A feature selection method and system

Abstract

This invention discloses a feature selection method and system, relating to the field of medical data technology. The method includes: constructing an initial feature set for patients; initializing an aurora particle population, generating candidate feature subsets, and calculating their fitness values; introducing a dynamic gradient interaction strategy and a multi-dimensional hybrid search strategy to update the continuous feature weight vectors of aurora particles and optimize the population position; performing binarization mapping and fitness evaluation on the updated aurora particle population again, and updating the current best individual and the globally best individual; iterating cyclically until a termination condition is met to obtain the optimal feature subset. This method can select compact and discriminative feature combinations from a multi-source information feature set to improve model interpretability and the stability of prediction results.

Description

Technical Field

[0001] This invention relates to the field of medical data technology, specifically to a feature selection method and system. Background Technology

[0002] With the development of medical artificial intelligence, machine learning, and radiomics technologies, constructing disease diagnostic models using patient clinical data, laboratory test data, and chest CT image texture features has become an important direction for assisting clinical decision-making. However, in actual modeling, the fused clinical data, laboratory test data, and image texture features often exhibit characteristics such as high dimensionality, small sample size, strong correlation, and significant noise. On the one hand, a large number of redundant and irrelevant features increase the model training complexity, reduce the model's generalization ability, and increase the risk of overfitting. On the other hand, complex correlations may exist between multiple clinical indicators, multiple image texture features, and between clinical indicators and image features. Directly inputting all features into the classifier can easily lead to decreased model interpretability and unstable prediction results. Therefore, it is necessary to select a small subset of features with stable discriminative ability from the initial high-dimensional feature set.

[0003] Existing feature selection methods typically fall into three categories: filtering, embedded, and wrapping. Filtering methods select features based on statistical correlation or information content between them and labels, offering high computational efficiency. However, they are usually independent of specific classifiers and struggle to fully reflect higher-order interactions between features. Embedded methods perform feature selection concurrently with model training, such as using regularized models or tree models to select features based on importance scores. However, their selection results are often influenced by specific model structures and parameter settings. Wrapping methods use the predictive performance of the classifier as the evaluation criterion for the quality of feature subsets, more directly reflecting the degree of matching between feature subsets and the classification model. This approach is generally suitable for scenarios like medical diagnosis, where high predictive performance and feature combination effects are crucial. However, wrapping feature selection requires searching through a large number of candidate feature subsets. As the feature dimension increases, the number of possible feature combinations grows exponentially, resulting in a huge search space and high computational costs. To improve the search efficiency of wrapping feature selection, metaheuristic optimization algorithms such as genetic algorithms, particle swarm optimization, gray wolf optimization, Harris Eagle optimization, slime mold optimization, and aurora optimizers are applied to feature selection tasks. Such algorithms can search complex non-convex spaces without relying on gradient information and alleviate the computational pressure of exhaustive search to some extent. The aurora optimizer is a metaheuristic optimization algorithm inspired by the aurora formation process. By simulating the motion of high-energy particles in the Earth's magnetic field and atmosphere, it can achieve a balance between global exploration and local exploitation in continuous optimization problems. However, existing aurora optimizers and other metaheuristic feature selection methods are mainly geared towards continuous optimization problems. When used for binary feature selection, the continuous search trajectory needs to be mapped to discrete decisions of feature "selection" or "non-selection." Existing mapping mechanisms still have limited control over the discrete search of high-dimensional medical features. Secondly, traditional aurora optimizers are easily limited by the overall vector update method in high-dimensional search spaces, making it difficult to effectively adjust key dimensions while maintaining global search capabilities, resulting in insufficient search efficiency and feature subset stability. Moreover, traditional aurora optimizers lack sufficient fine-grained local exploitation mechanisms in the later stages of iteration, making them prone to premature convergence in complex multi-peak search spaces, thus affecting the discriminative performance of the final feature subset.

[0004] In summary, traditional Aurora optimizers are easily limited by the overall vector update method in high-dimensional search spaces, making it difficult to effectively adjust key dimensions while maintaining global search capabilities. As a result, in high-dimensional medical feature spaces, the adjustment range of key feature dimensions is diluted by a large number of non-key dimensions, making it difficult to achieve accurate optimization. This leads to insufficient search efficiency and feature subset stability, thus affecting the discrimination performance of the final feature subset. Summary of the Invention

[0005] To address the shortcomings of existing technologies in high-dimensional search spaces, which are easily limited by the overall vector update method and struggle to effectively adjust key dimensions while maintaining global search capabilities, this invention proposes a feature selection method and system. It constructs key reference points through a dynamic gradient interaction strategy, enhancing the local development capability of the aurora particle population during iteration. Furthermore, it employs a multi-dimensional hybrid search strategy to explore, update, and locally optimize the population, improving the algorithm's search capability and stability in high-dimensional feature spaces, thereby resolving the problems existing in current technologies.

