A method and system for reducing key features of cognitive risk in hyperuricemia based on self-information and rough set fusion
By fusing self-information and rough set methods, multi-kernel learning and adaptive optimization algorithms are used to process the multi-source heterogeneous features of cognitive risk of hyperuricemia, which solves the problems of feature redundancy and uncertainty, improves prediction accuracy and model interpretability, and generates interpretable risk assessment rules.
Patent Information
- Application Number
- CN202611213754.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-11-17
AI Technical Summary
Existing cognitive risk prediction technologies for hyperuricemia face challenges such as high dimensionality and redundancy of multi-source heterogeneous features, uncertain relationships between features, sample imbalance, and insufficient model interpretability, resulting in high computational complexity, decreased model generalization ability, and overfitting.
We employ a method that combines self-information and rough set theory. By using a fuzzy neighborhood rough set model with multi-kernel learning to handle nonlinear relationships, and combining self-information entropy and adaptive optimization algorithms, we can filter key features and generate interpretable fuzzy decision rules.
It significantly reduces feature redundancy, improves the accuracy and computational efficiency of cognitive risk prediction, enhances the sensitivity of identifying early mild cognitive impairment, and generates interpretable risk assessment rules for easy clinical application.
Smart Images

Figure CN123120291A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical information processing and intelligent medical decision-making technology, specifically involving a method and system for reducing key features of cognitive risk in hyperuricemia based on the fusion of self-information and rough set. Background Technology
[0002] Hyperuricemia (HUA) has become a serious public health problem. Clinical and animal studies suggest that HUA is an independent risk factor for cognitive impairment. Although there is increasing research on cognitive impairment caused by HUA, most studies are retrospective or animal experiments, lacking predictive models for HUA cognitive impairment risk. Current technologies for predicting cognitive risk in hyperuricemia face the following core challenges: 1. High dimensionality and redundancy of multi-source heterogeneous features: Potential influencing factors of cognitive risk associated with hyperuricemia encompass multi-dimensional data including serum uric acid levels, uric acid variability, uric acid-lowering drug use, renal function indicators, comorbid metabolic syndrome, inflammatory factors, cognitive assessment scales, and lifestyle. These features contain significant redundancy and noise. Directly using the original feature set for modeling leads to high computational complexity, decreased model generalization ability, and a high risk of overfitting.
[0003] 2. Uncertain relationships exist between features, and effective reduction methods are lacking: The association between uric acid levels and cognitive function exhibits nonlinearity, a threshold effect, and individual differences (e.g., different genders, ages, and disease stages show varying sensitivities to uric acid neurotoxicity). Studies have shown that serum uric acid has a dual role in cognitive function; although considered a neuroprotective agent, its neuroprotective function gradually weakens or even becomes harmful when it exceeds a certain threshold. Traditional feature selection methods (such as filtering or wrapping) struggle to handle this imprecise, incomplete information and nonlinear associations. While rough set theory can handle uncertainty, single rough set methods are sensitive to the discretization of continuous attributes and easily lose key information; combining them with self-information metrics is necessary to effectively uncover the intrinsic discriminative power of features.
[0004] 3. Lack of adaptive identification and dynamic reduction mechanisms for key features: Existing studies mostly rely on prior knowledge or univariate analysis to screen risk factors, failing to adaptively identify the feature subsets that contribute most to cognitive risk based on data distribution. In the hyperuricemia population, key risk features differ among different subgroups (e.g., asymptomatic hyperuricemia versus gout, and whether or not hypertension is present). Fixed feature sets are difficult to adapt to various clinical scenarios, necessitating a dynamic feature selection strategy that integrates self-information entropy and rough set reduction.
[0005] 4. Difficulty in obtaining samples and imbalance between positive and negative samples: There are few prospective cohort studies on cognitive impairment in hyperuricemia, and cases of early cognitive impairment are relatively scarce, leading to a severe imbalance in the proportion of normal and mild cognitive impairment samples in the training data. Traditional prediction models tend to favor the majority class (normal) and lack sufficient sensitivity in identifying early cognitive risk. Furthermore, the progression of cognitive decline associated with hyperuricemia is slow, and the feature boundaries are blurred, further increasing the difficulty of classification. In addition, the protective effect of uric acid only exists within a certain threshold range; levels below or above this range may increase cognitive risk. This U-shaped or J-shaped association further blurs the classification boundary between normal and abnormal.
[0006] 5. Insufficient Model Interpretability and Clinical Acceptability: While complex models such as deep learning can improve prediction accuracy, their "black box" nature makes them difficult for clinicians to understand. Assessing the cognitive risk of hyperuricemia requires clearly identifying which key features (such as long-term uric acid levels, variability, and inflammatory markers) are the basis for decision-making. Rough set methods naturally possess interpretable rule generation capabilities, but after incorporating self-information, the readability and clinical rationality of the reduction results must still be ensured to avoid losing important clinical information due to over-reduction.
[0007] In conclusion, it is essential to develop a key feature reduction method and system that can address issues such as high-dimensional heterogeneous feature processing, nonlinear uncertainty relationship modeling, sample imbalance, and insufficient interpretability in cognitive risk prediction of hyperuricemia. Summary of the Invention
[0008] Therefore, it is necessary to provide a method and system for reducing key features of hyperuricemia cognitive risk based on the fusion of self-information and rough set in existing hyperuricemia cognitive risk prediction, which addresses issues such as high-dimensional heterogeneous feature processing, nonlinear uncertainty relationship modeling, sample imbalance, and insufficient interpretability.
