Fusion intervention targeting small sample marker enhancement method and system

Abstract

The application discloses a small sample marker enhancement method and system fusing intervention targeting, relates to the medical information processing technical field of small sample markers, and comprises the following steps: calculating the distribution difference data between groups by using a physiological marker blank baseline data set and a physiological marker intervention group data set, and converting the distribution difference data into a physiological sensitivity mask matrix; inputting a preset noise vector into a preset baseline generation network and a preset conditional offset network respectively to obtain a baseline state vector and an initial offset vector; multiplying the initial offset vector by the physiological sensitivity mask matrix to perform zero processing and obtain a target offset vector, and superimposing the target offset vector with the baseline state vector to generate candidate enhanced marker data; and filtering loss data by using the physiological sensitivity mask matrix in combination with an adversarial discriminant network to update network weights, outputting target enhanced marker data, and training a clinical fatigue state evaluation model. The application effectively eliminates feature pollution in the small sample generation process, and significantly improves the accuracy of medical auxiliary evaluation.

Description

Technical Field

[0001] This invention relates to the field of small-sample biomarker medical information processing technology, and more specifically, this application relates to a method and system for enhancing small-sample biomarkers with fusion intervention targeting. Background Technology

[0002] In many application scenarios such as sports rehabilitation and clinical pharmacology evaluation of new drugs, obtaining clinical body fluid samples from target subjects and extracting and monitoring multidimensional physiological marker data is the core foundation for accurately assessing the body's fatigue state and diagnosing the degree of tissue function impairment.

[0003] To train a medical aid assessment model with high-precision predictive capabilities, a large-scale medical training set is typically required. However, physiological marker data collected in actual medical activities often exhibit sparse characteristics with extremely small samples.

[0004] Existing technologies generally employ deep learning data augmentation models such as generative adversarial networks to expand the initial physiological biomarker data. Label information representing external medical intervention conditions is treated as a global attribute condition and is uniformly and fully connectedly mapped and injected into the feature data streams of various dimensions of the data generation network, thereby driving the artificial intelligence model to passively synthesize simulated medical data containing the specific intervention features.

[0005] Because existing technologies force all biomarker dimensions to respond simultaneously to the same global intervention label, the augmented data generated by the model suffers from severe feature data contamination. For example, non-sensitive biomarkers that should not exhibit fluctuations produce false shift features, while the changes in truly core sensitive biomarkers are diluted by the average weight. When this severely contaminated data is fed into downstream evaluation models, it directly destroys the diagnostic model's inherent causal decision-making logic.

[0006] In scenarios involving the expansion of small-sample medical data, there exists a problem of feature data contamination caused by false shifts in insensitive biomarkers due to the global indiscriminate feature mapping of intervention labels. Summary of the Invention

[0007] To address the aforementioned technical problems, this technical solution provides a method and system for enhancing small-sample biomarkers for targeted interventions. The solution resolves the issues raised in the background section.

[0008] In a first aspect, embodiments of this application provide a method for enhancing small-sample biomarkers with fusion intervention targeting, comprising the following steps: obtaining original clinical samples of the target object for concentration quantification detection to obtain a physiological biomarker blank baseline dataset and a physiological biomarker intervention group dataset containing preset medical intervention condition labels; calculating the inter-group distribution difference data between the physiological biomarker blank baseline dataset and the physiological biomarker intervention group dataset, and transforming it through a preset mapping function to obtain a physiological sensitivity mask matrix; inputting a preset noise vector into a preset baseline generation network and a preset conditional offset network respectively for feature mapping to obtain a baseline state vector and an initial offset vector; multiplying the initial offset vector element-wise using the physiological sensitivity mask matrix, and selecting the vector with a mask coefficient lower than the first... Data from a preset threshold of physiological biomarkers in a preset dimension are zeroed out to obtain a target offset vector. The baseline state vector and the target offset vector are then superimposed to generate candidate enhancement biomarker data. This candidate enhancement biomarker data is input into a preset adversarial discriminant network. A physiological sensitivity mask matrix is ​​used to filter loss data in a preset dimension where the mask coefficient is lower than a second preset threshold, resulting in target conditional distribution loss data. Based on this loss data, the network weights of the preset conditional offset network are updated using a backpropagation algorithm. Target enhancement biomarker data is then generated and output based on the updated preset conditional offset network. The preset original medical training set is expanded based on the target enhancement biomarker data, and the preset clinical fatigue state assessment model is iteratively trained to output medical auxiliary assessment results.

[0009] Secondly, embodiments of this application provide a small-sample biomarker enhancement system that integrates intervention targeting, comprising: a data quantification module: used to acquire original clinical samples of the target object for concentration quantification detection, obtaining a physiological biomarker blank baseline dataset and a physiological biomarker intervention group dataset containing preset medical intervention condition labels; a mask generation module: used to calculate the inter-group distribution difference data between the physiological biomarker blank baseline dataset and the physiological biomarker intervention group dataset, and transform it through a preset mapping function to obtain a physiological sensitivity mask matrix; a dual-branch mapping module: used to input a preset noise vector into a preset baseline generation network and a preset conditional offset network respectively for feature mapping, obtaining a baseline state vector and an initial offset vector; and a superposition filtering module: used to multiply the initial offset vector element-wise using the physiological sensitivity mask matrix, and filter the corresponding mask. The data of physiological biomarker labels with code coefficients lower than a first preset threshold in a preset dimension are zeroed out to obtain a target offset vector. The baseline state vector and the target offset vector are then superimposed to generate candidate enhancement biomarker data. The discrimination and update module is used to input the candidate enhancement biomarker data into a preset adversarial discriminant network. It combines the physiological sensitivity mask matrix to filter the loss data of preset dimensions with mask coefficients lower than a second preset threshold to obtain target conditional distribution loss data. Based on this, the network weights of the preset conditional offset network are updated through the backpropagation algorithm. The target enhancement biomarker data is generated and output based on the updated preset conditional offset network. The medical assessment output module is used to expand the preset original medical training set based on the target enhancement biomarker data, iteratively train the preset clinical fatigue state assessment model, and output the medical auxiliary assessment results.