[0006] A feature selection method includes the following steps: Obtain an initial feature set consisting of multiple samples and their corresponding feature vectors and class labels; Based on the total number of features in the initial feature set, a population of aurora particles is randomly initialized in the continuous search space. Each aurora particle in the population represents a candidate feature selection scheme. The initial aurora particle population is binarized to generate a subset of candidate features, and the fitness value of the subset of candidate features is calculated to determine the current global optimal position. Based on the average position of the current aurora particle population and the current position of the target aurora particle, a gradient proxy vector is determined. The first and second moment estimates of each aurora particle are updated based on the gradient proxy vector, and bias correction is applied to the updated first and second moment estimates. Key reference points are constructed based on the current global optimal position and the bias-corrected moment estimates. The continuous feature weight vector of the target aurora particle is updated based on the key reference points, the current position of the target aurora particle, and the positions of random neighboring aurora particles. The updated aurora particle population is then explored, updated, and locally optimized using a multi-dimensional hybrid search strategy to obtain the explored and optimized aurora particle population. The optimized aurora particle population is binarized and its fitness is evaluated. The current best individual and the global best individual are updated. The process is iterated until the termination condition is met, and the optimal feature subset is output.

[0007] Furthermore, the binarization mapping of the initial aurora particle population to generate a candidate feature subset specifically includes the following steps: A V-shaped binarization mapping function is used to map the continuous feature weight values ​​of each aurora particle to the feature selection probability; Based on a preset random number, if the random number is less than the selection probability of the feature, the binary code value of that feature is set to 1, indicating that the feature is selected; otherwise, it is set to 0, indicating that the feature is not selected. The binary encoding values ​​of each aurora particle are summarized to obtain the binary feature selection vector; Based on the features corresponding to the dimensions with a value of 1 in the binary feature selection vector, a subset of candidate features is obtained.

[0008] Further, the calculation of the fitness value of the candidate feature subset includes the following steps: A subset of candidate features is input into a support vector machine classifier, and the classification error rate is calculated using cross-validation. Determine the ratio of the number of features selected in the candidate feature subset to the total number of features in the initial feature set; The classification error rate and the ratio are assigned weight coefficients respectively, and the weighted sum is used as the fitness value, wherein the weight coefficient of the classification error rate is greater than the weight coefficient of the feature quantity ratio.

[0009] Furthermore, the gradient proxy vector is determined by the difference between the average position of the current aurora particle population and the current position of the target aurora particle.

[0010] Furthermore, the first-order moment estimate is updated by the weighted sum of the first-order moment estimate from the previous iteration and the current gradient surrogate vector, and the second-order moment estimate is updated by the weighted sum of the second-order moment estimate from the previous iteration and the square of the current gradient surrogate vector, wherein the initial values ​​of both the first-order moment estimate and the second-order moment estimate are zero.

[0011] Furthermore, the bias correction is performed by dividing the current first-order moment estimate and second-order moment estimate by a power-law compensation term of the attenuation coefficient, respectively.

[0012] Furthermore, the key reference point is obtained by subtracting the product of the bias-corrected ratio of the moment estimates and the learning rate from the current position of the globally optimal aurora particle, as expressed below: in, Indicates the first During the nth iteration A key reference point for aurora particles This indicates the current position of the globally optimal aurora particle. Indicates the learning rate. This indicates the smallest positive number whose denominator is zero. It is the first-order moment estimate after bias correction. It is the second-order moment estimate after bias correction.

[0013] Furthermore, the exploration, updating, and local optimization based on the multidimensional hybrid search strategy specifically includes: The total number of features in the initial feature set is compared with the preset dimension threshold. If the total number of features is greater than or equal to the preset dimension threshold, multiple dimensions are randomly selected from the continuous feature weight vector. Based on the difference between the current global optimal position and the target aurora particle in that dimension, combined with control parameters that change with the number of iterations and perturbation factors generated based on sine or cosine functions, a multi-dimensional search update is performed on the multiple dimensions. If the total number of features is less than the preset dimension threshold, a single dimension is selected from the continuous feature weight vector. Based on the difference information of two randomly selected different aurora particles in that dimension, combined with random weight coefficients, a single-dimensional search update is performed on that dimension. During the local optimization process, the position of the target aurora particle is updated based on the distance vector between the current global optimal position and multiple randomly selected aurora particles.

[0014] Furthermore, the updating of the current best individual and the global best individual adopts a greedy selection strategy, specifically including: comparing the fitness value of the updated aurora particle with the fitness value before the update; if the updated fitness value is better, then the updated aurora particle is retained, otherwise the aurora particle before the update is retained; comparing the fitness value of the updated aurora particle with the fitness value of the current global best aurora particle; if the updated fitness value is better, then the updated aurora particle is determined as the new global best individual.

[0015] The present invention also includes a feature selection system, comprising: The acquisition module is used to acquire an initial feature set consisting of multiple samples and their corresponding feature vectors and class labels; The initialization module is used to randomly initialize an aurora particle population in a continuous search space based on the total number of features in the initial feature set. Each aurora particle in the population represents a candidate feature selection scheme. The initial aurora particle population is binarized to generate a subset of candidate features, and the fitness value of the subset of candidate features is calculated to determine the current global optimal position. The optimization module is used to determine a gradient proxy vector based on the average position of the current aurora particle population and the current position of the target aurora particle; update the first-order moment estimate and second-order moment estimate of each aurora particle based on the gradient proxy vector, and perform bias correction on the updated first-order moment estimate and second-order moment estimate; construct key reference points based on the current global optimal position and the bias-corrected moment estimates; update the continuous feature weight vector of the target aurora particle based on the key reference points, the current position of the target aurora particle, and the positions of random neighboring aurora particles; and perform exploratory updates and local optimizations on the updated aurora particle population based on a multi-dimensional hybrid search strategy to obtain an exploratory and optimized aurora particle population. The output module is used to perform binarization mapping and fitness evaluation on the explored and optimized aurora particle population, and update the current best individual and the global best individual; it iterates until the termination condition is met, and outputs the optimal feature subset.

