A method and system for treatment response classification based on weak feature enhancement of myocarditis
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUBEI UNIV OF EDUCATION
- Filing Date
- 2026-07-10
- Publication Date
- 2026-08-07
AI Technical Summary
[0006]本发明旨在解决儿科心肌炎治疗应答评估依赖主观经验、弱特征挖掘不足、分类精度偏低的问题，通过多源临床数据处理与弱特征智能增强，实现治疗应答效果的量化分类，为临床个体化诊疗提供客观技术支撑，提升儿科心肌炎诊疗规范化与准确化水平
1.本发明的基于心肌炎弱特征增强的治疗应答分类方法，通过采集待测患儿的时序临床检验指标、心脏影像组学特征及心电图波形数据，对多源临床数据进行预处理，采用Z-score标准化方法统一数据量纲，利用线性插值法完成不同采样频率数据的时间对齐，构建出包含不同采样频率和维度的标准化患儿特征数据集。该技术特征能够整合临床诊疗中分散的多类型数据，消除数据量纲差异与采样频率不一致带来的分析障碍，规范数据格式与时间维度，为后续的特征筛选、特征增强及分类计算提供标准统一的数据支撑，保障后续各项数据处理与模型运算的基础稳定性，提升临床数据的适配性与可分析性。
Smart Images
Figure CN122531755A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of treatment response classification and prediction technology, and more specifically, relates to a treatment response classification method and system based on weak feature enhancement of myocarditis. Background Technology
[0002] Pediatric myocarditis is a critical and frequently occurring cardiovascular disease in children, characterized by insidious onset, rapid progression, and significant individual differences in treatment response. Assessing the child's treatment response is crucial for adjusting treatment plans, reducing the risk of severe illness, and improving long-term prognosis. Currently, clinical assessment of treatment response in pediatric myocarditis relies primarily on routine laboratory monitoring and physician subjective judgment, lacking a standardized, quantitative, and objective evaluation system. This makes it difficult to capture subtle changes in the early stages of the disease, often resulting in delayed adjustments to treatment plans.
[0003] Clinical diagnosis and treatment generate multi-dimensional data such as time-series tests, cardiac imaging, and electrocardiogram waveforms. This data is stored in a scattered manner and has not been effectively integrated, failing to form a complete basis for disease assessment. Traditional feature analysis methods only focus on highly significant dominant features such as troponin and cardiac function parameters, completely ignoring latent weak features related to treatment response. Although these weak features have a low direct correlation with treatment response, they contain potential pathological information about myocardial injury, inflammatory response, and cardiac function compensation, and are important criteria for distinguishing different response types.
[0004] Existing treatment response classification models lack prior knowledge of pediatric myocarditis pathology, and their feature selection and model training lack medical logic constraints, resulting in insufficient ability to discover and utilize weak features. The models exhibit poor discriminative power at classification boundaries, struggling to accurately identify the boundaries between effective, partial, and non-response responses, leading to significant classification bias. Furthermore, the models are poorly adapted to the physiological characteristics of children, exhibiting weak generalization ability and failing to meet the needs of actual clinical diagnosis and treatment.
[0005] Inaccurate response assessments can directly lead to overtreatment or undertreatment, increasing the medical suffering of children and the financial burden on families, while also hindering the development of individualized diagnosis and treatment for pediatric myocarditis. The numerous pain points at the current clinical and technical levels urgently require an intelligent classification technology capable of uncovering occult pathological features, integrating multi-source clinical data, and aligning with pediatric pathological mechanisms. This technology would enable accurate assessment of treatment response and provide technical support for the standardized and individualized diagnosis and treatment of pediatric myocarditis. Summary of the Invention
[0006] This invention aims to address the problems of reliance on subjective experience, insufficient weak feature mining, and low classification accuracy in the assessment of treatment response in pediatric myocarditis. By processing multi-source clinical data and intelligently enhancing weak features, it achieves quantitative classification of treatment response effects, providing objective technical support for individualized clinical diagnosis and treatment, and improving the standardization and accuracy of pediatric myocarditis diagnosis and treatment.
[0007] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides a treatment response classification method based on weak feature enhancement in myocarditis, comprising: S1. Collect multi-source clinical data of the children to be tested. The multi-source clinical data includes at least time-series clinical test indicators, cardiac radiomics features, and electrocardiogram waveform data. Preprocess the multi-source clinical data to construct a standardized child feature dataset containing different sampling frequencies and dimensions. S2. Using statistical difference analysis and mutual information algorithm, highly significant dominant features are removed from the standardized patient feature dataset, and hidden features with a correlation with treatment response below a preset threshold are selected as weak feature candidate set; a pathological prior knowledge graph is introduced to constrain the weak feature candidate set, and an initial weak feature subset with potential pathological indication significance is separated. S3. Construct a deep feature enhancement network based on an attention mechanism, input the initial weak feature subset into the deep feature enhancement network; through nonlinear transformation and feature cross-derivation, amplify the discriminative power of the initial weak feature subset on the treatment response classification boundary, and generate a treatment response enhancement feature vector containing high-order interaction information; S4. The treatment response enhancement feature vector is fused with the highly significant dominant feature and input into the pre-constructed treatment response classification model; the classification model outputs the response probability distribution of the child to the current treatment plan, and generates classification results based on the response probability distribution, the classification results including effective response, partial response and no response.