[0010] Thirdly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for enhancing small-sample biomarkers for targeted fusion intervention.

[0011] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. This invention calculates the inter-group distribution difference data between the baseline dataset and the intervention group dataset of physiological markers, transforms it into a physiological sensitivity mask matrix, and then uses the physiological sensitivity mask matrix to perform element-wise multiplication on the initial offset vector, forcibly zeroing the data of a preset dimension of physiological marker labels whose corresponding mask coefficients are below a preset threshold. This mechanism effectively prevents false shifts in insensitive features under medical intervention conditions, eliminates the feature contamination problem existing in existing small sample generation techniques, and ensures the biological targeting authenticity of the generated data.

[0012] 2. This invention inputs a preset noise vector into a preset baseline generation network and a preset conditional offset network for feature mapping, respectively, to derive a baseline state vector representing the individual's baseline characteristics and an initial offset vector representing the intervention intensity. The baseline state vector and the target offset vector are then superimposed to generate candidate enhancement biomarker data. This dual-branch decoupled generation mechanism independently processes the body's natural baseline metabolism and external medical intervention characteristics, greatly reducing the difficulty of model data fitting under extremely small sample conditions and improving the feature purity of the target enhancement biomarker data.

[0013] 3. In this invention, when inputting candidate enhancement biomarker data into a preset adversarial discriminant network, the loss data of a preset dimension with mask coefficients lower than a second preset threshold is filtered using a physiological sensitivity mask matrix. Based on the obtained target conditional distribution loss data, the network weights of the preset conditional offset network are specifically updated through a backpropagation algorithm. Through the dual physical constraints of forward zeroing and backward loss filtering, the output target enhancement biomarker data strictly conforms to the intervention mechanism. Expanding the preset original medical training set based on the target enhancement biomarker data significantly improves the generalization ability of the preset clinical fatigue state assessment model, thereby enhancing the accuracy and reliability of the final output medical auxiliary assessment results. Attached Figure Description

[0014] Figure 1 A schematic diagram of the structure of the small-sample biomarker enhancement method for fusion intervention targeting provided in the embodiments of this application; Figure 2 A schematic diagram of the logic flow of the small-sample biomarker enhancement method for fusion intervention targeting provided in the embodiments of this application; Figure 3 A schematic diagram of the structure of a small-sample biomarker enhancement system for targeted fusion intervention provided in this application embodiment. Detailed Implementation

[0015] This application's embodiments address the problem of feature data contamination caused by false shifts in insensitive biomarkers due to the global indiscriminate feature mapping of intervention labels in small-sample medical data expansion scenarios, through a method and system for enhancing small-sample biomarkers with fusion intervention targeting.

[0016] In applications of augmenting medical data with extremely small samples, existing data generation models, due to their indiscriminate mapping of global conditional labels, violate the objective biophysical principle that medical interventions have strong target specificity. This leads to severe feature contamination of physiological biomarker data, ultimately causing the training failure of downstream medical evaluation models. This solution addresses this core technical challenge by proposing an exploratory data processing approach based on an asymmetric physical truncation mechanism constructed from real physiological distribution differences.

[0017] This approach first acquires raw clinical samples from the target subjects for concentration quantification, separating a baseline dataset of physiological markers representing the individual's natural background state, and an intervention group dataset of physiological markers containing pre-defined medical intervention condition labels that reflects the response to external stimuli. To accurately capture the true targeting specificity of medical interventions within living organisms, this approach calculates the inter-group distribution differences between the baseline and intervention datasets, and transforms this data into a physiological sensitivity mask matrix representing the sensitivity of each marker dimension through a pre-defined mapping function. This data processing operation solidifies the objective distribution differences in clinical statistics into a structured constraint matrix that can be directly read by a deep learning network.

[0018] In the core data feature mapping stage, this scheme abandons the traditional global hybrid feature generation mode. Instead, it independently inputs a preset noise vector into a preset baseline generation network and a preset conditional offset network, respectively deriving a baseline state vector purely representing basal metabolic characteristics and an initial offset vector representing external intervention stimulus characteristics. To intercept feature contamination at the source of data flow, this scheme uses a physiological sensitivity mask matrix to perform element-wise multiplication on the initial offset vector. Data of preset dimensions of physiological biomarker labels with mask coefficients below a preset threshold are forced to zero, thereby purifying and filtering to obtain a targeted offset vector that strictly conforms to the targeted biochemical rules. Subsequently, the baseline state vector and this targeted offset vector are superimposed to generate candidate enhancement biomarker data with a fully compliant physiological logical structure.

[0019] In the adversarial discrimination and network optimization stages, this scheme inputs candidate enhancement biomarker data into a pre-defined adversarial discrimination network. Here, a physiological sensitivity mask matrix is ​​used again to filter out loss data in a pre-defined dimension whose mask coefficients are below a pre-defined threshold, resulting in purified target conditional distribution loss data. Based on this target conditional distribution loss data, the network weights of the pre-defined conditional offset network are precisely updated using a backpropagation algorithm, thereby generating and outputting target enhancement biomarker data. Finally, the pre-defined original medical training set is expanded based on the target enhancement biomarker data, and the pre-defined clinical fatigue state assessment model is iteratively trained to output accurate medical auxiliary assessment results.

[0020] This scheme constructs a physiological sensitivity mask matrix using real clinical distribution difference data, and through a bidirectional nonlinear constraint mechanism of zeroing forward dimension data and filtering backward loss data, it successfully breaks through the technical limitations of global condition indiscriminate mapping in existing technologies. It eliminates the feature contamination problem in the generation process of small sample physiological biomarker data from the root, and greatly ensures the targeting authenticity and assessment accuracy of the generated medical data.