[0016] This invention provides a feature selection method, which has the following beneficial effects: This invention employs a dynamic gradient interaction strategy to construct a gradient proxy vector based on the population average position and the current position of the target particle. This vector is then used to update the first and second moment estimates. After bias correction, it is combined with the current global optimal position to construct a key reference point. This allows the continuous feature weight vector of the target particle to be adaptively adjusted along the key discriminative dimension, overcoming the drawback of traditional aurora optimizers where the adjustment range of key feature dimensions is diluted by numerous non-critical clinical indicators and imaging features in the overall vector update method. Simultaneously, a multi-dimensional hybrid search strategy is used for differentiated exploration updates and local optimization, enhancing the fine-grained search capability near the optimal feature combination in the later stages of iteration. This effectively avoids missing feature pairs with collaborative discriminative power due to premature convergence. Finally, from a high-dimensional medical feature set that integrates multiple samples and their corresponding features, a compact, stable, and clinically interpretable optimal feature subset is selected, significantly improving the prediction accuracy and interpretability of the classification model. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the method for selecting diagnostic features for pulmonary aspergillosis and pulmonary cryptococcosis in an embodiment of the present invention. Figure 2 This is a flowchart of the algorithm for the multidimensional gradient interactive search aurora optimizer in an embodiment of the present invention; Figure 3 This is a framework diagram for the differential diagnosis of pulmonary aspergillosis and pulmonary cryptococcosis based on a binary multidimensional gradient interactive search aurora optimizer and support vector machine in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0019] This invention proposes a feature selection method, comprising the following steps: S1. Obtain an initial feature set consisting of multiple samples and their corresponding feature vectors and class labels.

[0020] S2. Based on the total number of features in the initial feature set, randomly initialize the aurora particle population in the continuous search space. Each aurora particle in the population represents a candidate feature selection scheme. Perform binarization mapping on the initial aurora particle population to generate a candidate feature subset and calculate the fitness value of the candidate feature subset to determine the current global optimal position.

[0021] The initial aurora particle population is binarized to generate a subset of candidate features. This process includes the following steps: A V-shaped binarization mapping function is used to map continuous feature weights to feature selection probabilities; based on a preset random number, if the random number is less than the feature selection probability, the binary code value of that feature dimension is set to 1, indicating that the feature is selected; otherwise, it is set to 0, indicating that the feature is not selected; the binary code values ​​of each dimension of each aurora particle are summarized to obtain a binary feature selection vector; and the subset of candidate features is obtained based on the features corresponding to the dimensions with a value of 1 in the binary feature selection vector.

[0022] The fitness value calculation process includes: inputting the candidate feature subset into the support vector machine classifier and calculating the classification error rate using cross-validation; determining the ratio of the number of selected features in the candidate feature subset to the total number of features in the initial feature set; assigning weight coefficients to the classification error rate and the ratio respectively, and using the weighted sum as the fitness value, wherein the weight coefficient of the classification error rate is greater than the weight coefficient of the feature ratio.

[0023] S3. Based on the average position of the current aurora particle population and the current position of the target aurora particle, determine the gradient surrogate vector; update the first-order moment estimate and second-order moment estimate of each aurora particle based on the gradient surrogate vector; perform bias correction on the first-order moment estimate and second-order moment estimate; construct key reference points based on the current global optimal position and the bias-corrected moment estimates; update the continuous feature weight vector of the target aurora particle based on the key reference points, the current position of the target aurora particle, and the positions of random neighboring aurora particles; for the updated aurora particle population, perform exploration updates and local optimization based on a multi-dimensional hybrid search strategy to obtain the explored and optimized aurora particle population.

[0024] Preferably, the gradient proxy vector is determined by the difference between the average position of the current aurora particle population and the current position of the target aurora particle.

[0025] Preferably, the first-order moment estimate is updated by the weighted sum of the first-order moment estimate from the previous iteration and the current gradient surrogate vector, and the second-order moment estimate is updated by the weighted sum of the second-order moment estimate from the previous iteration and the square of the current gradient surrogate vector, wherein the initial values ​​of the first-order moment estimate and the second-order moment estimate are both zero.

[0026] Preferably, the bias correction is performed by dividing the current first-order moment estimate and second-order moment estimate by a power-law compensation term of the attenuation coefficient.

[0027] Preferably, the exploration update and local optimization are based on a multi-dimensional hybrid search strategy, specifically including: comparing the total number of features in the initial feature set with the size of a preset dimension threshold; if the total number of features is greater than or equal to the preset dimension threshold, then multiple dimensions are randomly selected from the continuous feature weight vector, and based on the difference between the current global optimal position and the target aurora particle in that dimension, combined with control parameters that change with the number of iterations and a perturbation factor generated based on a sine or cosine function, the multiple dimensions are searched and updated in a multi-dimensional way; if the total number of features is less than the preset dimension threshold, then a single dimension is selected from the continuous feature weight vector, and based on the difference information of two randomly selected different aurora particles in that dimension, combined with random weight coefficients, the dimension is searched and updated in a single-dimensional way; during the local optimization process, the position of the target aurora particle is updated based on the distance vector between the current global optimal position and the randomly selected multiple aurora particles.

[0028] Preferably, the key reference point is obtained by subtracting the product of the moment estimation ratio after bias correction and the learning rate from the position of the current global best aurora particle.