[0009] This invention proposes a method for reducing key features of cognitive risk in hyperuricemia based on the fusion of self-information and rough set theory, including: S1. Multidimensional clinical data collection and preprocessing: Collect multi-source heterogeneous data of hyperuricemia patients, including multiple medical indicators such as serum uric acid level and fluctuation, history of uric acid-lowering drug use, renal function indicators, metabolic syndrome-related indicators, inflammatory factors, cognitive assessment scale scores and lifestyle data, and perform missing value processing and outlier correction. S2. Apply nonlinear transformation and standardization to correct the skewed distribution of each feature, making the data closer to a normal distribution; S3. Construct a fuzzy neighborhood rough set model based on multi-kernel learning, and use the combination of multiple kernel functions to handle the nonlinear, multi-scale and threshold effect relationships between multi-source heterogeneous features; S4. Introduce self-information entropy to measure the ability of each feature to distinguish uncertainty in cognitive risk decision-making classes, and calculate the self-information gain of the features. S5. Based on the self-information entropy and the lower approximation quality of the fuzzy neighborhood rough set, a self-information rough set fusion importance index (SRFI) is constructed. This index is used to comprehensively evaluate the contribution of each feature to cognitive risk decision-making, and the initial key feature candidate set is selected based on the SRFI value. S6. An adaptive optimization algorithm combined with class-balanced cross-entropy loss is used to iteratively optimize the weight parameters in the feature reduction process to alleviate the problem of imbalance between positive and negative samples. S7. Based on the reduced subset of key features, construct an interpretable set of cognitive risk classification rules and output the cognitive risk level and explanation of key risk factors for patients with hyperuricemia.
[0010] In one embodiment, the multidimensional clinical data acquisition and preprocessing described in step S1; Multi-source heterogeneous raw data were collected from individuals with hyperuricemia, including multiple medical indicators such as serum uric acid levels and fluctuations, history of uric acid-lowering drug use, renal function indicators, metabolic syndrome-related indicators, inflammatory factors, cognitive assessment scale scores, and lifestyle data. Missing values were filled using multiple imputation, and outliers were identified and corrected using box-robin analysis. To eliminate the influence of different dimensions, each feature was Min-Max normalized and mapped to the [0,1] interval.
[0011] in These are the original eigenvalues. These are the normalized eigenvalues.
[0012] In one embodiment, the specific method for applying nonlinear transformation and normalization in step S2 is as follows: S21. For any feature, nonlinear transformation and standardization can be applied, and the Yeo-Johnson transformation can be used to make it approximate a normal distribution:
[0013] in The maximum likelihood estimation is used to determine how to make the transformed data follow a normal distribution to the greatest extent possible. S22. Based on the data obtained in step S21, perform Z-score standardization:
[0014] in and These are the mean and standard deviation of the transformed features, respectively, after standardization. The mean is 0 and the variance is 1.
[0015] In one embodiment, the specific method for constructing a fuzzy neighborhood rough set model based on multi-kernel learning in step S3 includes the following steps: S31. For any two samples, define a fuzzy similarity relationship based on multi-kernel learning:
[0016] in For the first The weights of the kernel functions satisfy... , The three kernel functions are: the rational quadratic kernel, used to capture multi-scale global nonlinear associations; the wavelet kernel, used to detect risk mutations near the uric acid threshold; and the chi-square kernel, used to process histogram distribution features (such as the distribution of inflammatory factor levels and the distribution of cognitive scale scores).
[0017] Rational quadratic kernel: , Among them, scale parameter Take the median of the Euclidean distance between samples, shape parameter (Default value), parameter Can be set to The value in Can be set to ( The values in the range (distance from standard deviation) are selected through cross-validation. Wavelet kernel: , in The feature dimension, For the first 3D eigenvalues For scale parameters; The standard deviation of the Euclidean distance between samples is taken as . If a significant threshold effect is detected (such as a U-shaped association between uric acid levels and cognitive risk), further five-fold cross-validation can be used. Intra-range optimization; Chi-square core: , in For the sample In the The frequency of each histogram interval (obtained by binning the eigenvalues). For interval numbers, To prevent division by zero for small constants.
[0018] S32. Based on the fuzzy similarity relationship obtained in step S31, calculate the fuzzy lower approximation and upper approximation of the sample belonging to the decision class; The following approximate definition is:
[0019] The upper approximation is defined as:
[0020] S33. Calculate the quality of the rough set approximation of the fuzzy neighborhood of the feature subset with respect to the decision:
[0021] in Based on feature subsets Calculation of fuzzy similarity relationships For decision-making Balance weights ( For the first (Number of samples per class) to eliminate the effects of class imbalance.
[0022] In one embodiment, the specific method for introducing self-information entropy to measure the ability of each feature to distinguish uncertainty in cognitive risk decision-making classes in step S4 is as follows: S41. Define the self-information entropy of a feature: To calculate the self-information entropy of continuous features, the feature values need to be discretized. This paper uses an equal-frequency discretization method to divide the value range of each feature into... Each interval contains approximately the same number of samples to preserve as much information as possible about the original distribution. Number of discretized intervals. Based on sample size Adaptive determination, for example .
[0023]
[0024] in Features The set of value intervals obtained after the above discretization For sample values to fall within the interval The probability of it; S42. Define the conditional self-information entropy of a feature relative to a decision:
[0025] S43. Calculate the self-information gain of the feature:
[0026] in For decision-making prior entropy. The larger the value, the more characteristic it indicates. The stronger the ability to distinguish between perceived risks.
[0027] In one embodiment, the specific method for constructing the self-information-rough set fusion importance index and screening the initial key feature candidate set in step S5 is as follows: S51. Define SRFI for each feature:
[0028] in For balancing parameters, For a single feature The approximate quality of the rough set in the fuzzy neighborhood; S52. Based on the SRFI values of each feature obtained in step S51, sort the features from largest to smallest, select the number of features whose cumulative SRFI is not lower than a preset threshold, and obtain the initial key feature candidate set.
