A feature selection method and device for predicting preeclampsia risk of pregnant women

Abstract

The application discloses a feature selection method and device for predicting preeclampsia risk of pregnant women, and relates to the technical field of data processing.The application uses an adaptive network topology strategy to construct a topology structure, and dynamically constructs sparse adjacency relations between individuals based on fitness and similarity to generate a topology population.The topology structure reflects the internal interaction correlation between features, can guide the algorithm to perform self-organizing directional search according to a search state, and realizes the unified balance of exploration and development.Then, a positive sine search mechanism is used to perform periodic jump search between global search and local search under the guidance of a global optimal solution, and a multi-level information interaction strategy is used, global items promote convergence to the optimal solution, local items strengthen mining of potential local areas, global information dissemination and local neighborhood learning are fused and unified, and dynamic balance search between global search and local search is realized, so that key features are finally identified.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a feature selection method, apparatus, device, and medium for predicting the risk of preeclampsia in pregnant women. Background Technology

[0002] Preeclampsia is a complex systemic pregnancy complication that causes more than 500,000 fetal and neonatal deaths and more than 70,000 maternal deaths worldwide each year. It also significantly increases the incidence of preterm birth and the risk of neonatal health problems. According to the International Society for the Study of Hypertension in Pregnancy (ISSHP) guidelines, a diagnosis of preeclampsia is made after 20 weeks of gestation with new-onset hypertension (defined as systolic blood pressure ≥140 mmHg or diastolic blood pressure ≥90 mmHg) accompanied by at least one preeclampsia-related symptom or sign. Relevant clinical manifestations include proteinuria, acute kidney injury, and elevated alanine aminotransferase or aspartate aminotransferase levels (>40 IU / L). Based on pathophysiological characteristics, preeclampsia can be divided into early-onset and late-onset subtypes. Although the exact pathogenesis is not fully understood, recent reviews and mechanistic studies emphasize the central role of placental dysfunction and maternal endothelial activation: early-onset disease is closely related to impaired spiral artery remodeling and placental developmental abnormalities, while late-onset disease often reflects the interaction between a (relatively) normal placenta and the mother's limited cardiovascular or metabolic reserves.

[0003] Therefore, developing machine learning-based prediction models for preeclampsia in pregnant women has significant clinical and social implications. Feature selection and dimensionality reduction are essential steps before establishing an efficient prediction model. Currently, feature selection methods are generally classified into three categories: filter-based methods, wrapper-based methods, and embedded methods. Filter-based methods operate independently of any learning algorithm, selecting features based on statistical criteria such as information gain or correlation coefficients. Although computationally efficient, these methods fail to consider potential interactions between features. Embedded methods automatically perform feature selection during model training, but the selected features often depend on a specific learning model, thus limiting generalization ability. Wrapper-based methods directly link the feature selection process to classification performance, but feature selection is an NP-hard combinatorial optimization problem. As the feature dimension increases, the corresponding search space expands exponentially, making it difficult for traditional exhaustive search methods to obtain the optimal solution within an acceptable timeframe.

[0004] Given the problems existing in current feature selection methods, metaheuristic algorithms are widely used in feature selection due to their powerful global search capabilities and flexible adaptability. Among them, the Sine-Cosine Algorithm (SCA) is a metaheuristic algorithm based on mathematical principles. It uses sine and cosine functions to achieve periodic oscillations between global exploration and local development in the search space. However, when dealing with high-dimensional and complex feature spaces, SCA may be misled, causing it to oscillate repeatedly among multiple features that provide similar information, making it difficult to identify the most core and discriminative key feature subset. It is prone to getting trapped in local optima and the imbalance between exploration and development makes it difficult to achieve a dynamic balance between global and local search, ultimately making it difficult to effectively identify key features. Summary of the Invention

[0005] This invention provides a feature selection method and apparatus for predicting the risk of preeclampsia in pregnant women, which can solve the problems existing in the prior art.

