factor, and an adaptive entr