[0029] In one embodiment, the specific method for optimizing feature weights in step S6 using an adaptive optimization algorithm combined with class-balanced cross-entropy loss is as follows: S61. Define weighted fuzzy similarity relations; Define feature weight vector The weighted fuzzy similarity relations are as follows:
[0030] in The multi-core combination in step S31 is adopted (each feature independently selects the optimal core or a fixed core). S62. Construct a class-balanced cross-entropy loss function:
[0031] in For class-balanced weights, For soft classification probabilities based on weighted fuzzy similarity relations:
[0032] S63. Construct the objective function:
[0033] in and These are hyperparameters that control the strength of L1 sparse regularization and L2 weight decay regularization, respectively. The L1 norm of the feature weights, The L2 norm squared. This objective function achieves sparsity and stability control of feature weights by simultaneously minimizing the classification loss and the regularization term, where... This causes the weights of unimportant features to approach zero. Prevent overfitting.
[0034] S64. Update feature weights using an adaptive optimization algorithm, such as the RAdam optimizer; The update rule for RAdam is defined as follows: First, calculate the gradient. First-order moment estimation Second-order moment estimation Deviation correction item , and calculate ,in .
[0035] when Adaptive learning rate update is used:
[0036] Otherwise, non-adaptive updates are used (degenerate to SGD):
[0037] Among them, learning rate , , , RAdam uses dynamic variance correction to automatically degenerate into SGD in the early stages of training, and then activates adaptive momentum after the variance stabilizes, effectively mitigating the risk of overfitting under small, imbalanced data. S65. Feature weight convergence determination and key feature subset selection.
[0038] When the overall objective function The optimal feature weights are obtained when convergence or when the maximum number of iterations is reached. Remove those with weights below the threshold. (like The features of ) are used to obtain the final key feature subset. .
[0039] In one embodiment, the specific method for constructing an interpretable cognitive risk classification rule set in step S7 is as follows: S71. Based on the final key feature subset and weighted fuzzy similarity relationship, extract the fuzzy decision rules for each decision class:
[0040] The interval Through the Class samples in features The value distribution on the above is obtained by interval estimation (e.g., mean ± standard deviation); S72, Rule Selection and Quality Assessment; Calculate the support for each rule ( ) and confidence level ( ), Filter support 0.05 and confidence level Rule 0.7 serves as the final interpretation rule; S73. For new samples, output the cognitive risk level according to the matching rules, and give the key risk factors (i.e. the features that appear in the rules and their value ranges).
[0041] A key feature reduction system for cognitive risk of hyperuricemia based on the fusion of self-information and rough set theory includes: The data acquisition and preprocessing module is used to collect multi-source heterogeneous data from people with hyperuricemia and perform missing value imputation, outlier correction, Min-Max normalization and nonlinear transformation. The multi-kernel fuzzy neighborhood rough set module, whose input is connected to the data acquisition and preprocessing module, is used to calculate the fuzzy similarity relationship between samples based on the combination of multiple kernel functions (such as rational quadratic kernel, wavelet kernel and chi-square kernel), and to calculate the approximate quality of each feature subset; The self-information entropy and feature importance fusion module has its input terminals connected to the multi-kernel fuzzy neighborhood rough set module, which is used to calculate the self-information gain of each feature and fuse the approximate quality to obtain the feature importance score, and to screen the initial key feature candidate set. An adaptive feature weight optimization module, whose input is connected to the self-information entropy and feature importance fusion module, is used to iteratively optimize the feature weights using an adaptive optimization algorithm (such as RAdam) combined with class-balanced cross-entropy loss to obtain the final key feature subset. The interpretable rule generation and risk prediction module has its input connected to the adaptive feature weight optimization module. It is used to extract fuzzy decision rules based on the final key feature subset and output the cognitive risk level and explanation of key risk factors. The visualization and clinical decision support module, whose input is connected to the interpretable rule generation and risk prediction module, is used to display feature reduction results, risk level distribution, importance ranking of key risk factors, and generate individualized risk assessment reports. The model update and incremental learning module, whose input is connected to the data acquisition and preprocessing module, is used to quickly and incrementally update the feature weights and rule sets based on an adaptive optimization algorithm when new data is added.
[0042] The beneficial effects of this invention are: (1) This invention addresses the specific needs of cognitive risk assessment for hyperuricemia by integrating self-information entropy with rough set theory to adaptively screen key risk features from high-dimensional heterogeneous clinical data, significantly reducing feature redundancy and improving the accuracy and computational efficiency of cognitive risk prediction.
[0043] (2) The Self-Information-Rough Set Fusion Importance Index (SRFI) was proposed for the first time. This index overcomes the limitation of the separation between independent and collaborative discriminant indicators in traditional feature selection methods by adaptively weighting the fusion of self-information gain and rough set approximation quality. It can simultaneously quantify the information content of features and their ability to determine decision boundaries, providing a unified mathematical metric for identifying key features of hyperuricemia cognitive risk, and significantly improving the scientific nature and clinical interpretability of feature reduction.
[0044] (3) By learning fuzzy neighborhood rough sets through multi-kernel (such as rational quadratic kernel, wavelet kernel and chi-square kernel), it can effectively capture the nonlinear, multi-scale and threshold effect correlation between uric acid level and cognitive function, especially the risk mutation pattern near the critical point of the transformation of uric acid protective effect to damage effect.
[0045] (4) Nonlinear transformation (such as Yeo-Johnson transformation) is used to correct the skewed distribution of medical indicators such as blood uric acid and inflammatory factors. Combined with Z-score standardization, the stability of rough set for continuous attribute discretization is improved, and the loss of key information caused by traditional discretization methods is avoided.