[0006] This invention provides a feature selection method for predicting the risk of preeclampsia in pregnant women, comprising the following steps: Obtain a feature dataset of pregnant patients associated with preeclampsia, the feature dataset including multi-dimensional blood routine index features; The population is initialized based on the feature dataset. An adaptive network topology strategy is used to set each individual in the initialized population as a node in the network, construct an adaptive network topology structure, and dynamically obtain the connection weights between individuals based on the individual fitness value and its similarity in the feature space. Sparse adjacency relationships between individuals are formed according to the connection weights between individuals, and a topological population representing the physiological interaction between multi-dimensional blood routine index features is generated. By using a sine and cosine search mechanism, the local search direction and the global optimal individual for multidimensional blood routine indicators associated with preeclampsia are determined. Based on the local search direction and the global optimal individual, individuals in the topological population are guided to perform periodic jump search between global search and local search in the feature space. At the same time, a multi-level information interaction strategy is used to guide individuals to search towards the global optimal individual in the feature space during global search, and to guide individuals to search towards individuals with sparse adjacency relationships in the feature space during local search, thereby obtaining multiple feature subsets. Obtain the fitness value of each feature subset, and repeat the search based on the fitness value until the maximum number of iterations is reached to obtain the optimal feature subset.

[0007] Preferably, the initialization of the population includes: Set up population Each individual Represents a candidate feature subset scheme; dimension For the number of blood routine indicators; each dimension Indicates the importance tendency of the corresponding features: Randomly initialize all dimensions The continuous values ​​are mapped to binary choice results using the sigmoid transformation function, as follows: ; in: ; The transformation function is the error function, expressed as: ; This binarization mechanism characterizes the probability of blood indicators being retained or excluded during prediction.

[0008] Preferably, obtaining the topological population includes: Each individual in the initial population is set as a node in the network to construct an adaptive network topology; individual In the The fitness of the next iteration is Then its weight is expressed as: ; Similarity is constructed based on the spatial differences in feature encodings among different individuals, and the similarity between individuals is defined as follows: and Defined as : ; in: This represents the Euclidean distance of an individual in the search space at the t-th iteration. and Location in the search space ; It is a scale parameter that controls the rate of similarity decay; Based on individual weights and spatial similarity, the edge weight update in the network is represented as: ; in: This represents the adaptive adjustment coefficient; Retain the highest weight ( The edges form a sparse adjacency structure between individuals, represented as: ; in: The control parameter representing the edge retention ratio is ultimately used to represent the neighborhood of each individual as follows: ; A topological population is generated based on the sparse adjacency structure between individuals.

[0009] Preferably, the search process of the sine and cosine search mechanism includes: Set the current global optimal solution as Then the individual position update is represented as: ; in: and They represent the first The individual in the first generation and first The position of the solution Affects the search step size; Indicates the random weight parameters; Represents a binary random number; Used to control the search scope, represented as: ; in: Indicates the maximum number of evaluations; Indicates the initial value; in, Used to determine whether the solution update direction approaches or moves away from the target solution. Used for the search process that affects individuals based on the offset of the current optimal solution. Used to switch between sine and cosine update rules. This is used to gradually converge the search from a large area to a local region.

[0010] Preferably, the search process of the multi-level information interaction strategy includes: At the global level, the following formula guides us toward the globally optimal feature subset: ; in: Let represent a dynamically decaying global guidance coefficient used to control the intensity of an individual's movement toward the global optimum. Its adaptive update is expressed as: ; At the local level, let Represents an individual In the next iteration, the individual The set of neighbors, whose adjacency relationships are determined by an adaptive network topology strategy, is represented by the local information fusion update as: ; in: This indicates the intensity used to quantify local information exchange; Represents an individual and The normalized edge weights between them are defined as: ; in: Represents an individual and ;and Represents an individual and The weights of the connections between them; The multi-level information interaction strategy integrates global and local guidance, and introduces a Gaussian perturbation term to achieve comprehensive updates, as follows: ; in: Indicates the disturbance coefficient; This indicates that the mean is zero and the variance is... Gaussian noise.

[0011] Preferably, obtaining the optimal feature subset includes: For each individual's feature subset, a support vector machine is used for classification, and the prediction accuracy is obtained. Introducing the feature ratio as a penalty term into the fitness function, it is expressed as: ; in: Indicates classification accuracy; Indicates the number of feature subsets; Indicates feature dimension; and Indicates the weighting parameter; The fitness value of each feature subset is obtained based on the fitness function. The search is repeated based on the fitness value until the maximum number of iterations is reached to obtain the optimal feature subset.

[0012] Preferably, the feature dataset includes hematocrit, hemoglobin, platelet-weighted hematocrit, white blood cell count, and red blood cell count. The hematocrit is used to reflect the blood concentration state, the hemoglobin index is used to reflect the oxygen transport capacity, the plateletcrit index is used to reflect platelet activity and coagulation state, and the red blood cell count index is used to reflect the blood's oxygen-carrying capacity.