[0008] Furthermore, the time-series clinical test indicators in S1 specifically include the change sequences of troponin I, creatine kinase isoenzyme, and N-terminal B-type natriuretic peptide precursor continuously collected by the child during the treatment cycle; the acquisition process of the cardiac radiomics features is as follows: the child's cardiac ultrasound video stream is acquired, the region of interest at end-diastole and end-systole of the left ventricle is extracted through a semantic segmentation network, and the gray-level co-occurrence matrix texture features and wavelet transform high-frequency coefficients within the region are calculated.
[0009] Furthermore, the specific steps in S2 to select occult features with a correlation to treatment response below a preset threshold as a weak feature candidate set include: Calculate each feature Treatment response label Pearson correlation coefficient between Set the first threshold Eliminate The dominant characteristic of strong correlation; For the remaining features, an improved adaptive bandwidth KSG mutual information estimation algorithm is used to calculate the nonlinear dependency between features and labels; the improved algorithm introduces a local density weighting factor into the distance metric of the standard KSG algorithm. The calculation formula is revised as follows: in, For the first From the nth sample point to its nth sample point The original Euclidean distance between nearest neighbors, This represents the local density gradient at that point. This is the noise suppression coefficient; Based on the corrected distance The number of nearest neighbors was recounted, and the noise-resistant mutual information value was calculated. Set a second threshold. , retain satisfaction The features are used as hidden features.
[0010] Furthermore, the introduction of a pathological prior knowledge graph in S2 to constrain the weak feature candidate set specifically refers to: First, a medical entity relationship graph containing myocardial inflammatory response pathways, myocardial cell damage markers, and cardiac function compensation mechanisms is constructed, and the features in the weak feature candidate set are mapped to the starting nodes in the graph. Secondly, the random walk method is used to search for all valid multi-hop inference paths in the medical entity relationship graph that start from the starting node and terminate at the predefined set of core pathological target nodes. The path confidence score of each valid multi-hop inference path is calculated. The path confidence score is determined based on the weighted product of the relation type weights of each edge in the path and the information entropy of the nodes. Next, the path confidence scores of all valid multi-hop inference paths of each starting node are aggregated to obtain the global pathological association strength value of the starting node relative to the core pathological mechanism. Finally, based on the distribution histogram of the global pathological association strength values, the screening threshold is automatically determined using the maximum inter-class variance method. Features with global pathological association strength values lower than the screening threshold are removed, and the retained feature set is used as the initial weak feature subset.
[0011] Furthermore, the deep feature enhancement network based on the attention mechanism in S3 adopts a multi-head self-attention mechanism, which contains multiple parallel attention heads. Each attention head learns the dependencies of the initial weak feature subset in different feature subspaces. The network generates an intermediate feature representation that integrates global context information by concatenating the outputs of each attention head and performing linear projection.
[0012] Furthermore, the nonlinear transformation in S3 employs a hyperbolic tangent activation function or an exponential linear unit activation function to introduce unsaturated nonlinear characteristics to prevent gradient vanishing. The hyperbolic tangent activation function is calculated using the following formula: The formula for calculating the exponential linear unit activation function is as follows: In the formula, The input eigenvalues represent the nonlinear transformation. Preset hyperparameters for controlling the saturation of the negative half-axis.
[0013] Furthermore, the feature cross-derivation in S3 specifically includes: A learnable attention weight matrix is used to score the relevance of feature vectors in the initial weak feature subset, and feature pairs with scores higher than a preset threshold are selected. The selected feature pairs are then subjected to a Hadamard product to generate higher-order interaction features. These higher-order interaction features are then weighted and selected using a gated fusion unit, and the selected features are concatenated with the original weak features to form the treatment response enhancement feature vector. The operational logic of the gated fusion unit is as follows: In the formula, The second-order interactive features generated by the Hadamard product. It is a primitive weak feature. This represents vector concatenation. For the Sigmoid function, and These are trainable parameters.
[0014] Furthermore, the pre-built treatment response classification model in S4 adopts the gradient boosting decision tree ensemble method, and its loss function is defined as the multi-class log loss function. During the training process, the weights of the leaf nodes are restricted by introducing an L2 regularization term, and the early stopping method is used to monitor the classification accuracy of the validation set to prevent the model from overfitting. Finally, the response probability distribution of the child is output.