[0046] (5) The introduction of adaptive optimization algorithms (such as RAdam) and class-balanced cross-entropy loss effectively alleviated the imbalance of positive and negative samples in the prospective cohort of hyperuricemia cognitive impairment and improved the sensitivity of early mild cognitive impairment identification.
[0047] (6) Generate interpretable fuzzy decision rules, clearly list key risk factors and their value ranges (such as "blood uric acid level > 480 μmol / L and uric acid coefficient of variation > 0.25"), which is convenient for clinicians to understand and apply, and overcomes the "black box" defect of deep learning models.
[0048] (7) A complete system for reducing and predicting key features of cognitive risk in hyperuricemia was constructed, which realized full-process automation from data collection, feature reduction, risk prediction to clinical decision support, and provided intelligent tools for cognitive health management of hyperuricemia population. Attached Figure Description
[0049] Figure 1 This is a schematic diagram of the structure of the hyperuricemia cognitive risk key feature reduction system based on self-information and rough set fusion of the present invention; Figure 2This is a flowchart illustrating the method for reducing key features of cognitive risk in hyperuricemia based on the fusion of self-information and rough set in this invention. Figure 3 This is a schematic diagram of the method for performing nonlinear transformation using Yeo-Johnson transform and standardization according to the present invention; Figure 4 This is a schematic diagram illustrating the method for constructing a fuzzy neighborhood rough set model based on multi-kernel learning according to the present invention. Figure 5 This invention introduces a self-information entropy metric to measure the ability of each feature to distinguish uncertainty in cognitive risk decision-making, and calculates the self-information gain of the features. Figure 6 This is a schematic diagram of the method for screening an initial key feature candidate set based on the lower approximation quality of self-information entropy and fuzzy neighborhood rough set of the present invention, which constructs a fusion feature importance evaluation function. Figure 7 This is a schematic diagram of the method for iteratively optimizing the weight parameters in the feature reduction process using the RAdam optimizer combined with the class-balanced cross-entropy loss. Figure 8 This is a schematic diagram of the method for constructing an interpretable set of cognitive risk classification rules based on a reduced subset of key features according to the present invention. Figure 9 This invention provides convergence curves showing the objective function value as a function of iteration number under different datasets in embodiments of the invention. Figure 10 The parameters of the SVM classifier provided in the embodiments of the present invention and Sensitivity analysis; Figure 11 Parameters of the DA classifier provided in the embodiments of the present invention and Sensitivity analysis; The meanings of the numbers in the attached diagram are as follows: 100-A key feature reduction system for cognitive risk of hyperuricemia based on self-information and rough set fusion; 10-Data Acquisition and Preprocessing Module; 20-Multi-kernel fuzzy neighborhood rough set module; 30 - Module for fusing self-information entropy and feature importance; 40 - Adaptive Feature Weight Optimization Module; 50 - Explainable rule generation and risk prediction module; 60 - Visualization and Clinical Decision Support Module; 70 - Model Update and Incremental Learning Module; Detailed Implementation
[0050] 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.
[0051] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0052] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0053] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0054] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.
[0055] It should be noted that when an element is referred to as being "fixed to" or "set on" another element, it can be directly on the other element or there may be an intervening element. When an element is considered to be "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "upper," "lower," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.
[0056] Please see Figure 1 A key feature reduction system for cognitive risk of hyperuricemia based on the fusion of self-information and rough set theory 100, comprising: The data acquisition and preprocessing module 10 is used to collect multi-source heterogeneous data of people with hyperuricemia and perform missing value imputation, outlier correction, Min-Max normalization and nonlinear transformation. The multi-kernel fuzzy neighborhood rough set module 20, whose input end is connected to the data acquisition and preprocessing module 10, is used to calculate the fuzzy similarity relationship between samples based on the combination of rational quadratic kernel, wavelet kernel and chi-square kernel, and to calculate the approximate quality of each feature subset; The self-information entropy and feature importance fusion module 30 has its input end connected to the multi-kernel fuzzy neighborhood rough set module 20. It is used to calculate the self-information gain of each feature and fuse the approximate quality to obtain the feature importance score, and to screen the initial key feature candidate set. The adaptive feature weight optimization module 40, whose input is connected to the self-information entropy and feature importance fusion module 30, is used to iteratively optimize the feature weights using an adaptive optimization algorithm (such as RAdam) combined with class-balanced cross-entropy loss to obtain the final key feature subset. The interpretable rule generation and risk prediction module 50, whose input is connected to the adaptive feature weight optimization module 40, is used to extract fuzzy decision rules based on the final key feature subset and output cognitive risk level and explanation of key risk factors. The visualization and clinical decision support module 60, whose input is connected to the interpretable rule generation and risk prediction module 50, is used to display feature reduction results, risk level distribution, importance ranking of key risk factors, and generate a personalized risk assessment report; and The model update and incremental learning module 70, whose input is connected to the visualization and clinical decision support module 60, is used to quickly and incrementally update the feature weights and rule sets based on an adaptive optimization algorithm (RAdam may be used) when new data is added.
[0057] The following is combined Figures 2 to 8 The specific implementation steps of the method of the present invention will be described in detail.
[0058] Please see Figure 2 A method for reducing key features of cognitive risk in hyperuricemia based on the fusion of self-information and rough set theory, implemented based on the aforementioned system 100 for reducing key features of cognitive risk in hyperuricemia based on the fusion of self-information and rough set theory, includes the following steps: S1. Multidimensional clinical data collection and preprocessing: This includes multiple medical indicators, such as collecting multi-source heterogeneous data from people with hyperuricemia, including serum uric acid levels and fluctuations, history of uric acid-lowering drug use, renal function indicators, metabolic syndrome-related indicators, inflammatory factors, cognitive assessment scale scores, and lifestyle data, and performing missing value processing and outlier correction. Missing values were filled using multiple interpolation, and outliers were identified and corrected using box plotting. To eliminate the impact of different dimensions... To mitigate the impact of these effects, Min-Max normalization is applied to each feature, mapping it to the [0,1] interval:
[0059] in These are the original eigenvalues. These are the normalized eigenvalues.