[0021] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0022] like Figure 1 The diagram shown is a structural schematic of a small-sample biomarker enhancement method for fusion-targeted intervention provided in this application embodiment. This application embodiment provides a small-sample biomarker enhancement method for fusion-targeted intervention, comprising the following steps: obtaining original clinical samples of the target object for concentration quantification detection to obtain a physiological biomarker blank baseline dataset and a physiological biomarker intervention group dataset containing preset medical intervention condition labels; calculating the inter-group distribution difference data between the physiological biomarker blank baseline dataset and the physiological biomarker intervention group dataset, and transforming it through a preset mapping function to obtain a physiological sensitivity mask matrix; inputting a preset noise vector into a preset baseline generation network and a preset conditional offset network respectively for feature mapping to obtain a baseline state vector and an initial offset vector; using the physiological sensitivity mask matrix to perform element-wise multiplication of the initial offset vector, and setting the corresponding mask coefficient below a first preset value. The data of a preset dimension of physiological biomarker labels with a threshold value are zeroed out to obtain a target offset vector. The baseline state vector and the target offset vector are then superimposed to generate candidate enhancement biomarker data. The candidate enhancement biomarker data is input into a preset adversarial discriminant network. The loss data of a preset dimension with a physiological sensitivity mask matrix is ​​filtered out to obtain target conditional distribution loss data. Based on this, the network weights of the preset conditional offset network are updated using a backpropagation algorithm. The target enhancement biomarker data is then generated and output based on the updated preset conditional offset network. The preset original medical training set is expanded based on the target enhancement biomarker data, and the preset clinical fatigue state assessment model is iteratively trained to output medical auxiliary assessment results.

[0023] In this embodiment, Figure 2 A schematic diagram of the logic flow of the small-sample biomarker enhancement method for targeted fusion intervention provided in this application embodiment.

[0024] First, clinical body fluid samples from the target subjects were obtained and their concentrations were quantified to obtain a baseline dataset of physiological markers and an intervention group dataset of physiological markers containing pre-defined medical intervention condition labels. The medical intervention condition labels were obtained by reading the pre-defined prescription operation logs of clinical trial equipment, such as hyperbaric oxygen chambers; typical values ​​included categorical or continuous values ​​such as 1.0 ATA and 1.6 ATA.

[0025] Subsequently, the inter-group distributional differences between the baseline physiological marker dataset and the intervention group dataset are calculated, and the physiological sensitivity mask matrix is ​​obtained by transforming the data through a preset mapping function. In this step, the calculation of the distributional differences is a conventional statistical hypothesis testing technique in the field. Its input is two sets of high-dimensional datasets, and the output is a statistical scalar representing significant differences. This embodiment does not impose excessive restrictions on the specific operators of conventional hypothesis testing.

[0026] Next, the preset noise vector is input into the preset baseline generation network and the preset conditional offset network for feature mapping, respectively, to obtain the baseline state vector and the initial offset vector. The preset noise vector is obtained using a standard normal distribution random number generator, and its values ​​follow a range that conforms to... The continuous intervals are typically 128-dimensional in engineering. The preset baseline generation network and the preset conditional offset network both adopt the well-known multilayer perceptron (MLP) or one-dimensional convolutional neural network (1D-CNN) architecture. Their input is a low-dimensional vector, and their output is a high-dimensional feature vector consistent with the dimension of the physiological marker.

[0027] Then, the initial offset vector is multiplied element-wise using the physiological sensitivity mask matrix. Data of physiological marker dimensions with mask coefficients lower than a first preset threshold are zeroed out to obtain the target offset vector. The first preset threshold is obtained by statistically optimal solutions based on large-sample pre-simulation verification experiments in engineering. The setting rule is to intercept more than 95% of background noise fluctuations to ensure that only high-confidence target features are retained during forward propagation. The typical value range is [0.1, 0.5], and in this embodiment, it is preferably 0.3.

[0028] Then, the baseline state vector and the target offset vector are overlaid to generate candidate enhancement marker data.

[0029] Next, the candidate enhancement marker data is input into a pre-defined adversarial discriminant network. This network, combined with a physiological sensitivity mask matrix, filters out loss data in dimensions where the mask coefficients are below a second pre-defined threshold, resulting in the target conditional distribution loss data. The pre-defined adversarial discriminant network is a conventional discriminator structure; its input is the generated marker data, and its output is a probability value representing the data's authenticity and conditional matching degree.

[0030] Subsequently, based on the target conditional distribution loss data, the network weights of the preset conditional offset network are updated through the backpropagation algorithm, and the target enhancement marker data are output.

[0031] Finally, the pre-set original medical training set is expanded based on the target augmentation biomarker data, and the pre-set clinical fatigue state assessment model is iteratively trained to output medical auxiliary assessment results. The pre-set original medical training set is derived from a small-sample electronic medical record database previously stored in the hospital; the pre-set clinical fatigue state assessment model is obtained using a conventional support vector machine (SVM) or random forest classifier, with the expanded multidimensional physiological biomarker data as input and the fatigue level classification assessment results as output.

[0032] Through the above technical solution, this embodiment solves the technical problems of false shifts in insensitive biomarkers and feature contamination caused by indiscriminate global conditional label mapping in existing small-sample generation models by bidirectionally introducing a physiological sensitivity mask matrix based on inter-group distribution differences into the data streams of the generator and discriminator networks, and performing forced element-wise multiplication to zero after feature mapping. This achieves the internalization of the target-specific physical constraints of external intervention into neural network computational logic under extremely small sample conditions, ensuring that the output target enhancement biomarker data strictly conforms to the real biochemical metabolic cascade mechanism, fundamentally improving the diagnostic accuracy and reliability of downstream clinical fatigue state assessment models.