[0013] This invention also provides a feature selection device for predicting the risk of preeclampsia in pregnant women, comprising: The data acquisition module is used to acquire a feature dataset of pregnant patients associated with preeclampsia, the feature dataset including multi-dimensional blood routine index features; The feature selection module initializes the population based on the feature dataset, uses an adaptive network topology strategy to set each individual in the initialized population as a node in the network, constructs an adaptive network topology structure, and dynamically obtains the connection weights between individuals based on the individual fitness value and their similarity in the feature space. Based on the connection weights between individuals, sparse adjacency relationships between individuals are formed, generating a topological population that represents the physiological interactions between multi-dimensional blood routine index features. By using a sine and cosine search mechanism, the local search direction and the global optimal individual for multidimensional blood routine indicators associated with preeclampsia are determined. Based on the local search direction and the global optimal individual, individuals in the topological population are guided to perform periodic jump search between global search and local search in the feature space. At the same time, a multi-level information interaction strategy is used to guide individuals to search towards the global optimal individual in the feature space during global search, and to guide individuals to search towards individuals with sparse adjacency relationships in the feature space during local search, thereby obtaining multiple feature subsets. Obtain the fitness value of each feature subset, and repeat the search based on the fitness value until the maximum number of iterations is reached to obtain the optimal feature subset.

[0014] This invention also provides an electronic device, including a memory and a processor; The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the steps of a feature selection method for predicting the risk of preeclampsia in pregnant women as described above.

[0015] This invention also provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the steps of a feature selection method for predicting the risk of preeclampsia in pregnant women as described above.

[0016] This invention provides a feature selection method, apparatus, and device for predicting the risk of preeclampsia in pregnant women. Compared with the prior art, its advantages are as follows: This invention first utilizes an adaptive network topology strategy to construct an adaptive network topology structure, and dynamically builds sparse adjacency relationships between individuals based on fitness and similarity to generate a topological population representing the interaction of multi-dimensional blood routine index features. This process maps the physiological correlations implicit in the high-dimensional feature space into a structured information network for algorithm search. The constructed topology structure reflects the inherent interactive correlations between features, guiding the algorithm to perform self-organized directional searches based on the search state, achieving a unified balance between exploration and development. During the specific search, a sine and cosine search mechanism is used to enable each individual to perform periodic jump searches between global and local searches under the guidance of the global optimal solution, achieving a dynamic balance between global exploration and local development. Simultaneously, a multi-level information interaction strategy is used, where the global term promotes convergence to the optimal solution, and the local term strengthens the mining of potential local regions, integrating global information propagation and local neighborhood learning into a unified update process. This achieves a dynamic balance between global and local searches based on the balance between exploration and development, ultimately identifying key features for subsequent use. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of a feature selection method for predicting the risk of preeclampsia in pregnant women, provided by an embodiment of the present invention. Figure 2 This diagram illustrates the selection frequency of various features in a ten-fold cross-validation experiment for a feature selection method for predicting the risk of preeclampsia in pregnant women, as provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0019] Preeclampsia is a complex systemic pregnancy complication that causes more than 500,000 fetal and neonatal deaths and more than 70,000 maternal deaths worldwide each year. It also significantly increases the incidence of preterm birth and the risk of neonatal health problems. According to the International Society for the Study of Hypertension in Pregnancy (ISSHP) guidelines, a diagnosis of preeclampsia is made after 20 weeks of gestation when new-onset hypertension (defined as systolic blood pressure ≥140 mmHg or diastolic blood pressure ≥90 mmHg) is present, accompanied by at least one preeclampsia-related symptom or sign. Relevant clinical manifestations include proteinuria, acute kidney injury, and elevated alanine aminotransferase or aspartate aminotransferase levels (>40 IU / L). Based on pathophysiological characteristics, preeclampsia can be divided into early-onset and late-onset subtypes. Although the exact pathogenesis is not fully understood, recent reviews and mechanistic studies emphasize the central role of placental dysfunction and maternal endothelial activation: early-onset disease is closely related to impaired spiral artery remodeling and placental developmental abnormalities, while late-onset disease often reflects the interaction between a (relatively) normal placenta and the mother's limited cardiovascular or metabolic reserves. Although an imbalance between normal maternal perfusion and placental-fetal metabolic demands may lead to late-onset preeclampsia, the release of pro-inflammatory cytokines, reactive oxygen species, extracellular vesicles, and anti-angiogenic factors (such as soluble FMS-like tyrosine kinase-1, sFLt1) from syncytiotrophoblasts under stress into the maternal circulation is also considered a potential trigger for preeclampsia.