[0060] Specifically, the sample of the population with hyperuricemia These can be considered as observation records of patients from different outpatient clinics or physical examinations, or as observation records of the same patient at different follow-up time points. Characteristics This corresponds to clinically measurable physical quantities, such as serum uric acid concentration (μmol / L), uric acid coefficient of variation (standard deviation / mean), estimated glomerular filtration rate (eGFR, mL / min / 1.73m²), C-reactive protein (CRP, mg / L), and Montreal Cognitive Assessment (MoCA) score. Multiple imputation of missing values utilizes complete records from other patients with similar characteristics, avoiding sample waste caused by simple deletion. Box-line analysis effectively eliminates extreme values (e.g., serum uric acid >1200 μmol / L) caused by detection or data entry errors. Min-Max normalization unifies features with different dimensions to the same numerical range, ensuring that subsequent distance-based fuzzy similarity calculations are not biased towards features with large absolute values due to dimensional differences (e.g., uric acid variability and cognitive scores).
[0061] S2. Apply nonlinear transformation and standardization to correct the skewed distribution of each feature, making the data closer to a normal distribution; Please see Figure 3 Specifically, it includes the following steps: S21. For any feature, nonlinear transformation and standardization can be applied, and the Yeo-Johnson transformation can be used to make it approximate a normal distribution:
[0062] in The maximum likelihood estimation is used to determine how to make the transformed data follow a normal distribution to the greatest extent possible. S22. Based on the data obtained in step S21, perform Z-score standardization:
[0063] in and These are the mean and standard deviation of the transformed features, respectively, after standardization. The mean is 0 and the variance is 1.
[0064] Specifically, clinical laboratory indicators such as serum uric acid, CRP, and triglycerides typically exhibit a right-skewed distribution (i.e., most people have low levels, while a minority have abnormally high levels). Directly using raw values can lead to sparse samples in high-value areas and oversensitivity of the model to outliers. The Yeo-Johnson transform adaptively selects values through maximum likelihood estimation. It can correct skewed distributions to an approximate normal distribution, for example, blood uric acid... After transformation, the long tail is compressed, making the difference between the normal range and the high uric acid range more uniform. Subsequently, Z-score standardization places different features on the same scale, ensuring that when the kernel function calculates distance, the contribution of each feature is inversely proportional to its standard deviation—features with small fluctuations (such as age) are not amplified by their small dimensions, while features with large fluctuations (such as diurnal uric acid fluctuations) are not suppressed. This combined processing is a prerequisite for the effective discretization of continuous attributes using fuzzy rough sets, avoiding unreasonable discretization intervals caused by distribution skew (such as losing information because most samples are crowded into the same interval).
[0065] S3. Construct a fuzzy neighborhood rough set model based on multi-kernel learning, and use a combination of multiple kernel functions (such as rational quadratic kernel, wavelet kernel and chi-square kernel) to handle the nonlinear, multi-scale and threshold effect relationships between multi-source heterogeneous features. Please see Figure 4 Specifically, it includes the following steps: S31. For any two samples, define a fuzzy similarity relationship based on multi-kernel learning:
[0066] in For the first The weights of the kernel functions satisfy... , The three kernel functions are: the rational quadratic kernel, used to capture multi-scale global nonlinear associations; the wavelet kernel, used to detect risk mutations near the uric acid threshold; and the chi-square kernel, used to process histogram distribution features (such as the distribution of inflammatory factor levels and the distribution of cognitive scale scores).
[0067] Rational quadratic kernel: , Among them, scale parameter Take the median of the Euclidean distance between samples, shape parameter (Default value), parameter Can be set to The value in Can be set to ( The values in the range (distance from standard deviation) are selected through cross-validation. Wavelet kernel: , in The feature dimension, For the first 3D eigenvalues For scale parameters; The standard deviation of the Euclidean distance between samples is taken as . If a significant threshold effect is detected (such as a U-shaped association between uric acid levels and cognitive risk), further five-fold cross-validation can be used. Intra-range optimization; Chi-square core: , in For the sample In the The frequency of each histogram interval (obtained by binning the eigenvalues). For interval numbers, To prevent division by zero for small constants.
[0068] S32. Based on the fuzzy similarity relationship obtained in step S31, calculate the fuzzy lower approximation and upper approximation of the sample belonging to the decision class; The following approximate definition is:
[0069] The upper approximation is defined as:
[0070] S33. Calculate the quality of the rough set approximation of the fuzzy neighborhood of the feature subset with respect to the decision:
[0071] in Based on feature subsets Calculation of fuzzy similarity relationships For decision-making Balance weights ( For the first (Number of samples per class) to eliminate the effects of class imbalance.