[0033] Furthermore, the specific process for obtaining the inter-group distribution difference data is as follows: Based on a preset number of resampling cycles, the physiological marker blank baseline dataset and the physiological marker intervention group dataset are resampled using sampling with replacement to generate multiple parallel resampled validation datasets; the inter-group distribution difference data of each resampled validation dataset is calculated to construct a difference data distribution set; all data values ​​in the difference data distribution set are sorted in ascending order to obtain a numerical sorting sequence; based on the lower percentile value of the preset confidence interval, the value at the corresponding index position is extracted from the numerical sorting sequence and used as the statistical aggregate value; the statistical aggregate value is used as the noise-resistant corrected inter-group distribution difference data to obtain the inter-group distribution difference data of the physiological marker labels.

[0034] In this embodiment, the process of obtaining data on the inter-group distribution differences of physiological markers is further specified.

[0035] First, based on a preset number of resampling cycles, the baseline dataset and the intervention group dataset of physiological markers are resampled using sampling with replacement, generating multiple parallel resampled validation datasets. The preset number of resampling cycles is determined according to Bootstrap statistical industry standards, and its value typically ranges from [value missing]. The typical value for the project is 1000 times.

[0036] Secondly, calculate the inter-group distribution difference data for each resampled validation dataset and construct a set of difference data distributions.

[0037] Next, all data values ​​in the differential data distribution set are sorted in ascending order to obtain a sorted sequence. This is a well-known and conventional data sorting method in the field, with an unordered set as input and a monotonically increasing ordered array as output.

[0038] Subsequently, based on the lower percentile value of the preset confidence interval, the value at the corresponding index position is extracted from the numerical sorting sequence and used as the statistical aggregate value. The lower percentile value of the preset confidence interval is obtained by using a confidence level set according to the gold standard of medical clinical trials, such as a 95% confidence interval.

[0039] Finally, the statistically aggregated values ​​were used as noise-resistant corrected data for the inter-group distribution differences to obtain the inter-group distribution differences of physiological markers.

[0040] Through the above technical solution, this embodiment solves the technical problem of oscillation distortion of inter-group difference parameters caused by the random noise amplification effect due to the limited sensitivity of hardware sensors in clinical in vivo sampling by introducing a Bootstrap resampling mechanism and confidence interval lower limit extraction logic. It effectively smooths out extreme noise fluctuations in small samples, ensuring that the underlying data foundation for subsequent conversion into a sensitivity mask matrix has extremely high noise resistance and statistical confidence.

[0041] Furthermore, the specific processing procedure of the preset mapping function is as follows: the physiological marker baseline dataset and the physiological marker intervention group dataset contain data with physiological marker labels of different preset dimensions and different preset medical intervention condition labels; the inter-group distribution difference data of the physiological marker baseline dataset and the physiological marker intervention group dataset are calculated to obtain the inter-group distribution difference data of physiological marker labels of different preset dimensions under different preset medical intervention condition labels; the specific calculation formula of the preset mapping function is: ;in, For the first Physiological marker labels of preset dimensions in the first... Data on the inter-group distribution differences of pre-defined medical intervention condition labels. The first of the physiological sensitivity mask matrix Physiological marker labels of preset dimensions in the first... The mask coefficient of a preset medical intervention condition label, To preset a medical significance threshold, To preset smooth adjustment data, This represents the natural exponential function.

[0042] In this embodiment, the processing procedure of the preset mapping function is further specifically defined.

[0043] A preset medical significance threshold is defined, and its setting rules are derived from industry benchmarks for hypothesis testing in clinical biostatistics. The value range is [missing value]. The typical value for engineering projects is 0.05.

[0044] Pre-set smoothing adjustment data, obtained through gradient smoothness testing during model pre-training, aims to prevent non-differentiable hard steps in the mask. The value range is [range missing]. The typical value for engineering projects is 0.01.

[0045] Through the above technical solution, this embodiment solves the technical problem that discrete statistical difference parameters cannot be directly used as differentiable attention weights in deep learning networks for backpropagation by designing a generalized logistic inverse mapping function with a smooth adjustment mechanism. This achieves a smooth and rigorous transformation of biochemical significance criteria into... The network mask coefficients of the interval provide a solid mathematical model foundation for subsequent asymmetric gradient truncation.

[0046] Furthermore, the specific process for obtaining the baseline state vector and the initial offset vector includes: inputting a preset noise vector into a preset baseline generation network for natural distribution mapping to obtain a baseline state vector representing the basal metabolic homeostasis of the target object; extracting preset medical intervention condition labels corresponding to the physiological marker intervention group dataset, and using the conditional embedding layer of the preset conditional offset network to transform the preset medical intervention condition labels into conditional feature vectors; inputting the preset noise vector into the preset conditional offset network for shallow feature extraction to obtain initial hidden layer feature data; in the hidden layer of the preset conditional offset network, using the conditional feature vector to perform feature fusion modulation processing on the initial hidden layer feature data to obtain conditionally modulated hidden layer feature data; and performing deep feature inference mapping on the conditionally modulated hidden layer feature data to output an initial offset vector consistent with the preset dimension of the baseline state vector.

[0047] In this embodiment, the internal flow process of the generation network for the baseline state vector and the initial offset vector is further defined.

[0048] In the hidden layers of the pre-defined conditional offset network, the initial hidden layer feature data is fused and modulated using conditional feature vectors to obtain conditionally modulated hidden layer feature data. The complete execution process of the feature fusion and modulation process is as follows: First, the continuous conditional feature vectors are broadcast in higher dimensions to align with the tensor dimension of the initial hidden layer feature data. Then, the matching weights of the conditional vectors and hidden layer features in different channel dimensions are calculated through a spatial attention mechanism. Finally, the weight matrix is ​​multiplied by the hidden layer features to achieve deep injection of conditional information into the data stream.

[0049] Through the above technical solution, this embodiment solves the technical problem of excessive network coupling and difficulty in accurately capturing single external intervention features caused by traditional generative networks mixing all conditions into the input at once by constructing a dual-branch decoupled extraction logic that separates basal metabolic mapping from conditional modulation. It achieves refined modulation injection of conditional feature vectors at the hidden layer depth, greatly improving the model's accuracy in inferring the intensity of medical interventions and the feature representation capability of the generated vectors.