[0020] The most conservative predictive approach currently is to assess maternal preeclampsia risk using high-risk factors. These factors are included in several guidelines, such as the UK National Institute for Health and Care Excellence (NICE) and the American College of Obstetricians and Gynecologists (ACOG). Although these guidelines apply to all pregnant women, their predictive efficacy is relatively poor. Soluble fms-like tyrosine kinase-1 (sFlt-1) and placental growth factor (PlGF) act as anti-angiogenic and pro-angiogenic factors, respectively. The phenomenon of normal decline in maternal PlGF levels at the end of pregnancy often occurs earlier in pregnant women who subsequently develop preeclampsia than in women with normal pregnancies. Therefore, regular monitoring of maternal circulating PlGF levels can be an effective means of predicting the occurrence of this disease. Previous studies have shown that an elevated sFlt-1 / PlGF ratio may indicate impaired angiogenesis and an exacerbation of the pathophysiological process of preeclampsia. In late pregnancy, the sFlt-1 / PlGF ratio demonstrates superior sensitivity and specificity compared to PlGF concentration in the diagnosis and monitoring of preeclampsia. However, the additional cost of PlGF and sFlt-1 blood tests makes these methods impractical for pregnant women in low- and middle-income countries. In high-income countries, Doppler ultrasound assessment of uteroplacental circulation is an effective method for predicting the occurrence of preeclampsia. Currently, the only treatment for preeclampsia is delivery, and early detection and enhanced management remain the main clinical strategies. Accurate assessment of the risk of preeclampsia in pregnant women is crucial to reducing its adverse effects. Due to insufficient research on etiology, high-risk factors, and potential multiple pathogenic phenotypes, early prediction of preeclampsia remains challenging.

[0021] However, current methods for predicting the second trimester have limited accuracy and narrow applicability, mainly because they often require additional expensive screening tests. In fact, women in low- and middle-income countries are not only more prone to preeclampsia, but their maternal mortality rate is also much higher than that in high-income countries. With the development of smart healthcare, machine learning technology has shown significant advantages in the field of disease prediction and diagnosis, and has gradually replaced traditional statistical methods.

[0022] Therefore, developing machine learning-based prediction models for preeclampsia in pregnant women has significant clinical and social implications. However, medical datasets are typically characterized by high dimensionality, strong feature correlations, and noise interference, posing significant challenges to model construction. If feature selection is not performed before model training, relying solely on prediction algorithms may lead to the curse of dimensionality, increased model complexity, and decreased prediction performance. Furthermore, the presence of redundant and irrelevant features not only reduces the generalization ability of the classifier but also increases computational costs, thus limiting the practical application value of the model. Therefore, feature selection and dimensionality reduction are essential steps before establishing an efficient prediction model. Feature selection methods are generally classified into three categories: filter-based methods, wrapper-based methods, and embedded methods. Filter-based methods operate independently of any learning algorithm, selecting features based on statistical criteria such as information gain or correlation coefficients. Although computationally efficient, these methods fail to consider potential interactions between features. Embedded methods automatically perform feature selection during model training, but the selected features often depend on a specific learning model, thus limiting generalization ability. In contrast, wrapper-based methods directly link the feature selection process to classification performance. These methods iteratively evaluate the prediction accuracy of a subset of features in a given classifier, thereby guiding the search process toward the optimal solution. Although computationally demanding, wrapper-based methods can usually identify the most discriminative and representative subset of features, achieving a better balance between model simplicity and prediction accuracy.

[0023] However, feature selection based on wrappers is an NP-hard combinatorial optimization problem. As the feature dimension increases, the corresponding search space expands exponentially, making it difficult for traditional exhaustive search methods to obtain the optimal solution within an acceptable timeframe. To address this challenge, metaheuristic algorithms, with their powerful global search capabilities and flexible adaptability, are widely used in feature selection. By simulating observed group behavior or evolutionary mechanisms in nature, these algorithms can efficiently approximate the optimal solution within a finite computation time, achieving a good balance between search efficiency and solution quality. Among them, the Sine-Cosine Algorithm (SCA) is a metaheuristic algorithm based on mathematical principles. It utilizes sine and cosine functions to achieve periodic oscillations between global exploration and local exploitation within the search space. This algorithm has a simple structure, is easy to implement, and has a limited number of parameters. However, the original SCA algorithm suffers from premature convergence, susceptibility to local optima, and an imbalance between exploration and exploitation, especially when dealing with high-dimensional complex feature spaces. These defects hinder the algorithm's ability to dynamically balance global and local searches, leading to weakened robustness and reduced generalization ability of the obtained feature subset.