[0072] Specifically, in the cognitive risk assessment of hyperuricemia, both patients... and Fuzzy similarity This represents the degree of similarity in their overall risk patterns. The rational quadratic kernel passes through parameters... and It can simultaneously capture both large-scale trends in the relationship between uric acid and cognition (e.g., the overall trend of declining cognitive scores for every 100 μmol / L increase in long-term uric acid levels) and small-scale fluctuations (e.g., the transient impact of short-term uric acid fluctuations on attention). The wavelet kernel possesses local support properties, making it particularly suitable for detecting abrupt patterns near the protective-damaging transition threshold of uric acid—for example, when serum uric acid rises from 480 μmol / L to 540 μmol / L, cognitive risk may change from stable to a sharp increase. The wavelet kernel can assign higher similarity to samples within this neighborhood, thereby enhancing the model's sensitivity to threshold effects. The chi-square kernel uses the distribution histogram of inflammatory factors (such as IL-6 and TNF-α) as input to compare the similarity of the inflammatory profiles of two patients—even if both have the same mean IL-6 level, one may show a sustained, mild increase while the other shows intermittent bursts; the chi-square kernel can distinguish them through differences in distribution shape. The weights of the three kernels are discussed. The most suitable similarity metric is automatically determined through subsequent optimization, enabling the model to adaptively select the most appropriate similarity measure based on the data. (Approximation under fuzzy conditions) Indicates sample Determined to belong to the category of cognitive impairment The higher the value, the greater the degree. With all non The lower the similarity between the sample groups (i.e., significantly different from non-cognitive impairment patients), the more clinically relevant the patients with typical risk characteristics and clear diagnoses; the higher the similarity. Then it means It may belong to The degree of similarity corresponds to borderline cases that share some characteristics with patients with known cognitive impairment. Approximate quality Then evaluate the feature subset The extent to which the category of all samples can be determined is the core basis for feature selection.
[0073] S4. Introduce self-information entropy to measure the ability of each feature to distinguish uncertainty in cognitive risk decision-making classes, and calculate the self-information gain of the features. Please see Figure 5 Specifically, it includes the following steps: S41. Define the self-information entropy of a feature: To calculate the self-information entropy of continuous features, the feature values need to be discretized. This paper uses an equal-frequency discretization method to divide the value range of each feature into... Each interval contains approximately the same number of samples to preserve as much information as possible about the original distribution. Number of discretized intervals. Based on sample size Adaptive determination, for example .
[0074]
[0075] in Features The set of value intervals obtained after the above discretization For sample values to fall within the interval The probability of it; S42. Define the conditional self-information entropy of a feature relative to a decision:
[0076] S43. Calculate the self-information gain of the feature:
[0077] in For decision-making prior entropy. The larger the value, the more characteristic it indicates. The stronger the ability to distinguish between perceived risks.
[0078] Specifically, self-information entropy quantifies the uncertainty of features from an information theory perspective. Taking blood uric acid levels as an example: if the uric acid levels of all patients with hyperuricemia are concentrated in a narrow range (e.g., 450-480 μmol / L), then... A lower value indicates a lack of diversity in this characteristic, limiting its contribution to distinguishing different cognitive risk states; conversely, if uric acid levels are widely distributed between 300-800 μmol / L, then... Larger values imply more information. Conditional entropy. This indicates how much uncertainty remains regarding the cognitive risk category after knowing the uric acid level. If the uric acid level completely determines cognitive risk (e.g., everyone experiences cognitive impairment when uric acid >540 μmol / L, and is normal when uric acid <360 μmol / L), then... Information gain To reach its maximum. In actual clinical practice, the relationship between uric acid and cognition follows a U-shaped or J-shaped pattern: both excessively low and high levels increase risk, while moderate levels offer the strongest protective effect. Still greater than 0 but significantly less than the prior entropy , Medium. By calculating the sum of all candidate features This can initially screen out those clinical indicators that are most strongly associated with cognitive risk, such as serum uric acid levels, uric acid variability, and inflammatory factors, which often have high information gain.
[0079] S5. Based on the lower approximation quality of self-information entropy and fuzzy neighborhood rough set, construct the self-information rough set fusion importance index (SRFI) to screen the initial key feature candidate set; Please see Figure 6 Specifically, it includes the following steps: S51. Define SRFI for each feature:
[0080] in For balancing parameters, For a single feature The approximate quality of the rough set in the fuzzy neighborhood; S52. Based on the SRFI values of each feature obtained in step S51, sort the features from largest to smallest, select the number of features whose cumulative SRFI is not lower than a preset threshold, and obtain the initial key feature candidate set.
[0081] Among them, the preset threshold It can be set to 0.9, which is the smallest subset of features that retain the top 90% of the most important features, to obtain the initial candidate set of key features. If the total number of features is small, all features can be retained directly. Its characteristics.
[0082] Specifically, self-information gain The ability of features to independently distinguish risks is measured by the approximate quality of rough sets. This measures the ability of a feature to determine the decision boundary after interacting with other features. Both have their own emphasis: High values for a particular trait may contribute little due to high redundancy with other traits (e.g., daily and weekly mean values of uric acid are almost identical). High-scoring features may be crucial because they complement other features. (Fusing score) Through parameters To balance the two, for example, take Simultaneously considering both the independent and cooperative discriminant power of features. In clinical applications, serum uric acid levels typically have a high [discriminatory power / intelligence]. Uric acid variability may synergistically influence cognitive risk in conjunction with renal function indicators (i.e., patients with renal insufficiency are more sensitive to uric acid fluctuations), in which case variability... It may be higher than its By fusing and ranking features and selecting a subset with a cumulative importance of 90%, a large number of irrelevant or redundant features can be eliminated (such as sleep duration in lifestyle factors, which has extremely low importance scores if it is not significantly associated with cognitive impairment), resulting in a reduced candidate set. This lays the foundation for subsequent fine-tuning.
[0083] S6. An adaptive optimization algorithm combined with class-balanced cross-entropy loss is used to iteratively optimize the weight parameters in the feature reduction process to alleviate the problem of imbalance between positive and negative samples. Please see Figure 7 Specifically, it includes the following steps: S61. Define weighted fuzzy similarity relations; Define feature weight vector The weighted fuzzy similarity relations are as follows:
[0084] in The multi-core combination in step S31 is adopted (each feature independently selects the optimal core or a fixed core). S62. Construct a class-balanced cross-entropy loss function:
[0085] in For class-balanced weights, For soft classification probabilities based on weighted fuzzy similarity relations:
[0086] S63. Construct the objective function:
[0087] in and These are hyperparameters that control the strength of L1 sparse regularization and L2 weight decay regularization, respectively. The L1 norm of the feature weights, The L2 norm squared. This objective function achieves sparsity and stability control of feature weights by simultaneously minimizing the classification loss and the regularization term, where... This causes the weights of unimportant features to approach zero. Prevent overfitting.