[0050] Furthermore, after obtaining the physiological sensitivity mask matrix, the process includes: extracting time-series feature data with timestamp attributes from the physiological biomarker intervention group dataset; using the Granger causality test algorithm to calculate the causal test F-statistic between the time nodes corresponding to the preset medical intervention condition labels and the corresponding time nodes of different physiological biomarker labels in the time-series feature data, and using the causal test F-statistic as the Granger causality strength data; comparing the Granger causality strength data with a preset causal significance threshold; if the Granger causality strength data is greater than the causal significance threshold, then the corresponding time node is recorded as a response mutation point. ; Calculate the time span value between the time node corresponding to the preset medical intervention condition label and the response mutation point; if the time span value is greater than the preset delay threshold, the corresponding physiological marker label dimension is determined to be the lagged physiological marker label dimension, and the time period corresponding to the time span value is defined as the response lag time window; within the response lag time window, the dynamic decay weight data is calculated using the exponential decay function, and the corresponding mask coefficients of the lagged physiological marker label dimension in the physiological sensitivity mask matrix are multiplied with the dynamic decay weight data to replace the corresponding mask coefficients, thus obtaining the corrected physiological sensitivity mask matrix.

[0051] In this embodiment, the random environmental noise amplification effect caused by the limited sensitivity of the clinical in vivo sampling hardware sensor causes the calculated inter-group distribution difference data to distort and oscillate, which further limits the causal correction logic for the time lag characteristics after mask generation.

[0052] The data is compared with a preset causal significance threshold. The preset causal significance threshold is obtained by consulting the standard F-distribution table and determining the critical F-value based on the preset degrees of freedom. The typical engineering value usually corresponds to the F-statistic when p=0.05.

[0053] The preset delay threshold is set based on prior pharmacokinetic knowledge of the half-life of physiological substances, and its value range is typically [range missing]. If the typical value is 15 minutes, it is determined to be a lagging physiological marker dimension, and this time period is defined as the response lag time window.

[0054] Finally, within the response lag time window, the dynamic decay weight data is calculated using an exponential decay function. This decay function is engineered to exhibit a non-linear decreasing trend with the time window percentage, and the dynamic decay weight data value can smoothly decay from 1.0 to 0.1. The corresponding mask coefficient for this dimension is then multiplied and replaced with this weight to obtain the corrected physiological sensitivity mask matrix.

[0055] Through the above technical solution, this embodiment solves the technical problem that the mask coefficients extracted at the same time point cannot accurately reflect the temporal physiological misalignment and distortion due to the inherent temporal lag of biological metabolism. This is achieved by introducing Granger causality testing and a time-response window detection mechanism. Dynamic adaptive correction of the mask coefficients along the time dimension is realized, ensuring that the evolution of the enhanced data on the time axis fully conforms to the lag characteristics of real pharmacokinetics and physiological responses.

[0056] Furthermore, the preset baseline generation network also includes an update step, specifically including: calculating the Pearson correlation coefficient between data of different preset dimensions in the physiological biomarker blank baseline dataset to obtain the Pearson correlation coefficient of the data pair; if the Pearson correlation coefficient of any data pair is greater than a preset collinearity threshold, then pairing the corresponding physiological biomarker labels of preset dimensions to determine the target physiological dimension pair; calculating the first covariance matrix of the target physiological dimension pair in the physiological biomarker blank baseline dataset, and using the first covariance matrix as the background covariance feature data; when candidate enhancement biomarker data is input into the preset adversarial discriminant network, calculating the candidate enhancement biomarker... The second covariance matrix of the target physiological dimension pair in the object data is obtained; the matrix 2 norm difference between the first and second covariance matrices is calculated, and the matrix 2 norm difference is used as the physiological covariance loss data; the basic generation loss data of the preset baseline generation network is obtained; the physiological covariance loss data and the basic generation loss data are weighted and fused using the obtained preset weight coefficients to obtain the comprehensive baseline loss data; based on the comprehensive baseline loss data, the gradient data of the preset baseline generation network weights is calculated using the chain rule; and the network weights of the preset baseline generation network are numerically iteratively updated according to the gradient data using a preset optimization algorithm.

[0057] In this embodiment, since the body's energy metabolism and oxidative stress system are restricted by the physical law of mass conservation, there is a strong physical collinearity among some preset dimensions of physiological marker labels. In order to prevent the mask coefficient from returning to zero and causing internal logical breaks in the generated data, the update steps of the preset baseline generation network are further limited.

[0058] First, Pearson correlation coefficients are calculated between data from different dimensions in the baseline dataset of physiological markers. If any Pearson correlation coefficient is greater than a preset collinearity threshold, it is identified as a target physiological dimension pair. The industry standard for the preset collinearity threshold is derived from the medical statistics strong correlation definition criteria, and its value range is [insert range here]. The typical value is 0.8.

[0059] Next, the first covariance matrix of this dimension pair in the baseline dataset and the second covariance matrix in the candidate enhancement biomarker data when input into the adversarial discriminant network are calculated respectively; the matrix L2 difference between the two is calculated as the physiological covariance loss data.

[0060] Subsequently, the basic generation loss data of the preset baseline generation network is obtained; using preset weight coefficients, the two are weighted and fused to obtain the comprehensive baseline loss data. The preset weight coefficients are determined based on a grid search optimization strategy, and their values ​​range from [value range missing]. The typical value is 0.1.

[0061] Finally, based on the comprehensive baseline loss data, the gradient data of the network weights is calculated using the chain rule; and a preset optimization algorithm, such as the Adam optimization algorithm, is used to perform numerical iterative updates based on the gradient data. The convergence logic of this iterative update is as follows: gradient descent continues to be executed until the rate of change of the comprehensive baseline loss data is less than the typical engineering boundary value or the maximum set number of iterations is reached, at which point the iteration stops to ensure that the model can converge stably.