[0015] As a second aspect of the present invention, a treatment response classification system based on weak feature enhancement of myocarditis is also provided, comprising: The data acquisition and standardization unit is used to collect multi-source clinical data of the children to be tested. The multi-source clinical data includes at least time-series clinical test indicators, cardiac radiomics features, and electrocardiogram waveform data. The multi-source clinical data is preprocessed to construct a standardized child feature dataset containing different sampling frequencies and dimensions. The weak feature screening and separation unit is used to remove highly significant dominant features from the standardized patient feature dataset using statistical difference analysis and mutual information algorithm, and to screen out hidden features with a correlation with treatment response below a preset threshold as a weak feature candidate set; a pathological prior knowledge graph is introduced to constrain the weak feature candidate set, and an initial weak feature subset with potential pathological indication significance is separated. A deep feature enhancement unit is used to construct a deep feature enhancement network based on an attention mechanism. The initial weak feature subset is input into the deep feature enhancement network. Through nonlinear transformation and feature cross-derivation, the discriminative power of the initial weak feature subset on the treatment response classification boundary is amplified, and a treatment response enhancement feature vector containing high-order interaction information is generated. The fusion classification and output unit is used to fuse the treatment response enhancement feature vector with the highly significant dominant feature and input it into the pre-constructed treatment response classification model. The classification model outputs the response probability distribution of the child to the current treatment plan and generates a classification result based on the response probability distribution. The classification result includes effective response, partial response and no response.
[0016] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor as described in any one of the claims, a treatment response classification method based on weak feature enhancement of myocarditis.
[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The treatment response classification method for myocarditis based on weak feature enhancement of the present invention collects time-series clinical laboratory indicators, cardiac radiomics features, and electrocardiogram waveform data of the children to be tested. It preprocesses multi-source clinical data, uses Z-score standardization to unify data dimensions, and employs linear interpolation to align data at different sampling frequencies, constructing a standardized child feature dataset containing different sampling frequencies and dimensions. This technology can integrate scattered multi-type data from clinical diagnosis and treatment, eliminate analytical obstacles caused by differences in data dimensions and inconsistent sampling frequencies, standardize data format and time dimension, provide standardized data support for subsequent feature selection, feature enhancement, and classification calculations, ensure the basic stability of subsequent data processing and model operations, and improve the adaptability and analyzability of clinical data.
[0018] 2. The treatment response classification method for myocarditis based on weak feature enhancement of the present invention, through statistical difference analysis and an improved adaptive bandwidth KSG mutual information algorithm, removes highly significant dominant features from the standardized feature dataset, selects hidden features to form a weak feature candidate set, introduces a pathological prior knowledge graph related to myocardial inflammation response, myocardial cell damage, and cardiac function compensation, and separates an initial weak feature subset with potential pathological indication significance through multi-hop inference path calculation and Otsu's inter-class variance method. This technique can extract weak features ignored in traditional analysis, filter out noisy features with no practical value based on pathological knowledge, and retain effective weak features related to the pathological mechanism of pediatric myocarditis, providing pure and pathologically valuable feature materials for subsequent feature enhancement, and avoiding the influence of invalid features on the learning and judgment of subsequent models.
[0019] 3. The treatment response classification method for myocarditis based on weak feature enhancement of the present invention constructs a deep feature enhancement network based on a multi-head self-attention mechanism. It performs nonlinear transformation and feature cross-derivation processing on an initial subset of weak features, amplifying the discriminative power of weak features at the treatment response classification boundary. This generates an enhanced feature vector containing high-order interaction information. The enhanced feature vector is then fused with highly significant dominant features and input into a pre-constructed classification model, outputting a treatment response probability distribution and generating classification results for effective response, partial response, and no response. This technique fully exploits the potential discriminative power of weak features, combining the complementary advantages of strong features and enhanced weak features to improve the accuracy of the classification model at the response boundary, stably outputting objective and quantitative treatment response classification results, and providing a reliable evaluation basis for the clinical diagnosis and treatment of pediatric myocarditis. Attached Figure Description
[0020] Figure 1 This is a flowchart of a treatment response classification method based on weak feature enhancement of myocarditis according to an embodiment of the present invention; Figure 2 This is a schematic diagram showing the direction and intensity distribution of the influence of each input feature on the treatment response classification result in an embodiment of the present invention; Figure 3 This is a schematic diagram of the system units in an embodiment of the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. Example
[0022] Please refer to Figure 1This embodiment 1 provides a treatment response classification method based on weak feature enhancement in myocarditis, including: S1. Collect multi-source clinical data of the children to be tested. The multi-source clinical data includes at least time-series clinical test indicators, cardiac radiomics features, and electrocardiogram waveform data. Preprocess the multi-source clinical data to construct a standardized child feature dataset containing different sampling frequencies and dimensions. S2. Using statistical difference analysis and mutual information algorithm, highly significant dominant features are removed from the standardized patient feature dataset, and hidden features with a correlation with treatment response below a preset threshold are selected as weak feature candidate set; a pathological prior knowledge graph is introduced to constrain the weak feature candidate set, and an initial weak feature subset with potential pathological indication significance is separated. S3. Construct a deep feature enhancement network based on an attention mechanism, input the initial weak feature subset into the deep feature enhancement network; through nonlinear transformation and feature cross-derivation, amplify the discriminative power of the initial weak feature subset on the treatment response classification boundary, and generate a treatment response enhancement feature vector containing high-order interaction information; S4. The treatment response enhancement feature vector is fused with the highly significant dominant feature and input into the pre-constructed treatment response classification model; the classification model outputs the response probability distribution of the child to the current treatment plan, and generates classification results based on the response probability distribution, the classification results including effective response, partial response and no response.
[0023] This embodiment 1 further elaborates on the above steps.