[0088] S64. Update feature weights using an adaptive optimization algorithm, such as the RAdam optimizer; The update rule for RAdam is defined as follows: First, calculate the gradient. First-order moment estimation Second-order moment estimation Deviation correction item , and calculate ,in .
[0089] when Adaptive learning rate update is used:
[0090] Otherwise, non-adaptive updates are used (degenerate to SGD):
[0091] Among them, learning rate , , , RAdam uses dynamic variance correction to automatically degenerate into SGD in the early stages of training, and then activates adaptive momentum after the variance stabilizes, effectively mitigating the risk of overfitting under small, imbalanced data. S65. Feature weight convergence determination and key feature subset selection; When the overall objective function The optimal feature weights are obtained when convergence or when the maximum number of iterations is reached. Remove those with weights below the threshold. (like The features of ) are used to obtain the final key feature subset. .
[0092] S7. Based on the reduced subset of key features, construct an interpretable set of cognitive risk classification rules and output the cognitive risk level and explanation of key risk factors for patients with hyperuricemia.
[0093] Please see Figure 8 Specifically, it includes the following steps: S71. Based on the final key feature subset and weighted fuzzy similarity relationship, extract the fuzzy decision rules for each decision class:
[0094] The interval Through the Class samples in features The value distribution on the above is obtained by interval estimation (e.g., mean ± standard deviation); S72, Rule Selection and Quality Assessment; Calculate the support for each rule ( ) and confidence level ( ), Filter support 0.05 and confidence level Rule 0.7 serves as the final interpretation rule; S73. For new samples, output the cognitive risk level according to the matching rules, and give the key risk factors (i.e. the features that appear in the rules and their value ranges).
[0095] verify: In clinical studies of cognitive risk in hyperuricemia, the acquisition of cognitive impairment samples is affected by factors such as the long duration of prospective cohort studies and the scarcity of early cognitive impairment cases. Normal cognitive samples far outnumber cognitive impairment samples, and the types of cognitive impairment are diverse and labeling costs are high, making it difficult to construct a large-scale labeled dataset covering all scenarios in a short period. To verify the general effectiveness of the feature reduction and risk classification framework proposed in this invention, multiple publicly available real-world medical datasets were used for algorithm performance evaluation. These datasets are diverse in sample size, feature dimensions, category distribution, and data type (as shown in Table 1), enabling a comprehensive test of the method's feature selection ability and classification accuracy under imbalanced data, high-dimensional features, and nonlinear relationships. This invention uses classification accuracy as the evaluation metric, employing both SVM and DA classifiers, and evaluates them through cross-validation. The experimental results are shown in Tables 2 and 3.
[0096] Table 1. Detailed information about the experimental dataset.
[0097] Table 2. Classification accuracy of the algorithm of this invention on the SVM classifier.
[0098] Table 3. Classification accuracy of the algorithm of the present invention on the DA classifier.
[0099] To verify the convergence performance of the proposed method, the curve of the objective function value changing with the number of training iterations was plotted, and the results are as follows: Figure 9 As shown. By Figure 9 It can be seen that the objective function values of the four datasets DNA, GS, SM, and SQ all show a smooth exponential downward trend, without oscillation or rebound, and finally converge to near 0 within 200 epochs, verifying the effectiveness and stability of the optimization algorithm used.
[0100] To verify the parameter robustness of the proposed method, the hyperparameters of the SVM classifier were analyzed. , Sensitivity analysis was performed on four datasets: DNA, GS, SM, and SQ. and Calculate the classification accuracy, and the results are as follows: Figure 10 As shown. Figure 10 (a) shows that, on the DNA dataset, the curves initially decline rapidly before rebounding. After reaching its peak, it declines significantly. and It tends to stabilize and maintains superior performance in the later stages, while The worst overall performance. Figure 10 (b) shows that on the GS dataset, the curves remain stable in the early stages. , and They performed relatively well initially, but then all dropped sharply to their lowest levels, only... and After a slight decrease, it tended to stabilize, indicating a smaller [indicator / condition]. The dataset exhibits extremely high sensitivity in its later stages. Figure 10 (c) shows that the curves on the SM dataset exhibit minimal overall fluctuations and remain essentially stable. The small differences between the values indicate that the dataset is not sensitive to parameter changes and has good robustness. Figure 10 (d) shows that the curves on the SQ dataset rise rapidly in the early stages. After reaching its peak, it gradually declines. It performed the worst overall and then dropped sharply in the later stages, while In the later stages, the SVM classifier tends to stabilize and becomes a better choice. In summary, the SVM classifier's parameters... and It exhibits a certain sensitivity, with significant differences between different datasets: the DNA dataset is suitable for larger datasets. The GS dataset should avoid being too small. In later parameter selection, the SM dataset exhibits excellent robustness, while the SQ dataset is suitable for moderate-to-moderate use. This verifies that the method requires parameter selection based on the specific characteristics of the data.