[0024] To overcome the above limitations, this invention proposes an improved sine and cosine algorithm (MIISCA) based on multi-level information interaction, and combines it with support vector machine (SVM) to construct a bMIISCA-SVM method for predicting preeclampsia. Figure 1 As shown, MIISCA, by introducing a multi-level information interaction mechanism, effectively maintains population diversity while improving the algorithm's convergence accuracy, thereby achieving efficient selection of key feature subsets. This invention further employs SVM as a classifier, incorporating both classification error and feature quantity into the optimization objective, guiding the feature selection process towards an optimal and compact feature set, and improving the model's generalization ability and interpretability. Furthermore, the bMIISCA-SVM based on this invention demonstrates high accuracy and stability in the preeclampsia risk prediction task, verifying the application value and implementation potential of this method in real-world medical scenarios; specifically, it includes the following steps:

[0025] Step 1: Dataset Acquisition.

[0026] First, clinical data required for predicting preeclampsia in pregnant women were obtained. This invention collected data from 155 pregnant women diagnosed with preeclampsia in the obstetrics department of a hospital between January 1, 2019 and December 31, 2019, and collected 17 blood routine indicators. The subjects were divided into two groups: normal pregnancy group (n = 84) and mid-preeclampsia group (n = 71). The 17 hematological characteristics collected included hematocrit, hemoglobin, plateletcrit, white blood cell count, red blood cell count, etc. These indicators are not only highly available, but also reflect physiological processes related to the onset of preeclampsia, such as maternal infection, hemoconcentration, and vascular lesions.

[0027] Step 2: Dataset partitioning.

[0028] The dataset of pregnant women with preeclampsia was divided into a training set and a test set. The training set was used to build the bMIISCA-SVM model and perform feature selection; the test set was used to evaluate the generalization ability of the model.

[0029] Step 3: Initialization phase.

[0030] This invention models feature selection as a binary combinatorial optimization problem; assuming a population... Each individual Represents a candidate feature subset scheme; dimension For the number of blood routine indicators; each dimension The numerical value indicates the importance tendency of the corresponding feature: the larger the value, the more likely the feature is to be used in the preeclampsia classification model; to maintain the diversity of the search space, all dimensions are randomly initialized. Subsequently, a sigmoid transformation function is used to map continuous values ​​to binary selection results, thereby determining whether a certain feature is included. .

[0031] in: The transformation function takes the error function: .

[0032] This binarization mechanism reflects the probability characteristics of blood indicators being retained or excluded when predicting the risk of preeclampsia, which is beneficial for searching for feature subsets with clear structures.

[0033] Step 4: Adaptive network topology strategy.

[0034] After the initial population generation, the algorithm needs to conduct a more structured search in the blood routine index space where features are strongly correlated. Since the formation of preeclampsia involves multiple factors such as blood concentration and inflammatory response, there are inherent physiological correlations between different blood indicators. Therefore, in order to avoid the algorithm being pulled by a few high-fit individuals in the early stages of the search, which would lead to population homogenization, this invention introduces an adaptive network topology strategy, which enables the population structure to self-organize and evolve according to the search state.

[0035] In this strategy, each individual is considered a node in the network, and its influence in the network is established through its fitness value; if we denote an individual... In the The fitness of the next iteration is Then its weight is defined as: .

[0036] This weight reflects the predictive value of the feature subset. The higher the fitness, the greater the importance of the individual in the network, and the more likely it is to pass on its structural information to other individuals.

[0037] Meanwhile, this invention constructs similarity based on the spatial differences in feature encoding between different individuals, defining the similarity between individuals as... and Defined as : .

[0038] in: This represents the Euclidean distance of an individual in the search space at the t-th iteration. and Location in the search space ,and It is a scale parameter that controls the rate of similarity decay. The higher the similarity, the closer the spatial locations of the individuals, and the greater the potential value of information exchange.

[0039] Based on individual weights and spatial similarity, the edge weights in the network are updated using the following formula: .

[0040] in: It is an adaptive adjustment coefficient (set to 0.1 in this invention) used to balance the response speed and structural stability of the next-generation topology.

[0041] To avoid overly dense topologies that reduce population diversity, only the species with the highest weights are retained. ( Edges form a sparse adjacency structure: .

[0042] in: Ultimately, the neighborhood of each individual can be defined as the control parameter for the edge retention ratio: .

[0043] The adaptive topology constructed in the above manner enables the algorithm to possess both structural stability and information diversity throughout the search process, thereby improving the model's ability to identify feature subsets related to preeclampsia.