[0029] S4. Perform binarization mapping and fitness evaluation on the optimized aurora particle population, and update the current best individual and the global best individual; iterate until the termination condition is met, and output the optimal feature subset.

[0030] Preferably, the update of the current best individual and the global best individual adopts a greedy selection strategy, which specifically includes: comparing the fitness value of the updated aurora particle with the fitness value before the update; if the updated fitness value is better, then the updated aurora particle is retained, otherwise the aurora particle before the update is retained; comparing the fitness value of the updated aurora particle with the fitness value of the current global best aurora particle; if the updated fitness value is better, then the updated aurora particle is determined as the new global best individual.

[0031] This invention constructs an initial feature set containing multi-source information, enabling a more comprehensive characterization of the differences between two types of patients; it converts the continuous feature weight vector of aurora particles into a binary feature selection vector through a binarization mapping function, allowing the multi-dimensional gradient interactive search aurora optimizer to adapt to discrete feature selection tasks; it constructs key reference points through a dynamic gradient interaction strategy, improving the local exploitation capability of the aurora particle population during the iteration process; it explores, updates, and locally optimizes the population through a multi-dimensional hybrid search strategy, improving the algorithm's search capability and stability in high-dimensional feature spaces; and it uses a support vector machine classifier and a fitness function that balances classification error rate and feature quantity to perform a packaged evaluation of candidate feature subsets, thereby obtaining a compact, stable, and discriminative optimal feature subset.

[0032] Based on the above inventive concept, the present invention proposes the following embodiments: Pulmonary aspergillosis and pulmonary cryptococcosis are both common pulmonary fungal infections, and they share significant overlap in clinical symptoms, laboratory test results, and chest CT imaging findings. For example, both diseases can present as pulmonary nodules, masses, or consolidation, making stable and accurate differential diagnosis difficult to achieve solely through manual image interpretation or a single laboratory indicator. While pathological examination, pathogen culture, and fungal antigen detection can be used for definitive diagnosis, some of these procedures are invasive, time-consuming, or highly dependent on sample quality, thus failing to meet the needs for rapid auxiliary differential diagnosis.

[0033] With the development of medical artificial intelligence and machine learning technologies, constructing a differential diagnostic model for pulmonary fungal infections using patient clinical data, laboratory test data, and chest CT image texture features has become a feasible auxiliary diagnostic method. However, the fusion of these multi-source data usually results in a high-dimensional feature set, which may contain a large number of redundant, irrelevant, and strongly correlated features. Directly inputting all features into a classifier can easily lead to model overfitting, decreased generalization ability, and insufficient interpretability. Therefore, a method is needed to automatically select a compact, stable, and discriminative feature subset from high-dimensional medical features.

[0034] Based on this, this invention proposes a feature selection method for the differential diagnosis of pulmonary aspergillosis and pulmonary cryptococcosis. This method uses a multidimensional gradient interactive search aurora optimizer to search a high-dimensional feature space, and converts continuous search results into discrete feature selection results using a binarization mapping function. Then, a support vector machine classifier is used to perform a packaged evaluation of the candidate feature subset, thereby obtaining the optimal feature subset and diagnostic model for the differential diagnosis of pulmonary aspergillosis and pulmonary cryptococcosis. This scheme corresponds to the MGSPLO algorithm, bMGSPLO-SVM diagnostic framework, clinical and imaging feature construction methods, and model comparison experimental results in the original content.

[0035] like Figure 1 As shown, the method specifically includes the following steps: S1. Collect clinical data, laboratory test data, and chest CT imaging data of patients with pulmonary aspergillosis and pulmonary cryptococcosis to construct an initial feature set.

[0036] Specifically, clinical data, laboratory test data, and chest CT imaging data of patients with pulmonary aspergillosis and pulmonary cryptococcosis were collected to construct an initial feature set.

[0037] In one specific embodiment, patients with pulmonary aspergillosis and pulmonary cryptococcosis can be diagnosed through pathological examination, pathogen culture, blood culture, cryptococcal capsular antigen detection, or Aspergillus antigen detection. Clinical data includes patient age, hypertension, diabetes, and other information; laboratory test data includes one or more of the following: C-reactive protein, white blood cell count, neutrophil percentage, eosinophil percentage, basophil percentage, monocyte percentage, lymphocyte percentage, absolute eosinophil count, absolute neutrophil count, absolute monocyte count, absolute lymphocyte count, absolute basophil count, red blood cell count, hemoglobin, hematocrit, mean corpuscular volume, mean corpuscular hemoglobin content, mean corpuscular hemoglobin concentration, red blood cell distribution width, standard deviation of red blood cell distribution width, platelet count, plateletcrit, mean platelet volume, platelet distribution width, and large platelet ratio.

[0038] In one specific embodiment, chest CT image data is acquired using a CT scanning device. The patient can be in a supine position during the scan, and the whole-lung CT scan is completed while holding their breath at the end of inspiration. Scanning parameters may include a tube voltage of 120kV, a tube current of 200mAs, a slice thickness of 5mm, a slice spacing of 5mm, and an image matrix size of 512×512. These parameters are merely one possible implementation and do not constitute a limitation on the scope of protection of this invention.

[0039] For chest CT image data, the voxel intensity values ​​are first converted to Henle units and truncated according to the preset lung window range; then, the truncated chest CT image data is normalized; further, the chest CT images and corresponding lesion masks are resampled to reduce the impact of anisotropic resolution on the calculation of image texture features.