[0024] (1) Data collection and standardization The diagnosis and treatment of pediatric myocarditis generates diverse clinical data in various formats. This data suffers from mixed dimensions, different units, and significant differences in sampling frequency, making it unsuitable for direct use in subsequent feature analysis and treatment response classification. Therefore, data collection and standardization are necessary. First, pediatric patients diagnosed with acute myocarditis were selected as research subjects. Peripheral blood samples were continuously collected during the standardized treatment cycle to obtain time-series clinical laboratory data, specifically including the variation sequences of troponin I, creatine kinase isoenzymes, and N-terminal pro-B-type natriuretic peptide. Simultaneously, echocardiographic video stream data and 12-lead electrocardiogram waveform data at corresponding time points were collected, forming a multi-source clinical dataset containing time-series clinical indicators, cardiac radiomics features, and electrocardiogram waveforms.
[0025] Next, the time-series clinical laboratory indicators were serialized. The collected detection values of troponin I, creatine kinase isoenzyme, and N-terminal B-type natriuretic peptide precursor were sorted by timestamp to construct a three-dimensional time-series tensor. The first dimension is the sample number, the second dimension is the time step, and the third dimension is the indicator type, forming an initial time-series feature matrix.
[0026] Secondly, cardiac radiomics features were extracted. A U-Net semantic segmentation network based on an attention mechanism was used to process the cardiac ultrasound video stream frame by frame. The input ultrasound image was 512×512 pixels. The encoder extracted multi-scale features, and skip connections were used to fuse shallow detail information with deep semantic information. Finally, a pixel-level segmentation mask was generated at the decoder output layer. For keyframes at left ventricular end-diastole and end-systole, the Region of Interest (ROI) of the left ventricular wall and myocardial tissue was extracted based on the mask. The ROI was uniformly cropped into 256×256 pixel square regions, and histogram equalization was performed to enhance image contrast. Then, the gray-level co-occurrence matrix (GLCM) texture features were calculated. The preprocessed ROI image was converted into a 256-level grayscale image, and four directional features were constructed. The gray-level co-occurrence matrix is set to 16 gray levels to balance computational efficiency and texture detail. For each direction of the gray-level co-occurrence matrix, five basic texture features are calculated: energy, entropy, contrast, correlation, and homogeneity. The specific calculation formulas are as follows: energy: , entropy: , Contrast Ratio: , Correlation: , Homogeneity: , in For the elements of the normalized gray-level co-occurrence matrix, For gray levels, The mean, The standard deviation is used; simultaneously, a second-level wavelet transform is performed on the ROI region to extract high-frequency coefficients as a supplement to the texture features. A second-level wavelet decomposition is performed on the ROI region image using a biorthogonal wavelet basis. The first-level decomposition divides the image into low-frequency approximation components and high-frequency detail components. The second-level decomposition performs the same decomposition again on the low-frequency approximation components to obtain second-level low-frequency components and second-level high-frequency components. All high-frequency detail components are extracted as a supplement to the texture features of myocardial tissue. For each high-frequency component, its mean, standard deviation, skewness, and kurtosis are calculated as four statistical features, and finally, a radiomics feature vector is formed.
[0027] Subsequently, the ECG waveform data underwent time-frequency domain transformation. Continuous wavelet transform was used to convert the one-dimensional ECG waveform into a time-frequency graph, extracting key morphological features such as QRS complex duration, ST segment offset, and T wave inversion depth to form an ECG feature vector. The time-series clinical test indicators, cardiac radiomics features, and ECG waveform features were initially aligned. Data from different modalities were mapped onto a unified time axis based on the acquisition timestamps. Missing time nodes were filled using nearest neighbor interpolation, forming a preliminarily aligned multi-source dataset.
[0028] Next, data standardization preprocessing was performed. For clinical laboratory indicators with different dimensions, the Z-score standardization method was used for normalization. The specific calculation formula is as follows: ,in The original detection value. This represents the mean of the metric in the training set. To achieve the standard deviation, all clinical laboratory indicators were mapped to a standard normal distribution interval with a mean of 0 and a standard deviation of 1, eliminating the influence of dimensional differences on model training. The Min-Max normalization method was used for cardiac radiomics features and electrocardiogram features to scale the gray-level co-occurrence matrix texture features and wavelet transform high-frequency coefficients to the [0,1] interval, ensuring that different modal features are at the same order of magnitude.
[0029] Finally, time alignment and resampling were performed. To address the inconsistency in sampling frequencies between ECG waveform data (sampling frequency 1000Hz) and clinical laboratory indicators (sampling frequency 3 times / week), linear interpolation was used for unified resampling. The preset time step was set to 24 hours. High-frequency ECG feature sequences were averaged and pooled according to a time window (±12 hours), and low-frequency clinical laboratory indicators were linearly interpolated at missing time points, so that the three types of data were fully aligned in the time dimension. The processed time-series clinical laboratory indicators, cardiac radiomics features, and ECG waveform features were then concatenated to form a standardized child feature dataset. This dataset can be directly input into the subsequent weak feature screening module for processing.