[0044] Step 5: Sine and Cosine Search Mechanism.

[0045] After the network structure is determined, the algorithm needs to further update the individual positions to explore better candidate subsets in the hematological feature space. This invention adopts the core idea of ​​SCA, enabling each individual to make periodic jumps under the guidance of the global optimal solution, achieving a dynamic balance between global exploration and local development. Let the current global optimal solution be... The individual position is then updated as follows: .

[0046] in: and They represent the first The individual in the first generation and first The position of the solution It affects the search step size and determines whether the solution update direction approaches or moves away from the target solution. As a random weight parameter, it influences the individual's search process based on the offset of the current optimal solution; This variable, as a binary random number (with a value range of [0,1]), is used to switch between sine and cosine update rules, thereby enhancing the diversity of the search; Used to control the search range, gradually converging the search from a large area to a local region, defined as: .

[0047] in: To the maximum number of evaluations, This is the initial value (set to 2).

[0048] Step Six: Multi-level Information Interaction Strategy.

[0049] Although the sine and cosine mechanisms can enhance an individual's exploratory ability, their search behavior is mainly driven by the global optimal solution, which can easily lead to the population concentrating on a single structure. The occurrence mechanism of preeclampsia involves multiple pathological pathways, so it is necessary to further enhance the learning ability of local structures so that the algorithm can mine complementary information from different feature combinations. To this end, this invention introduces a multi-level information interaction strategy, which integrates global information propagation and local neighborhood learning into a unified update process.

[0050] At the global level, the following formula guides us toward the globally optimal feature subset: .

[0051] in: It is a dynamically decaying global guiding coefficient used to control the intensity of an individual's movement towards the global optimum. Its adaptive update formula is as follows: .

[0052] This design allows the algorithm to have greater performance in the early stages. To enhance global exploration capabilities, and The value is gradually reduced in the later stages to improve the convergence and stability of the algorithm.

[0053] At the local level, each individual exchanges information only with its neighborhood to enhance its local search capabilities; let... Represents an individual In the next iteration, the individual The neighbor set, whose adjacency relationships are determined by the adaptive network topology strategy; the local information fusion update formula is as follows: .

[0054] in: Used to quantify the intensity of local information exchange; Represents an individual and The normalized edge weights between them are defined as follows: .

[0055] in: Represents an individual and ,and The weights of the connections between them; to avoid the denominator being zero, a minimal constant is introduced. To collaboratively utilize global and local information, a multi-level information interaction strategy integrates the above elements and introduces a Gaussian perturbation term to maintain population diversity; the comprehensive update formula is as follows: .

[0056] in: It is the disturbance coefficient. This indicates that the mean is zero and the variance is... Gaussian noise.

[0057] Step 7: Fitness assessment and update.

[0058] After the position update is completed, the algorithm needs to evaluate the feature subset corresponding to each individual to determine its retention priority in the next iteration. This invention regards feature selection as a dual-objective process that simultaneously optimizes classification performance and model simplicity. To this end, for the feature subset of each individual, the training set is classified using a Support Vector Machine (SVM), and the prediction accuracy is calculated. Meanwhile, a larger number of features often means decreased model interpretability and increased clinical testing costs. Therefore, the feature ratio is introduced as a penalty term into the fitness function. .

[0059] in: Indicates classification accuracy; The number of feature subsets; For feature dimensions; and The weighting parameters are set to 0.95 and 0.05 respectively.

[0060] Step 8: Output the optimal feature set and classification model.

[0061] When the number of iterations reaches the set maximum value, the algorithm ends and outputs the optimal feature subset and the final classification model. The optimal feature subset usually includes several hematological indicators that play an important role in the pathological process, such as hematocrit, which reflects the blood concentration state; hemoglobin, which reflects the oxygen transport capacity; plateletcrit, which reflects platelet activity and coagulation state; and red blood cell count, which has been repeatedly identified as a key indicator in this invention.

[0062] The SVM classifier trained on this feature subset will be used to predict the test set, and the model performance will be comprehensively evaluated by indicators such as accuracy, sensitivity, specificity, precision, Matthews correlation coefficient, and F-value. As shown in Table 1, the experimental results show that the bMIISCA-SVM constructed in this invention exhibits a high selection frequency for multiple hematological features. Among them, hemoglobin, red blood cell count, hematocrit, and plateletcrit are frequently selected during the feature selection process, indicating that the above indicators have high importance and stability in model construction. These features are all derived from routine blood tests and are closely related to pathophysiological processes such as hemoconcentration, hemolysis, and platelet dysfunction associated with preeclampsia. This further illustrates that while reducing the number of features, bMIISCA-SVM can effectively retain key information with clear clinical significance, thereby improving the predictive performance and interpretability of the model.