[0040] During region-of-interest (ROI) segmentation, radiologists can delineate the ROI of lung lesions on chest CT images. For multiple slices containing lesions, the lesion outlines are marked separately, and a three-dimensional ROI mask is generated. If there are discrepancies in the lesion boundaries, the final ROI can be determined through review or consensus discussion.

[0041] After obtaining the region of interest (ROI), the local binary pattern operator is used to encode the lesion texture within the ROI. Specifically, a rotation-invariant uniform local binary pattern can be used, with radius and sampling point parameters set to encode the texture structure within the ROI. Then, the ROI encoding results are summarized using histogram statistics, and the binning ratio of each histogram is used as the image texture feature.

[0042] Finally, the image texture features are fused with clinical data and laboratory test data to form an initial feature set. Each patient in the initial feature set corresponds to a feature vector, along with a corresponding diagnostic label, which indicates whether the patient has pulmonary aspergillosis or pulmonary cryptococcosis.

[0043] S2. Divide the initial feature set into a training set and a test set, and initialize the aurora particle population in the continuous search space.

[0044] The training set is used for feature selection and fitness evaluation, while the test set is used for the final validation of the diagnostic model.

[0045] Let the feature dimension of the initial feature set be... The aurora particle population size is The maximum number of iterations is In the first During the nth iteration, the 1st The continuous feature weight vector of each aurora particle is denoted as:

[0046] in, Indicates the first During the nth iteration The aurora particles in the first Continuous weight values ​​on dimensional features, .

[0047] Random initialization in a continuous search space We obtain an initial aurora particle population by generating a number of aurora particles. The continuous feature weight vector for each aurora particle represents a potential feature selection scheme. Since feature selection is a binary discrete decision problem, we subsequently need to convert the continuous feature weight vector into a binary feature selection vector using a binarization mapping function.

[0048] S3. Binarize the initial aurora particle population to generate a subset of candidate features, and calculate the fitness value based on SVM.

[0049] Specifically, a V-shaped binarization mapping function is used to map continuous feature weight values ​​to feature selection probabilities. For the th During the nth iteration The first aurora particle The weight values ​​of continuous features are determined based on the following formula to determine the feature selection probability:

[0050] in, Represents the V-shaped binary mapping function. This represents the error function.

[0051] Then, the first is determined based on the following formula. During the nth iteration The aurora particles in the first Binary encoded values ​​on dimensional features: in, Indicates the first During the nth iteration The aurora particles in the first Binary encoded values ​​on dimensional features, The range is indicated in A random number within the range. When When, it indicates the selection of the first. One characteristic; when When, it means not to select the first option. One characteristic.

[0052] The first The summation of the binary encoding values ​​of each aurora particle yields the binary feature selection vector: Based on the features corresponding to dimensions with a value of 1 in the binary feature selection vector, a subset of candidate features is obtained. .

[0053] For each candidate feature subset The data is then input into a support vector machine classifier to obtain classification prediction results for pulmonary aspergillosis and pulmonary cryptococcosis. The support vector machine classifier distinguishes between the two types of patients by constructing an optimal classification hyperplane, making it suitable for handling small sample and high-dimensional medical data scenarios.

[0054] In one specific embodiment, cross-validation is used to calculate the classification error rate of the candidate feature subset. The classification error rate is determined based on the following formula:

[0055] in, Indicates the first During the nth iteration The classification error rate of a subset of candidate features. This represents the number of true positive samples. This represents the number of true negative samples. Indicates the number of false positive samples. This indicates the number of false negative samples.

[0056] Then, the fitness value of the candidate feature subset is calculated based on the following formula: in, Indicates the first During the nth iteration The fitness value of each candidate feature subset. This indicates the number of features selected from the candidate feature subset. This represents the total number of features in the initial feature set. and Let be the weighting coefficient, and satisfy: In medical diagnostic tasks, classification error rate is of high importance, therefore, it can be set... Greater than This approach prioritizes reducing diagnostic errors during the optimization process while appropriately penalizing the number of features. A smaller fitness value indicates a better subset of candidate features. Based on the fitness values ​​of each aurora particle, the initial aurora particle population is sorted, and the current best individual and the global best individual are determined.

[0057] S4. Construct key reference points based on the dynamic gradient interaction strategy and update the continuous feature weight vector of aurora particles.

[0058] The original aurora optimizer performs optimization searches by simulating the motion of high-energy particles in the Earth's magnetic field and atmospheric environment, including mechanisms such as cyclotron motion, aurora egg walking, and particle collisions. This algorithm has a certain global exploration capability in continuous optimization problems, but in complex high-dimensional feature selection scenarios, it is prone to insufficient development capability and premature convergence issues in the later stages of iteration.

[0059] To improve the algorithm's local exploitation capability in the later stages of iteration, this embodiment of the invention designs a dynamic gradient interaction strategy. For the... During the nth iteration The aurora particle, firstly based on the average position of the current aurora particle population and the... The current position of each aurora particle determines the gradient proxy vector:

[0060] in, Indicates the first During the nth iteration Gradient proxy vectors of auroral particles Indicates the first The average position of the aurora particle population at the next iteration. Indicates the first During the nth iteration The current position of each aurora particle.