[0030] (2) Weak feature screening and separation After standardizing the data on the characteristics of the children, the feature set needs to be screened to remove highly significant dominant features and extract latent features that are less than a preset threshold in relation to treatment response as a weak feature candidate set.
[0031] First, highly significant dominant features are removed. Based on a standardized dataset of patient features, each feature is calculated... Treatment response label Pearson correlation coefficient between The formula for calculating the Pearson correlation coefficient is as follows: ,in and Each represents a feature and tags The sample mean. Set a first threshold. The absolute value obtained from the calculation The features identified as strongly correlated dominant features are directly removed from the feature set. Next, for the remaining feature set after removing dominant features, an improved adaptive bandwidth KSG mutual information estimation algorithm is used to capture the complex nonlinear dependencies between features and treatment response. Based on the standard KSG algorithm, this embodiment introduces a local density weighting factor. The specific calculation process is as follows: First calculate the... sample points To its first The original Euclidean distance of the nearest neighbor ,in The value is set to ( (This represents the total number of samples). Subsequently, the local density gradient at that point is calculated. This gradient, obtained through kernel density estimation, reflects the rate of change in the data distribution around the sample point. A noise suppression coefficient is introduced. Set its value to 0.5, and use the correction formula. Calculate the corrected distance The core of this step lies in appropriately expanding the neighborhood distance when sample points are located in regions with drastic density changes, thereby smoothing out the interference of local noise on mutual information estimation. Subsequently, based on the corrected distance... The number of nearest neighbors within the hypersphere is recounted and substituted into the formula of the KSG mutual information estimator to calculate the noise-resistant mutual information value. The formula for calculating this value is: ,in It is a double gamma function. and They are based on corrected distance exist Space and The number of nearest neighbors within the spatial projection. This represents the average over all sample points. This calculation process accurately quantifies the features. Includes information about tags The amount of non-linear information. Finally, a second threshold is set. Used to filter hidden features, all calculated noise-resistant mutual information values are used. and Compare and retain those that meet the requirements. These features, although showing weak correlations in Pearson correlation analysis, exhibit statistically significant associations with treatment response on a non-linear dimension. Therefore, they are defined as latent features and included in the weak feature candidate set for subsequent pathological knowledge graph constraint screening.
[0032] After extracting the weak feature candidate set, a pathological prior knowledge graph is introduced to constrain the weak feature candidate set, separating an initial weak feature subset with potential pathological indicative significance. First, a medical entity-relationship graph is constructed, extracting entities and relations related to myocardial inflammatory response pathways, myocardial cell damage markers, and cardiac function compensation mechanisms as triples and storing them in a graph database. For each edge in the graph, a relation type weight is assigned based on its medical confidence level. For each node, calculate the information entropy based on its frequency of occurrence in the standard medical corpus. Map the weak feature candidate set to the set of starting nodes in the graph. And predefine the set of core pathological target nodes. .
[0033] When searching for valid multi-hop inference paths, a random walk method with restart is employed. The specific process is as follows: starting from each initial node... Starting with probability Randomly select a neighboring node of the current node for redirection, with probability. Return to the starting node During the walk, the edges and nodes traversed at each step are recorded, forming a path sequence. The endpoint of the walk belongs to the set of core pathological target nodes. At that time, the path is marked as a valid multi-hop inference path. The maximum number of steps traversed is limited to [number]. To avoid an infinite loop, the system collects data from the starting node through multiple repeated random walks. To the core pathological target node set All valid path sets .
[0034] For each valid multi-hop inference path Calculate its path confidence score The calculation formula is:
[0035] in, Representing a path The edges in and Representing edges respectively Connect the head entity and the tail entity. For the edge Relationship type weights For nodes The information entropy. This formula quantifies the joint effect of the edge weights and node information entropy in the path through a product form.
[0036] Next, for each starting node The path confidence scores of all valid multi-hop inference paths are aggregated to obtain the global pathological association strength value. The aggregation formula is:
[0037] in, For the node The set of all valid multi-hop inference paths originating from, For set The number of paths in the middle, For path The path confidence score is calculated using an average aggregation method.
[0038] Finally, based on the distribution histogram of global pathological association strength values of all starting nodes... The screening threshold is automatically determined using the maximum inter-class variance method. The calculation formula is:
[0039] in, and Histograms At the threshold The pixel ratio on the left and right sides, and These are the means of the left and right sides, respectively. Global pathological association strength values below [a certain value] are considered. Feature removal, retaining features higher than or equal to The features are used as the initial weak feature subset.
[0040] Please refer to Figure 2This figure illustrates the direction and intensity distribution of the influence of various input features on the classification results of myocarditis treatment response. The figure uses a two-way scatter plot, with the horizontal axis representing the "degree of influence on the classification result," with 0 as the dividing line; the left side represents negative influence, and the right side represents positive influence. The vertical axis lists multi-dimensional clinical features, including liver and kidney function indicators, core ECG changes, and myocardial enzyme markers. The color gradient bar on the right visually reflects the "feature value" of each feature, ranging from blue (low value) to red (high value). As can be seen in the figure, some myocardial enzyme indicators, such as "myocardial enzyme_NT-proBNP" and "myocardial enzyme_cTnI," are located in the positive influence region and are predominantly red, indicating that these are strong features significantly correlated with treatment response. Features such as "core ECG changes_atrial flutter" and "core ECG changes_ventricular fibrillation" are located in the region with an influence degree close to 0. Although their direct impact on the classification result is relatively small, they still have potential pathological significance as verified by the pathological knowledge graph, and belong to the hidden weak features that this embodiment needs to explore.