[0062] Through the above technical solution, this embodiment solves the technical problem that the forced zeroing operation of the mask in a specific dimension may cause the generated data to violate the law of conservation of matter and energy in the body and destroy the original physical collinearity between biomarkers by constructing a Pearson correlation detection and matrix L2 difference loss feedback link for target physiological dimension pairs. It achieves the technical effect of actively maintaining the internal logical compliance of the biochemical cascade during iterative updates of the baseline generation network, eliminating physical logical breaks.

[0063] Furthermore, the specific process for obtaining the target conditional distribution loss data is as follows: During the forward propagation processing of candidate enhancement marker data by the preset adversarial discriminant network, the intermediate conditional judgment feature matrix output by the hidden layer of the preset adversarial discriminant network is extracted; the intermediate conditional judgment feature matrix and the physiological sensitivity mask matrix are subjected to Adama multiplication, and in the calculation result matrix, the feature elements whose corresponding mask coefficients are lower than the second preset threshold are forced to zero to obtain the masked feature matrix; the masked feature matrix is ​​input into the output layer of the preset adversarial discriminant network, and the cross-entropy loss data is calculated in combination with the preset medical intervention condition labels, and the cross-entropy loss data is used as the target conditional distribution loss data.

[0064] In this embodiment, firstly, during the forward propagation processing of candidate enhancement marker data by the preset adversarial discriminant network, the intermediate condition judgment feature matrix output by the hidden layer of the preset adversarial discriminant network is extracted. This is a feature map extraction technique for deep neural networks, whose input is a multi-dimensional network node tensor and whose output is a truncated two-dimensional feature matrix representation.

[0065] Secondly, the intermediate condition judgment feature matrix and the physiological sensitivity mask matrix are subjected to an Adama product calculation. In the resulting matrix, feature elements with mask coefficients lower than a second preset threshold are forcibly set to zero, thus obtaining the masked feature matrix. The second preset threshold is determined by testing the adversarial convergence speed of the pre-set adversarial discriminant network. Its setting rule aims to eliminate high-frequency noise gradients caused by early errors in the discriminator, preventing invalid or weak gradient backpropagation from interfering with the update direction of network weights. The typical range of the second preset threshold is... In this embodiment, 0.25 is preferred. In actual engineering implementation, to simplify the complexity of hyperparameter tuning, the second preset threshold can also be set to a value equal to the first preset threshold.

[0066] Finally, the masked feature matrix is ​​input into the output layer of the pre-defined adversarial discriminant network, and the cross-entropy loss data is calculated in conjunction with the medical intervention condition labels. The specific execution logic for this complex loss determination process is as follows: First, the predicted probability distribution of the current data belonging to each medical intervention condition label is calculated based on the output layer features. Then, the real medical intervention condition labels are obtained and converted into a one-hot encoding format. Finally, the relative entropy between the predicted probability distribution and the real one-hot encoding, i.e., the cross-entropy, is calculated to quantify the judgment error of the output network, and the cross-entropy loss data is used as the target condition distribution loss data.

[0067] Through the above technical solution, this embodiment solves the technical problem that traditional discriminators blindly collect irrelevant dimensional errors when evaluating conditional generation data, resulting in backpropagation gradients containing a large amount of background noise, by deeply embedding a masked Adama product blocking mechanism within the adversarial discriminant network. This achieves precise physical limitation on the field of view for discriminant network error calculation, ensuring that the backpropagated loss gradient is 100% derived from the actual medical target dimension, greatly accelerating the Nash equilibrium convergence speed of the adversarial network.

[0068] Furthermore, the specific process for acquiring target enhancement biomarker data is as follows: Based on the target conditional distribution loss data, the loss gradient data of each network layer node of the preset conditional offset network is calculated using the chain rule; using a preset optimization algorithm, the network weights of the preset conditional offset network are numerically iteratively updated according to the loss gradient data to obtain the updated network weights; the preset noise vector and preset medical intervention condition labels are input into the preset conditional offset network loaded with the updated network weights for feature secondary deduction to obtain the final offset vector; the physiological sensitivity mask matrix is ​​used to perform element-wise multiplication filtering with the final offset vector, and the data of the dimension with the corresponding mask coefficient below the third preset threshold is zeroed and superimposed with the baseline state vector to obtain and output the target enhancement biomarker data.

[0069] In this embodiment, firstly, based on the target conditional distribution loss data, the loss gradient data of each network layer node of the preset conditional offset network is calculated using the chain rule. Secondly, using a pre-defined optimization algorithm, the initial network weights of the pre-defined conditional offset network are numerically updated iteratively based on the loss gradient data to obtain the updated network weights. The convergence logic of this pre-defined optimization algorithm also follows the standard gradient descent termination condition criterion.

[0070] Next, the preset noise vector and medical intervention condition labels are re-input into the preset condition offset network loaded with updated network weights, and feature secondary derivation is performed to obtain the final offset vector.

[0071] Finally, after applying the same filtering process to the final offset vector using a physiological sensitivity mask matrix, it is superimposed with the baseline state vector to obtain and output the target enhancement marker data. The third preset threshold is determined based on high-fidelity requirements for clinical medical data. Its setting rule is to perform a final rigorous check before outputting the final enhancement data to ensure that the final output target offset features do not contain any statistically weak spurious feature fluctuations. Since it is the final output layer, its value is typically equal to or slightly stricter than the first preset threshold, with a typical range of values. In this embodiment, 0.3 is preferred.

[0072] Through the above technical solution, this embodiment solves the technical problem that relying solely on a single forward propagation cannot dynamically correct the model's deduction direction using adversarial errors, thus limiting the accuracy of the generated data. This is achieved by clearly defining a complete data processing loop involving loss gradient backpropagation, weight iteration, and secondary feature deduction. Dynamic fine-tuning of model parameters based on pure targeting loss is realized, ensuring that the final output target enhancement biomarker data fully absorbs the optimization results of adversarial training, achieving the data fidelity requirements that meet clinical medical standards.