[0061] Then, the first-order moment estimate and the second-order moment estimate are updated based on the gradient surrogate vector: in, Indicates the first During the nth iteration First-order moment estimation of an aurora particle Indicates the first During the nth iteration Second-order moment estimation of auroral particles and This represents the attenuation rate parameter.

[0062] To reduce initialization bias, bias corrections are applied to the first-order moment estimates and second-order moment estimates.

[0063] Based on the current globally optimal position and the moment estimate after bias correction, key reference points are constructed: in, Indicates the first During the nth iteration A key reference point for aurora particles This indicates the current position of the globally optimal aurora particle. Indicates the learning rate. This represents the smallest positive number that prevents the denominator from being zero.

[0064] After obtaining the key reference point, the continuous feature weight vector of the target aurora particle is updated based on the key reference point, the current position of the target aurora particle, and the positions of random neighboring aurora particles. Thus, the dynamic gradient interaction strategy can utilize current population distribution information and historical search information to improve the convergence accuracy and local exploitation capability of the aurora particle population in high-dimensional complex spaces.

[0065] S5. Based on a multi-dimensional hybrid search strategy, the updated aurora particle population is explored, updated, and locally optimized.

[0066] Specifically, when the feature dimension of the initial feature set Greater than or equal to the preset dimension threshold At that time, multiple dimensions are randomly selected from the continuous feature weight vector, and a multi-dimensional search update is performed on these multiple dimensions based on the current global optimal position. The multi-dimensional search update is performed based on the following formula: in, Indicates the dimension selected randomly. Indicates the updated number 2D continuous feature weight values, This indicates that the current globally optimal aurora particle is at the [number]th [position]. Dimensional continuous feature weight values, This represents the control parameters that change with the number of iterations. This represents the perturbation factor generated based on the sine or cosine function.

[0067] When the feature dimension of the initial feature set Less than the preset dimension threshold At that time, a single dimension is selected from the continuous feature weight vector, and a one-dimensional search update is performed based on the difference information between two randomly selected aurora particles. The one-dimensional search update is performed based on the following formula:

[0068] in and This indicates the numbers of two different aurora particles randomly selected from the aurora particle population. and These represent the two aurora particles at the [number]th [time]. Dimensional continuous feature weight values, This represents the random weighting coefficient.

[0069] Furthermore, during the local optimization process, the target aurora particle's position is updated based on the distance vector between the current global optimal position and multiple randomly selected aurora particles, so that the target aurora particle can be guided from multiple directions and maintain a certain degree of randomness during the local development process.

[0070] Through the above-mentioned multidimensional hybrid search strategy, the present invention can take into account the search breadth, search efficiency and local optimization capability in a high-dimensional feature space, and avoid inefficient overall updates of all dimensions in each iteration.

[0071] S6. Perform binarization mapping and fitness evaluation on the updated aurora particle population again, and update the current best individual.

[0072] Specifically, boundary processing is performed on each aurora particle in the updated aurora particle population to ensure that the continuous feature weight vector of the aurora particle lies within a preset search boundary. Then, the binarization mapping in step S3 is performed again on the boundary-processed continuous feature weight vector to obtain the updated binary feature selection vector.

[0073] Based on the updated binary feature selection vector, the features corresponding to the dimensions with a value of 1 are used to generate an updated subset of candidate features. Subsequently, the fitness values ​​of the updated subset of candidate features are recalculated using a support vector machine classifier.

[0074] Furthermore, a greedy selection strategy is used to update the aurora particle population. If the fitness value of the updated aurora particle is better than that of the unupdated aurora particle, the updated aurora particle is retained; otherwise, the unupdated aurora particle is retained. If the fitness value of the updated aurora particle is better than that of the current global best aurora particle, the updated aurora particle is determined as the new current global best individual.

[0075] S7. Iterate until the termination condition is met, output the optimal feature subset, and construct a differential diagnostic model for pulmonary aspergillosis and pulmonary cryptococcosis.

[0076] Repeat steps S4 to S6 until a termination condition is met. Termination conditions include reaching the maximum number of iterations, reaching the maximum number of function evaluations, or the optimal fitness improvement falling below a preset threshold for several consecutive iterations.

[0077] After the iteration terminates, the features corresponding to the dimensions with a value of 1 in the binary feature selection vector corresponding to the current globally optimal aurora particle are determined as the optimal feature subset. Then, the support vector machine classifier is retrained based on the optimal feature subset to obtain a differential diagnostic model for pulmonary aspergillosis and pulmonary cryptococcosis.

[0078] For patients awaiting diagnosis, their clinical data, laboratory test data, and chest CT image data are obtained. Image texture features are extracted according to step S1, and a patient feature vector is constructed. Subsequently, corresponding features are selected from the patient feature vector based on the optimal feature subset, and input into the differential diagnostic model for pulmonary aspergillosis and pulmonary cryptococcosis, outputting the auxiliary differential diagnosis results for pulmonary aspergillosis or pulmonary cryptococcosis.

[0079] like Figure 1 As shown, the method provided in this embodiment of the invention is executed sequentially according to S1 to S7, realizing a complete process from multi-source data acquisition, initial feature set construction, aurora particle population initialization, binarization mapping, fitness evaluation, continuous feature weight vector update to final diagnostic model construction.

[0080] like Figure 2 As shown, the multidimensional gradient interactive search aurora optimizer iteratively updates the aurora particle population through a dynamic gradient interaction strategy and a multidimensional hybrid search strategy, thereby improving its search capability and stability in high-dimensional complex feature spaces.