[0041] (3) Deep feature enhancement The initial weak feature subset obtained after screening has a weak signal strength and is difficult to form a clear discriminative difference in the boundary region of treatment response classification. It cannot directly provide a basis for the classification model and needs to be deeply enhanced by a dedicated network to improve its representation ability and classification discrimination.
[0042] First, the basic architecture of the network is constructed by inputting an initial subset of weak features into a multi-head self-attention mechanism module. This module contains... There are 4 parallel attention heads, each with its own independent linear projection parameter matrix. The input feature vector is first mapped to the query, key, and value space through a linear transformation, and the attention score within each head is calculated. The specific formula is as follows: ,in Let be the dimension of the key vector. Through this parallel mechanism, each attention head can capture the dependencies of weak feature subsets in different subspaces. For example, one head might focus on the dynamic correlation of temporal test indicators, while another head focuses on the potential connection between imaging texture and electrocardiogram morphology. Subsequently, the output vectors of the eight attention heads are concatenated and then dimensionality-reduced and fused through a linear projection layer to generate an intermediate feature representation that incorporates global contextual information. This process solves the problem of capturing long-distance dependencies between weak features.
[0043] Next, a nonlinear transformation is performed on the intermediate feature representation to introduce non-saturated nonlinearity to prevent gradient vanishing, thereby enhancing the model's sensitivity to weak signals. This embodiment uses either the hyperbolic tangent activation function Tanh or the exponential linear unit activation function ELU. If the Tanh function is used, its operation logic is as follows: This function maps eigenvalues to An interval can centralize the data distribution; if the ELU function is used, its operation logic is: when the input... hour, When input hour, ,Right now ,in The input eigenvalues represent the nonlinear transformation. The preset hyperparameters control the saturation of the negative half-axis. The soft saturation characteristic of the ELU function on the negative half-axis makes it more robust to noise, and its mean is close to zero, which helps to accelerate convergence. The feature vector after nonlinear activation has stronger expressive power, laying the foundation for subsequent feature crossing.
[0044] Subsequently, feature cross-derivation is performed, using a learnable attention weight matrix to score the relevance of feature vectors in the initial weak feature subset. Feature pairs with scores higher than a preset threshold are selected by calculating the dot product between each pair of feature vectors or by using a multilayer perceptron network to calculate the association score. These selected highly correlated feature pairs are then subjected to a Hadamard product operation (element-wise multiplication) to generate second-order higher-order interaction features. This transforms weak features that are originally linearly inseparable into higher-order features with clear pathological indications.
[0045] Finally, a gated fusion unit performs weighted filtering of the high-order interaction features and fuses them with the original weak features. The gated fusion unit receives the concatenated vector. As input, its operational logic strictly follows the formula: .in It is a primitive weak feature. For the Sigmoid function, and For trainable parameters, This represents element-wise multiplication. This mechanism adaptively determines which higher-order interaction features to retain and how much original information to preserve, ultimately outputting a treatment response enhancement feature vector that includes higher-order interaction information. The vector has a significantly higher discriminative power on the classification boundary than the original weak features.
[0046] (4) Integration of classification and output After obtaining the treatment response enhancement feature vector, it needs to be integrated with the highly significant dominant features and fed into the classification model for final evaluation. First, deep fusion of feature vectors is performed, concatenating the treatment response enhancement feature vector with the highly significant dominant feature vector. Before concatenation, to prevent large differences in the numerical range of features with different dimensions from affecting model convergence, all features are Z-score standardized, using the formula: ,in The original detection value. This represents the mean of the metric in the training set. The standard deviation is used. The two standardized feature sets are concatenated column-wise to form the final fused feature vector. This fusion strategy retains the direct indicative role of known highly significant biomarkers in clinical practice, while supplementing the subtle pathological changes carried by deeply mined occult and weak features, thereby constructing input data that comprehensively characterizes the child's condition.
[0047] Next, the fused feature vector is input into a pre-built treatment response classification model. In this embodiment, a gradient boosting decision tree ensemble method is used. Model construction includes... A set of regression trees is trained using a forward stepwise algorithm. For a three-class classification task of "effective response," "partial response," and "no response" in pediatric myocarditis treatment response, a multi-class logarithmic loss function is defined. Assume the sample belongs to the... The true probability of class is The model predicts the probability as Then the loss function for a single sample is defined as ,in The model optimizes the splitting nodes of the decision tree by minimizing this loss function, and uses a second-order Taylor expansion to approximate the loss function, thereby accurately finding the optimal splitting gain.