[0073] Figure 3 This is a schematic diagram of the structure of a small-sample biomarker enhancement system for fusion-based intervention targeting provided in this application embodiment. The system includes: a data quantization module for acquiring original clinical samples of the target object and performing concentration quantification detection to obtain a physiological marker baseline dataset and a physiological marker intervention group dataset containing preset medical intervention condition labels; a mask generation module for calculating the inter-group distribution difference data between the physiological marker baseline dataset and the physiological marker intervention group dataset, and transforming it through a preset mapping function to obtain a physiological sensitivity mask matrix; a dual-branch mapping module for inputting a preset noise vector into a preset baseline generation network and a preset conditional offset network respectively for feature mapping to obtain a baseline state vector and an initial offset vector; and a superposition filtering module for applying the physiological sensitivity mask matrix to the initial offset vector. The algorithm performs element-wise multiplication to zero out the data of the preset dimension of physiological biomarker labels whose mask coefficients are lower than the first preset threshold, obtaining the target offset vector. The baseline state vector and the target offset vector are then superimposed to generate candidate enhancement biomarker data. The discrimination and update module inputs the candidate enhancement biomarker data into a preset adversarial discriminant network, filters the loss data of the preset dimension whose mask coefficients are lower than the second preset threshold using the physiological sensitivity mask matrix, obtains the target conditional distribution loss data, and updates the network weights of the preset conditional offset network using the backpropagation algorithm. Based on the updated preset conditional offset network, the algorithm generates and outputs the target enhancement biomarker data. The medical assessment output module expands the preset original medical training set based on the target enhancement biomarker data, iteratively trains the preset clinical fatigue state assessment model, and outputs the medical auxiliary assessment results.

[0074] This application also provides a computer-readable storage medium for storing a computer program that, when executed by a processor, implements a small-sample biomarker enhancement method for targeted fusion intervention.

[0075] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0076] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0079] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0080] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for enhancing small-sample biomarkers for targeted intervention, characterized in that, Includes the following steps: The original clinical samples of the target subjects were obtained and their concentrations were quantified to obtain a baseline dataset of physiological markers and an intervention group dataset of physiological markers containing labels of preset medical intervention conditions. The intergroup distribution difference data between the physiological marker blank baseline dataset and the physiological marker intervention group dataset is calculated, and the physiological sensitivity mask matrix is ​​obtained by transforming it through a preset mapping function; The preset noise vector is input into the preset baseline generation network and the preset conditional offset network respectively for feature mapping to obtain the baseline state vector and the initial offset vector. The initial offset vector is multiplied element-wise using the physiological sensitivity mask matrix. Data of the preset dimension of physiological marker labels whose corresponding mask coefficients are lower than the first preset threshold are zeroed to obtain the target offset vector. The baseline state vector and the target offset vector are then superimposed to generate candidate enhancement marker data. The candidate enhancement marker data is input into the preset adversarial discriminant network. The loss data of the preset dimension with the mask coefficient lower than the second preset threshold is filtered by the physiological sensitivity mask matrix to obtain the target conditional distribution loss data. Based on this, the network weights of the preset conditional offset network are updated by the backpropagation algorithm. The target enhancement marker data is generated and output based on the updated preset conditional offset network. The preset original medical training set is expanded based on target augmentation biomarker data, and the preset clinical fatigue state assessment model is iteratively trained to output medical auxiliary assessment results.

2. The method for enhancing small-sample biomarkers for targeted fusion intervention according to claim 1, characterized in that, The specific process for obtaining the inter-group distribution difference data is as follows: Based on a preset number of resampling cycles, the baseline dataset of physiological markers and the intervention group dataset of physiological markers are resampled using a sampling with replacement method to generate multiple parallel resampled validation datasets. Calculate the inter-group distribution difference data for each resampled validation dataset and construct a set of difference data distributions; Sort all data values ​​in the differential data distribution set in ascending order to obtain a sorted sequence of values. Based on the lower percentile value of the preset confidence interval, extract the value at the corresponding index position in the numerical sorting sequence and use it as the statistical aggregate value. By using the statistically aggregated values ​​as noise-resistant and corrected data for the inter-group distribution differences, we obtained the inter-group distribution differences data for physiological marker labels.

3. The method for enhancing small-sample biomarkers for targeted fusion intervention according to claim 2, characterized in that, The specific processing procedure of the preset mapping function is as follows: The physiological marker blank baseline dataset and the physiological marker intervention group dataset contain data with different preset dimensions of physiological marker labels and different preset medical intervention condition labels; The intergroup distribution difference data of physiological marker blank baseline dataset and physiological marker intervention group dataset is calculated to obtain the intergroup distribution difference data of physiological marker labels of different preset dimensions under different preset medical intervention condition labels. The specific calculation formula for the preset mapping function is as follows: ； in, For the first Physiological marker labels of preset dimensions in the first... Data on the inter-group distribution differences of pre-defined medical intervention condition labels. The first of the physiological sensitivity mask matrix Physiological marker labels of preset dimensions in the first... The mask coefficient of a preset medical intervention condition label, To preset a medical significance threshold, To preset smooth adjustment data, This represents the natural exponential function.

4. The method for enhancing small-sample biomarkers for targeted fusion intervention according to claim 3, characterized in that, The specific process for obtaining the baseline state vector and the initial offset vector includes: The preset noise vector is input into the preset baseline generation network for natural distribution mapping to obtain the baseline state vector characterizing the baseline metabolic homeostasis of the target object; Extract the preset medical intervention condition labels corresponding to the physiological marker intervention group dataset, and use the conditional embedding layer of the preset conditional offset network to transform the preset medical intervention condition labels into conditional feature vectors; The preset noise vector is input into the preset conditional offset network for shallow feature extraction to obtain the initial hidden layer feature data; In the hidden layer of the pre-defined conditional offset network, the initial hidden layer feature data is fused and modulated using the conditional feature vector to obtain the conditionally modulated hidden layer feature data. The conditional modulation hidden layer feature data is subjected to deep feature inference mapping output to obtain an initial offset vector with a preset dimension consistent with the baseline state vector.