[0063] Table 1. Selection frequency of each feature in the bMIISCA-SVM during ten 10-fold cross-validation experiments. like Figure 2 As shown in the figure, the experimental results demonstrate that the bMIISCA-SVM constructed in this invention can maintain high predictive ability under conditions with fewer features, providing an efficient and interpretable tool for early risk identification of preeclampsia.

[0064] This invention dynamically constructs sparse connections based on fitness and similarity, effectively preventing premature convergence and population homogenization. Its topological structure reflects the intrinsic relationships between features, guiding the algorithm to perform directional searches in physiologically meaningful feature subspaces, thereby improving search efficiency and feature subset quality. This invention integrates sine and cosine mechanisms with multi-level information interaction, and achieves a dynamic balance between exploration and development through a triple-drive mechanism of global guidance, local learning, and random perturbation, avoiding getting trapped in local optima.

[0065] This invention dynamically constructs and updates the connection relationships between individuals in a population through an adaptive network topology strategy. The connection weights combine individual fitness and the similarity of feature encodings between individuals. This process cleverly maps the physiological connections implicit in the high-dimensional feature space into a structured information network for algorithm search. This guides the algorithm not to oscillate blindly across the entire space, but to perform targeted mining in a high-potential subspace composed of high-quality solutions and similar solutions.

[0066] This invention integrates three levels—global guidance, local neighborhood learning, and random perturbation—through a multi-level information interaction strategy. The global guidance decays exponentially, perfectly simulating the natural transition of the search process from bold exploration to refined development, thus resolving the contradiction of the fixed oscillation mode. In the early stage, strong global guidance ensures rapid location of potential areas. In the later stage, weakened global guidance combined with enhanced local neighborhood learning allows individuals to fine-tune under the influence of their high-quality neighbors, achieving rapid and stable convergence.

[0067] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A feature selection method for predicting the risk of preeclampsia in pregnant women, characterized in that, Includes the following steps: Obtain a feature dataset of pregnant patients associated with preeclampsia, the feature dataset including multi-dimensional blood routine index features; The population is initialized based on the feature dataset. An adaptive network topology strategy is used to set each individual in the initialized population as a node in the network, construct an adaptive network topology structure, and dynamically obtain the connection weights between individuals based on the individual fitness value and its similarity in the feature space. Sparse adjacency relationships between individuals are formed according to the connection weights between individuals, and a topological population representing the physiological interaction between multi-dimensional blood routine index features is generated. By using a sine and cosine search mechanism, the local search direction and the global optimal individual for multidimensional blood routine indicators associated with preeclampsia are determined. Based on the local search direction and the global optimal individual, individuals in the topological population are guided to perform periodic jump search between global search and local search in the feature space. At the same time, a multi-level information interaction strategy is used to guide individuals to search towards the global optimal individual in the feature space during global search, and to guide individuals to search towards individuals with sparse adjacency relationships in the feature space during local search, thereby obtaining multiple feature subsets. Obtain the fitness value of each feature subset, and repeat the search based on the fitness value until the maximum number of iterations is reached to obtain the optimal feature subset.

2. The feature selection method for predicting the risk of preeclampsia in pregnant women according to claim 1, characterized in that, The initialization of the population includes: Set up population Each individual Represents a candidate feature subset scheme; dimension For the number of blood routine indicators; each dimension Indicates the importance tendency of the corresponding features: Randomly initialize all dimensions The continuous values ​​are mapped to binary choice results using the sigmoid transformation function, as follows: ； in: ; The transformation function is the error function, expressed as: ； This binarization mechanism characterizes the probability of blood indicators being retained or excluded during prediction.

3. The feature selection method for predicting the risk of preeclampsia in pregnant women according to claim 2, characterized in that, The acquisition of the topological population includes: Each individual in the initial population is set as a node in the network to construct an adaptive network topology; individual In the The fitness of the next iteration is Then its weight is expressed as: ； Similarity is constructed based on the spatial differences in feature encodings among different individuals, and the similarity between individuals is defined as follows: and Defined as : ； in: This represents the Euclidean distance of an individual in the search space at the t-th iteration. and Location in the search space ; It is a scale parameter that controls the rate of similarity decay; Based on individual weights and spatial similarity, the edge weight update in the network is represented as: ； in: This represents the adaptive adjustment coefficient; Retain the highest weight ( The edges form a sparse adjacency structure between individuals, represented as: ； in: The control parameter representing the edge retention ratio is ultimately used to represent the neighborhood of each individual as follows: ； A topological population is generated based on the sparse adjacency structure between individuals.