[0048] During model training, to prevent overfitting on the limited pediatric dataset, an L2 regularization term is introduced to limit the weights of the leaf nodes. The specific objective function is defined as: ,in For loss function, The number of leaf nodes. For the first The weight score of each leaf node. This is the L2 regularization coefficient (set to 1.0). A penalty coefficient (set to 0.1) is used to control tree complexity. This regularization term smooths the model's output weights, preventing the model from over-relying on certain noisy features. Simultaneously, early stopping is employed to monitor the classification accuracy on the validation set. A patience parameter of 10 epochs is set; if the validation set accuracy does not improve within 10 consecutive iterations, training automatically terminates and rolls back to the optimal number of iterations, ensuring the model has optimal generalization ability.
[0049] Finally, the model outputs the probability distribution of the child's response to the current treatment plan. The trained ensemble model infers from the input feature vector and outputs a probability vector containing three elements. , respectively, correspond to the probability values of valid response, partial response, and no response, and satisfy the following conditions: The final classification result is generated based on this probability distribution, using the maximum a posteriori probability criterion, that is, selecting the class with the highest probability value as the prediction result: For example, if the output If the condition is not met, the child is classified as "unresponsive," suggesting that the clinician may need to adjust the current treatment plan. This process achieves a complete closed loop from multimodal feature fusion to classification decision-making.
[0050] Example 2
[0051] Please refer to Figure 3 This embodiment 2 provides a treatment response classification system based on weak feature enhancement of myocarditis, including: The data acquisition and standardization unit is used to collect multi-source clinical data of the children to be tested. The multi-source clinical data includes at least time-series clinical test indicators, cardiac radiomics features, and electrocardiogram waveform data. The multi-source clinical data is preprocessed to construct a standardized child feature dataset containing different sampling frequencies and dimensions. The weak feature screening and separation unit is used to remove highly significant dominant features from the standardized patient feature dataset using statistical difference analysis and mutual information algorithm, and to screen out hidden features with a correlation with treatment response below a preset threshold as a weak feature candidate set; a pathological prior knowledge graph is introduced to constrain the weak feature candidate set, and an initial weak feature subset with potential pathological indication significance is separated. A deep feature enhancement unit is used to construct a deep feature enhancement network based on an attention mechanism. The initial weak feature subset is input into the deep feature enhancement network. Through nonlinear transformation and feature cross-derivation, the discriminative power of the initial weak feature subset on the treatment response classification boundary is amplified, and a treatment response enhancement feature vector containing high-order interaction information is generated. The fusion classification and output unit is used to fuse the treatment response enhancement feature vector with the highly significant dominant feature and input it into the pre-constructed treatment response classification model. The classification model outputs the response probability distribution of the child to the current treatment plan and generates a classification result based on the response probability distribution. The classification result includes effective response, partial response and no response.
[0052] Example 3
[0053] This embodiment 3 also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement any step of a treatment response classification method based on weak feature enhancement of myocarditis.
[0054] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0055] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.
[0056] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A treatment response classification method based on weak feature enhancement in myocarditis, characterized in that, include: S1. Collect multi-source clinical data of the children to be tested. The multi-source clinical data includes at least time-series clinical test indicators, cardiac radiomics features, and electrocardiogram waveform data. Preprocess the multi-source clinical data to construct a standardized child feature dataset containing different sampling frequencies and dimensions. S2. Using statistical difference analysis and mutual information algorithm, highly significant dominant features are removed from the standardized patient feature dataset, and hidden features with a correlation with treatment response below a preset threshold are selected as weak feature candidate set; a pathological prior knowledge graph is introduced to constrain the weak feature candidate set, and an initial weak feature subset with potential pathological indication significance is separated. S3. Construct a deep feature enhancement network based on an attention mechanism, and input the initial weak feature subset into the deep feature enhancement network; By using nonlinear transformation and feature cross-derivation, the discriminative power of the initial weak feature subset on the treatment response classification boundary is amplified, generating a treatment response enhanced feature vector containing higher-order interaction information. S4. The treatment response enhancement feature vector is fused with the highly significant dominant feature and input into the pre-constructed treatment response classification model; the classification model outputs the response probability distribution of the child to the current treatment plan, and generates classification results based on the response probability distribution, the classification results including effective response, partial response and no response.
2. The treatment response classification method based on weak feature enhancement in myocarditis according to claim 1, characterized in that, The time-series clinical test indicators in S1 specifically include the change sequences of troponin I, creatine kinase isoenzyme, and N-terminal B-type natriuretic peptide precursor continuously collected during the treatment cycle of the child; the acquisition process of the cardiac radiomics features is as follows: the child's cardiac ultrasound video stream is acquired, the region of interest at end-diastole and end-systole of the left ventricle is extracted through a semantic segmentation network, and the gray-level co-occurrence matrix texture features and wavelet transform high-frequency coefficients in the region are calculated.
3. The treatment response classification method based on weak feature enhancement in myocarditis according to claim 1, characterized in that, The specific steps in S2 to select latent features with a correlation to treatment response below a preset threshold as a weak feature candidate set include: Calculate each feature Treatment response label Pearson correlation coefficient between Set the first threshold Eliminate The dominant characteristic of strong correlation; For the remaining features, an improved adaptive bandwidth KSG mutual information estimation algorithm is used to calculate the nonlinear dependency between features and labels; the improved algorithm introduces a local density weighting factor into the distance metric of the standard KSG algorithm. The calculation formula is revised as follows: in, For the first From the nth sample point to its nth sample point The original Euclidean distance between nearest neighbors, This represents the local density gradient at that point. This is the noise suppression coefficient; Based on the corrected distance The number of nearest neighbors was recounted, and the noise-resistant mutual information value was calculated. Set a second threshold. , retain satisfaction The features are used as hidden features.