5. The method for enhancing small-sample biomarkers for targeted fusion intervention according to claim 1, characterized in that, After obtaining the physiological sensitivity mask matrix, the following is also included: Extract time-series feature data with timestamp attributes from the physiological biomarker intervention group dataset; Using the Granger causality test algorithm, the causality test F-statistic between the time nodes corresponding to the preset medical intervention condition labels and the corresponding time nodes of the preset dimensions of different physiological marker labels in the time series feature data is calculated, and the causality test F-statistic is used as Granger causality strength data. The Granger causality strength data is compared with a preset causality significance threshold to determine their magnitude. If the Granger causality strength data is greater than the causal significance threshold, the corresponding time point will be recorded as the response mutation point. Calculate the time span between the time node corresponding to the preset medical intervention condition label and the response mutation point; If the time span value is greater than the preset delay threshold, the corresponding physiological marker label dimension is determined to be the lagging physiological marker label dimension, and the time period corresponding to the time span value is defined as the response lag time window. Within the response lag time window, dynamic decay weight data is calculated using the exponential decay function. The corresponding mask coefficients for the lag physiological biomarker label dimension in the physiological sensitivity mask matrix are multiplied with the dynamic decay weight data, and the corresponding mask coefficients are replaced to obtain the corrected physiological sensitivity mask matrix.

6. The method for enhancing small-sample biomarkers for targeted fusion intervention according to claim 3, characterized in that, The preset baseline generation network also includes an update step, specifically including: Calculate the Pearson correlation coefficient between data from different preset dimensions in the physiological marker blank baseline dataset to obtain the Pearson correlation coefficient of the data pairs; If the Pearson correlation coefficient of any data pair is greater than the preset collinearity threshold, then the corresponding physiological marker labels of the preset dimensions are paired and determined as the target physiological dimension pair. Calculate the first covariance matrix of the target physiological dimension in the blank baseline dataset of physiological markers, and use the first covariance matrix as the background covariance feature data; When candidate enhancement biomarker data is input into a preset adversarial discriminant network, the second covariance matrix of the target physiological dimension pairs in the candidate enhancement biomarker data is calculated; Calculate the matrix L2 norm difference between the first covariance matrix and the second covariance matrix, and use the matrix L2 norm difference as the physiological covariance loss data; Obtain the basic generation loss data of the preset baseline generation network; Using the acquired preset weight coefficients, the physiological covariance loss data and the basic generated loss data are weighted and fused to obtain the comprehensive baseline loss data; Based on the comprehensive baseline loss data, the gradient data of the preset baseline generation network weights is calculated using the chain rule; Using a preset optimization algorithm, the network weights of the network generated from the preset baseline are numerically updated iteratively based on gradient data.

7. The method for enhancing small-sample biomarkers for targeted fusion intervention according to claim 1, characterized in that, The specific process for obtaining the target conditional distribution loss data is as follows: During the forward propagation process of candidate enhancement marker data by the pre-defined adversarial discriminant network, the intermediate condition judgment feature matrix output by the hidden layer of the pre-defined adversarial discriminant network is extracted. The intermediate condition judgment feature matrix and the physiological sensitivity mask matrix are multiplied by the Adama product. In the result matrix, the feature elements whose corresponding mask coefficients are lower than the second preset threshold are forced to zero to obtain the mask feature matrix. The masked feature matrix is ​​input into the output layer of the preset adversarial discriminant network, and the cross-entropy loss data is calculated in combination with the preset medical intervention condition labels. The cross-entropy loss data is used as the target conditional distribution loss data.

8. The method for enhancing small-sample biomarkers for targeted fusion intervention according to claim 1, characterized in that, The specific process for acquiring the target enhancement marker data is as follows: Based on the target conditional distribution loss data, the loss gradient data of each network layer node of the preset conditional offset network is calculated using the chain rule. Using a preset optimization algorithm, the network weights of the preset condition offset network are numerically iteratively updated based on the loss gradient data to obtain the updated network weights. The preset noise vector and preset medical intervention condition labels are input into the preset conditional offset network loaded with updated network weights to perform feature secondary inference and obtain the final offset vector. The physiological sensitivity mask matrix and the final offset vector are multiplied element-wise for filtering. Data in dimensions with mask coefficients below the third preset threshold are zeroed out and then superimposed with the baseline state vector to obtain and output target enhancement marker data.

9. A small-sample biomarker enhancement system for fusion-based intervention, characterized in that, include: Data quantification module: used to acquire the original clinical samples of the target object for concentration quantification detection, and obtain the physiological marker blank baseline dataset and the physiological marker intervention group dataset containing the preset medical intervention condition labels; Mask generation module: used to calculate the inter-group distribution difference data between the physiological marker blank baseline dataset and the physiological marker intervention group dataset, and to obtain the physiological sensitivity mask matrix through a preset mapping function; Dual-branch mapping module: used to input the preset noise vector into the preset baseline generation network and the preset conditional offset network respectively for feature mapping, to obtain the baseline state vector and the initial offset vector; The overlay filtering module is used to multiply the initial offset vector element by element using the physiological sensitivity mask matrix, and to zero out the data of the preset dimension of the physiological marker label whose corresponding mask coefficient is lower than the first preset threshold, so as to obtain the target offset vector. The baseline state vector and the target offset vector are overlaid to generate candidate enhancement marker data. The discrimination and update module is used to input candidate enhancement marker data into a preset adversarial discrimination network, filter loss data of preset dimensions with mask coefficients lower than the second preset threshold by combining the physiological sensitivity mask matrix, obtain target conditional distribution loss data, update the network weights of the preset conditional offset network through the backpropagation algorithm, and generate and output target enhancement marker data based on the updated preset conditional offset network. Medical assessment output module: used to expand the preset original medical training set based on target augmentation biomarker data, and to iteratively train the preset clinical fatigue state assessment model to output medical auxiliary assessment results.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-8.

Images