[0081] like Figure 3 As shown, the diagnostic framework based on the binarized multidimensional gradient interactive search aurora optimizer and support vector machine includes multi-source data input, feature fusion, bMGSPLO feature selection, SVM fitness evaluation, and final diagnostic model output.

[0082] Implementation Results Description In one specific embodiment, to verify the effectiveness of the present invention, the proposed bMGSPLO-SVM method is compared with other binary metaheuristic wrapper methods and commonly used machine learning classification models. The comparison methods include bPSO-SVM, Eagle-optimized support vector machine, Turtle-optimized support vector machine, Snake-optimized support vector machine, LightGBM, XGBoost, kernel extreme learning machine, and ordinary support vector machine. All methods use the same data partitioning method and cross-validation protocol, and accuracy, recall, precision, and F1 score are used as evaluation metrics.

[0083] Table 1 Performance comparison of bMGSPLO-SVM with other methods in ten independent runs. As shown in Table 1, bMGSPLO-SVM achieved superior overall performance in terms of accuracy, precision, and F1 score, indicating that the present invention can screen out a more discriminative subset of features from highly similar clinical, laboratory, and imaging features, thereby improving the differential diagnostic performance between pulmonary aspergillosis and pulmonary cryptococcosis.

[0084] Table 2. Ranking of the top 10 clinical and imaging features by frequency of selection As shown in Table 2, age, hemoglobin, standard deviation of red blood cell distribution width, red blood cell count, red blood cell distribution width, C-reactive protein, and some local binary pattern image texture features have a high selection frequency, indicating that the present invention can not only improve classification performance, but also output stable feature combinations with clinical interpretability.

[0085] The present invention has the following beneficial effects: This invention addresses the issues of similar clinical and imaging manifestations, excessive redundancy of high-dimensional features, and insufficient stability of traditional feature selection methods in the differential diagnosis of pulmonary aspergillosis and pulmonary cryptococcosis. It proposes a packaged feature selection method based on a multidimensional gradient interactive search aurora optimizer and support vector machine.

[0086] This invention constructs an initial feature set containing multi-source information by fusing clinical data, laboratory test data, and chest CT imaging data from patients with pulmonary aspergillosis and pulmonary cryptococcosis, enabling a more comprehensive characterization of the differences between the two types of pulmonary fungal infections. A V-shaped binarization mapping function is used to convert the continuous feature weight vector of the aurora particles into a binary feature selection vector, allowing the multi-dimensional gradient interactive search aurora optimizer to adapt to discrete feature selection tasks. A dynamic gradient interaction strategy is used to construct key reference points, improving the local exploitation capability of the aurora particle population during iteration. A multi-dimensional hybrid search strategy is used to explore, update, and locally optimize the population, improving the algorithm's search capability and stability in high-dimensional feature spaces. A support vector machine classifier and a fitness function that balances classification error rate and feature quantity are used to perform a packaged evaluation of candidate feature subsets, thereby obtaining a compact, stable, and discriminative optimal feature subset, improving the accuracy, stability, and interpretability of the differential diagnostic model for pulmonary aspergillosis and pulmonary cryptococcosis.

[0087] Based on the same inventive concept, this invention also proposes a feature selection system, comprising: The acquisition module is used to obtain an initial feature set consisting of multiple samples and their corresponding feature vectors and class labels.

[0088] The initialization module is used to randomly initialize an aurora particle population in a continuous search space based on the total number of features in the initial feature set. Each aurora particle in the population represents a candidate feature selection scheme. The initial aurora particle population is binarized to generate a subset of candidate features, and the fitness value of the subset of candidate features is calculated to determine the current global optimal position.

[0089] The optimization module is used to determine the gradient surrogate vector based on the average position of the current aurora particle population and the current position of the target aurora particle; update the first-order moment estimate and second-order moment estimate of each aurora particle based on the gradient surrogate vector; perform bias correction on the first-order moment estimate and second-order moment estimate; construct key reference points based on the current global optimal position and the bias-corrected moment estimates; update the continuous feature weight vector of the target aurora particle based on the key reference points, the current position of the target aurora particle, and the positions of random neighboring aurora particles; and perform exploratory updates and local optimization on the updated aurora particle population based on a multi-dimensional hybrid search strategy to obtain the exploratory-optimized aurora particle population.

[0090] The output module is used to perform binarization mapping and fitness evaluation on the explored and optimized aurora particle population, and update the current best individual and the global best individual; it iterates until the termination condition is met, and outputs the optimal feature subset.

[0091] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A feature selection method characterized by, Includes the following steps: Obtain an initial feature set consisting of multiple samples and their corresponding feature vectors and class labels; Based on the total number of features in the initial feature set, a population of aurora particles is randomly initialized in the continuous search space. Each aurora particle in the population represents a candidate feature selection scheme. The initial aurora particle population is binarized to generate a subset of candidate features, and the fitness value of the subset of candidate features is calculated to determine the current global optimal position. Based on the average position of the current aurora particle population and the current position of the target aurora particle, a gradient proxy vector is determined; the first-order moment estimate and second-order moment estimate of each aurora particle are updated based on the gradient proxy vector, and the updated first-order moment estimate and second-order moment estimate are biased and corrected; based on the current global optimal position and the biased and corrected moment estimate, a key reference point is constructed. Based on the key reference point, the current position of the target aurora particle, and the positions of random neighboring aurora particles, the continuous feature weight vector of the target aurora particle is updated; the updated aurora particle population is then explored, updated, and locally optimized using a multi-dimensional hybrid search strategy to obtain the explored and optimized aurora particle population. The optimized aurora particle population is binarized and its fitness is evaluated. The current best individual and the global best individual are updated. The process is iterated until the termination condition is met, and the optimal feature subset is output.