4. The treatment response classification method based on weak feature enhancement in myocarditis according to claim 1, characterized in that, The introduction of a pathological prior knowledge graph in S2 to constrain the weak feature candidate set specifically refers to: First, a medical entity relationship graph containing myocardial inflammatory response pathways, myocardial cell damage markers, and cardiac function compensation mechanisms is constructed, and the features in the weak feature candidate set are mapped to the starting nodes in the graph. Secondly, the random walk method is used to search for all valid multi-hop inference paths in the medical entity relationship graph that start from the starting node and terminate at the predefined set of core pathological target nodes. The path confidence score of each valid multi-hop inference path is calculated. The path confidence score is determined based on the weighted product of the relation type weights of each edge in the path and the information entropy of the nodes. Next, the path confidence scores of all valid multi-hop inference paths of each starting node are aggregated to obtain the global pathological association strength value of the starting node relative to the core pathological mechanism. Finally, based on the distribution histogram of the global pathological association strength values, the screening threshold is automatically determined using the maximum inter-class variance method. Features with global pathological association strength values lower than the screening threshold are removed, and the retained feature set is used as the initial weak feature subset.
5. The treatment response classification method based on weak feature enhancement in myocarditis according to claim 1, characterized in that, The deep feature enhancement network based on the attention mechanism in S3 adopts a multi-head self-attention mechanism, which contains multiple parallel attention heads. Each attention head learns the dependencies of the initial weak feature subset in different feature subspaces. The network generates an intermediate feature representation that integrates global context information by concatenating the outputs of each attention head and performing linear projection.
6. The treatment response classification method based on weak feature enhancement in myocarditis according to claim 1, characterized in that, The nonlinear transformation in S3 uses a hyperbolic tangent activation function or an exponential linear unit activation function to introduce non-saturated nonlinearity to prevent gradient vanishing. The hyperbolic tangent activation function is calculated using the following formula: The formula for calculating the exponential linear unit activation function is as follows: In the formula, The input eigenvalues represent the nonlinear transformation. Preset hyperparameters for controlling the saturation of the negative half-axis.
7. The treatment response classification method based on weak feature enhancement in myocarditis according to claim 1, characterized in that, The feature cross-derivation in S3 specifically includes: A learnable attention weight matrix is used to score the relevance of feature vectors in the initial weak feature subset, and feature pairs with scores higher than a preset threshold are selected. The selected feature pairs are then subjected to a Hadamard product to generate higher-order interaction features. These higher-order interaction features are then weighted and selected using a gated fusion unit, and the selected features are concatenated with the original weak features to form the treatment response enhancement feature vector. The operational logic of the gated fusion unit is as follows: In the formula, The second-order interactive features generated by the Hadamard product. It is a primitive weak feature. This represents vector concatenation. For the Sigmoid function, and These are trainable parameters.
8. The treatment response classification method based on weak feature enhancement in myocarditis according to claim 1, characterized in that, The pre-built treatment response classification model in S4 adopts the gradient boosting decision tree ensemble method, and its loss function is defined as the multi-class log loss function. During training, the weights of the leaf nodes are restricted by introducing an L2 regularization term, and the early stopping method is used to monitor the classification accuracy of the validation set to prevent the model from overfitting. Finally, the response probability distribution of the child is output.
9. A treatment response classification system based on weak feature enhancement in myocarditis, characterized in that, include: The data acquisition and standardization unit is used to collect multi-source clinical data of the children to be tested. The multi-source clinical data includes at least time-series clinical test indicators, cardiac radiomics features, and electrocardiogram waveform data. The multi-source clinical data is preprocessed to construct a standardized child feature dataset containing different sampling frequencies and dimensions. The weak feature screening and separation unit is used to remove highly significant dominant features from the standardized patient feature dataset using statistical difference analysis and mutual information algorithm, and to screen out hidden features with a correlation with treatment response below a preset threshold as a weak feature candidate set; a pathological prior knowledge graph is introduced to constrain the weak feature candidate set, and an initial weak feature subset with potential pathological indication significance is separated. A deep feature enhancement unit is used to construct a deep feature enhancement network based on an attention mechanism, and inputs the initial weak feature subset into the deep feature enhancement network; By using nonlinear transformation and feature cross-derivation, the discriminative power of the initial weak feature subset on the treatment response classification boundary is amplified, generating a treatment response enhanced feature vector containing higher-order interaction information. The fusion classification and output unit is used to fuse the treatment response enhancement feature vector with the highly significant dominant feature and input it into the pre-constructed treatment response classification model. The classification model outputs the response probability distribution of the child to the current treatment plan and generates a classification result based on the response probability distribution. The classification result includes effective response, partial response and no response.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as a treatment response classification method based on weak feature enhancement of myocarditis, as described in any one of claims 1-8.