4. The feature selection method for predicting the risk of preeclampsia in pregnant women according to claim 3, characterized in that, The search process of the sine and cosine search mechanism includes: Set the current global optimal solution as Then the individual position update is represented as: ； in: and They represent the first The individual in the first generation and first The position of the solution Affects the search step size; Indicates the random weight parameters; Represents a binary random number; Used to control the search scope, represented as: ； in: Indicates the maximum number of evaluations; Indicates the initial value; in, Used to determine whether the solution update direction approaches or moves away from the target solution. Used for the search process that affects individuals based on the offset of the current optimal solution. Used to switch between sine and cosine update rules. This is used to gradually converge the search from a large area to a local region.

5. The feature selection method for predicting the risk of preeclampsia in pregnant women according to claim 4, characterized in that, The search process of the multi-level information interaction strategy includes: At the global level, the following formula guides us toward the globally optimal feature subset: ； in: Let represent a dynamically decaying global guidance coefficient used to control the intensity of an individual's movement toward the global optimum. Its adaptive update is expressed as: ； At the local level, let Represents an individual In the next iteration, the individual The set of neighbors, whose adjacency relationships are determined by an adaptive network topology strategy, is represented by the local information fusion update as: ； in: This indicates the intensity used to quantify local information exchange; Represents an individual and The normalized edge weights between them are defined as: ； in: Represents an individual and ;and Represents an individual and The weights of the connections between them; The multi-level information interaction strategy integrates global and local guidance, and introduces a Gaussian perturbation term to achieve comprehensive updates, as follows: ； in: Indicates the disturbance coefficient; This indicates that the mean is zero and the variance is... Gaussian noise.

6. The feature selection method for predicting the risk of preeclampsia in pregnant women according to claim 5, characterized in that, The acquisition of the optimal feature subset includes: For each individual's feature subset, a support vector machine is used for classification, and the prediction accuracy is obtained. Introducing the feature ratio as a penalty term into the fitness function, it is expressed as: ； in: Indicates classification accuracy; Indicates the number of feature subsets; Indicates feature dimension; and Indicates the weighting parameter; The fitness value of each feature subset is obtained based on the fitness function. The search is repeated based on the fitness value until the maximum number of iterations is reached to obtain the optimal feature subset.

7. The feature selection method for predicting the risk of preeclampsia in pregnant women according to claim 1, characterized in that, The feature dataset includes hematocrit, hemoglobin, platelet-weighted hematocrit, white blood cell count, and red blood cell count. The hematocrit is used to reflect the blood concentration state, the hemoglobin index is used to reflect the oxygen transport capacity, the plateletcrit index is used to reflect platelet activity and coagulation state, and the red blood cell count index is used to reflect the blood's oxygen-carrying capacity.

8. A feature selection device for predicting the risk of preeclampsia in pregnant women, characterized in that, include: The data acquisition module is used to acquire a feature dataset of pregnant patients associated with preeclampsia, the feature dataset including multi-dimensional blood routine index features; The feature selection module initializes the population based on the feature dataset, uses an adaptive network topology strategy to set each individual in the initialized population as a node in the network, constructs an adaptive network topology structure, and dynamically obtains the connection weights between individuals based on the individual fitness value and their similarity in the feature space. Based on the connection weights between individuals, sparse adjacency relationships between individuals are formed, generating a topological population that represents the physiological interactions between multi-dimensional blood routine index features. By using a sine and cosine search mechanism, the local search direction and the global optimal individual for multidimensional blood routine indicators associated with preeclampsia are determined. Based on the local search direction and the global optimal individual, individuals in the topological population are guided to perform periodic jump search between global search and local search in the feature space. At the same time, a multi-level information interaction strategy is used to guide individuals to search towards the global optimal individual in the feature space during global search, and to guide individuals to search towards individuals with sparse adjacency relationships in the feature space during local search, thereby obtaining multiple feature subsets. Obtain the fitness value of each feature subset, and repeat the search based on the fitness value until the maximum number of iterations is reached to obtain the optimal feature subset.

9. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the steps of the feature selection method for predicting the risk of preeclampsia in pregnant women as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, which, when executed by a processor, implements the steps of a feature selection method for predicting the risk of preeclampsia in pregnant women as described in any one of claims 1 to 7.

Images