2. The feature selection method of claim 1, wherein, The process of binarizing the initial aurora particle population to generate a candidate feature subset includes the following steps: A V-shaped binarization mapping function is used to map the continuous feature weight values ​​of each aurora particle to the feature selection probability; Based on a preset random number, if the random number is less than the selection probability of the feature, the binary code value of the feature is set to 1, indicating that the feature is selected; otherwise, it is set to 0, indicating that the feature is not selected. The binary encoding values ​​of each aurora particle are summarized to obtain the binary feature selection vector; Based on the features corresponding to the dimensions with a value of 1 in the binary feature selection vector, a subset of candidate features is obtained.

3. The feature selection method of claim 1, wherein, The calculation of the fitness value of the candidate feature subset includes the following steps: A subset of candidate features is input into a support vector machine classifier, and the classification error rate is calculated using cross-validation. Determine the ratio of the number of features selected in the candidate feature subset to the total number of features in the initial feature set; The classification error rate and the ratio are assigned weight coefficients respectively, and the weighted sum is used as the fitness value, wherein the weight coefficient of the classification error rate is greater than the weight coefficient of the feature quantity ratio.

4. The feature selection method of claim 1, wherein, The gradient proxy vector is determined by the difference between the average position of the current aurora particle population and the current position of the target aurora particle.

5. The feature selection method of claim 1, wherein, The first-order moment estimate is updated by the weighted sum of the first-order moment estimate from the previous iteration and the current gradient surrogate vector, and the second-order moment estimate is updated by the weighted sum of the second-order moment estimate from the previous iteration and the square of the current gradient surrogate vector, wherein the initial values ​​of the first-order moment estimate and the second-order moment estimate are both zero.

6. The feature selection method of claim 1, wherein, The bias correction is achieved by dividing the current first-order moment estimate and second-order moment estimate by a power-law compensation term of the attenuation coefficient.

7. The feature selection method of claim 1, wherein, The key reference point is obtained by subtracting the product of the bias-corrected ratio of the moment estimate and the learning rate from the current position of the globally optimal aurora particle, and is expressed as follows: in, Indicates the first During the nth iteration A key reference point for aurora particles This indicates the position of the current globally optimal aurora particle. Indicates the learning rate. This indicates the smallest positive number whose denominator is zero. It is the first-order moment estimate after bias correction. It is the second-order moment estimate after bias correction.

8. The feature selection method according to claim 1, characterized in that, The exploration, updating, and local optimization based on the multidimensional hybrid search strategy specifically includes: The total number of features in the initial feature set is compared with the preset dimension threshold. If the total number of features is greater than or equal to the preset dimension threshold, multiple dimensions are randomly selected from the continuous feature weight vector. Based on the difference between the current global optimal position and the target aurora particle in that dimension, combined with control parameters that change with the number of iterations and perturbation factors generated based on sine or cosine functions, a multi-dimensional search update is performed on the multiple dimensions. If the total number of features is less than the preset dimension threshold, a single dimension is selected from the continuous feature weight vector. Based on the difference information of two randomly selected different aurora particles in that dimension, combined with random weight coefficients, a single-dimensional search update is performed on that dimension. During the local optimization process, the position of the target aurora particle is updated based on the distance vector between the current global optimal position and multiple randomly selected aurora particles.

9. The feature selection method according to claim 1, characterized in that, The updating of the current best individual and the global best individual adopts a greedy selection strategy, which specifically includes: comparing the fitness value of the updated aurora particle with the fitness value before the update; if the updated fitness value is better, the updated aurora particle is retained; otherwise, the aurora particle before the update is retained; comparing the fitness value of the updated aurora particle with the fitness value of the current global best aurora particle; if the updated fitness value is better, the updated aurora particle is determined as the new global best individual.

10. A feature selection system, characterized in that, include: The acquisition module is used to acquire an initial feature set consisting of multiple samples and their corresponding feature vectors and class labels; The initialization module is used to randomly initialize an aurora particle population in a continuous search space based on the total number of features in the initial feature set. Each aurora particle in the population represents a candidate feature selection scheme. The initial aurora particle population is binarized to generate a subset of candidate features, and the fitness value of the subset of candidate features is calculated to determine the current global optimal position. The optimization module is used to determine the gradient proxy vector based on the average position of the current aurora particle population and the current position of the target aurora particle; update the first-order moment estimate and second-order moment estimate of each aurora particle based on the gradient proxy vector, and perform bias correction on the updated first-order moment estimate and second-order moment estimate; and construct key reference points based on the current global optimal position and the bias-corrected moment estimate. Based on the key reference point, the current position of the target aurora particle, and the positions of random neighboring aurora particles, the continuous feature weight vector of the target aurora particle is updated; the updated aurora particle population is then explored, updated, and locally optimized using a multi-dimensional hybrid search strategy to obtain the explored and optimized aurora particle population. The output module is used to perform binarization mapping and fitness evaluation on the explored and optimized aurora particle population, and update the current best individual and the global best individual; it iterates until the termination condition is met, and outputs the optimal feature subset.

Images