[0101] To verify the parameter robustness of the proposed method, the hyperparameters of the DA classifier were investigated. , Sensitivity analysis was performed on four datasets: DNA, GS, SM, and SQ. and Calculate the classification accuracy, and the results are as follows: Figure 11 As shown. Figure 11 (a) shows that on the DNA dataset, the curves rise rapidly in the early stages. After reaching its peak, it declines significantly. and It tends to stabilize and maintains superior performance in the later stages, while The worst overall performance. Figure 11 (b) shows that on the GS dataset, the curves remain stable in the early stages, with relatively small... They performed relatively well initially, but then all dropped sharply to their lowest levels, only... After a slight decrease, it tended to stabilize, indicating a smaller [indication / condition]. The dataset exhibits extremely high sensitivity in its later stages. Figure 11 (c) shows that the curves on the SM dataset exhibit minimal overall fluctuations and remain essentially stable. The small differences between the values indicate that the dataset is not sensitive to parameter changes and has good robustness. Figure 11 (d) shows that the curves fluctuate significantly on the SQ dataset. It reaches its peak early and then gradually declines. Overall, it performed the worst, with the curves converging in the later stages. In summary, the DA classifier has poor parameter performance. and It exhibits a certain sensitivity, with significant differences between different datasets: the DNA dataset is suitable for larger datasets. The GS dataset should avoid being too small. In later parameter selection, the SM dataset exhibits excellent robustness, while the SQ dataset is suitable for medium to small datasets. This verifies that the method requires parameter selection based on the specific characteristics of the data.
[0102] In summary, the key feature reduction method for cognitive risk of hyperuricemia based on the fusion of self-information and rough set proposed in this invention shows good performance in terms of risk prediction accuracy, feature reduction ability, convergence speed and robustness. Experiments have verified its effectiveness and stability as a solution for cognitive risk assessment of hyperuricemia.
[0103] It should be noted that the core innovation of this invention lies in combining self-information entropy measurement with fuzzy neighborhood rough set reduction to construct an adaptive feature reduction and classification framework suitable for medical risk prediction scenarios. Specifically, self-information entropy quantifies the ability of features to distinguish uncertainties in cognitive risk decision classes, enabling preliminary screening of key risk factors; fuzzy neighborhood rough set combined with multi-kernel learning (rational quadratic kernel, wavelet kernel, chi-square kernel) can capture the nonlinearity, threshold effect, and individual variability correlation between serum uric acid levels and cognitive function, enhancing the ability to identify the critical point of uric acid protection-damage transition; class-balanced cross-entropy loss can effectively handle the class imbalance problem between normal cognitive and cognitively impaired samples. Experiments on publicly available datasets aim to verify the effectiveness, stability, and transferability of the above mechanisms in a general data environment, rather than directly replacing field validation in actual clinical cohorts.
[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0105] 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 method and system for reducing key features of cognitive risk in hyperuricemia based on the fusion of self-information and rough set theory, characterized in that, Includes the following steps: S1. Multidimensional clinical data collection and preprocessing: Collect multi-source heterogeneous data of hyperuricemia patients, including multiple medical indicators such as serum uric acid level and fluctuation, history of uric acid-lowering drug use, renal function indicators, metabolic syndrome-related indicators, inflammatory factors, cognitive assessment scale scores and lifestyle data, and perform missing value processing and outlier correction. S2. Apply nonlinear transformation and standardization to correct the skewed distribution of each feature, making the data closer to a normal distribution; S3. Construct a fuzzy neighborhood rough set model based on multi-kernel learning, and use the combination of multiple kernel functions to handle the nonlinear, multi-scale and threshold effect relationships between multi-source heterogeneous features; S4. Introduce self-information entropy to measure the ability of each feature to distinguish uncertainty in cognitive risk decision-making classes, and calculate the self-information gain of the features. S5. Based on the self-information entropy and the lower approximation quality of the fuzzy neighborhood rough set, a self-information rough set fusion importance index (SRFI) is constructed. This index is used to comprehensively evaluate the contribution of each feature to cognitive risk decision-making, and the initial key feature candidate set is selected based on the SRFI value. S6. An adaptive optimization algorithm combined with class-balanced cross-entropy loss is used to iteratively optimize the weight parameters in the feature reduction process to alleviate the problem of imbalance between positive and negative samples. S7. Based on the reduced subset of key features, construct an interpretable set of cognitive risk classification rules and output the cognitive risk level and explanation of key risk factors for patients with hyperuricemia.
2. The method for reducing key features of cognitive risk in hyperuricemia based on the fusion of self-information and rough set theory according to claim 1, characterized in that, The multidimensional clinical data acquisition and preprocessing described in step S1; Multi-source heterogeneous raw data were collected from individuals with hyperuricemia, including multiple medical indicators such as serum uric acid levels and fluctuations, history of uric acid-lowering drug use, renal function indicators, metabolic syndrome-related indicators, inflammatory factors, cognitive assessment scale scores, and lifestyle data. Missing values were filled using multiple imputation, and outliers were identified and corrected using box-robin analysis. To eliminate the influence of different dimensions, each feature was Min-Max normalized and mapped to the [0,1] interval. in These are the original eigenvalues. These are the normalized eigenvalues.
3. The method for reducing key features of hyperuricemia cognitive risk based on the fusion of self-information and rough set theory according to claim 1, characterized in that, Step S2 applies nonlinear transformation and normalization, including the following steps: S21. For any feature, nonlinear transformation and standardization can be applied, and the Yeo-Johnson transform can be used to make it approximate a normal distribution: in The maximum likelihood estimation is used to determine how to make the transformed data follow a normal distribution to the greatest extent possible. S22. Based on the data obtained in step S21, perform Z-score standardization: in and These are the mean and standard deviation of the transformed features, respectively, after standardization. The mean is 0 and the variance is 1.
4. The method for reducing key features of hyperuricemia cognitive risk based on the fusion of self-information and rough set theory according to claim 1, characterized in that, The fuzzy neighborhood rough set model based on multi-kernel learning described in step S3 includes the following steps: S31. For any two samples, define a fuzzy similarity relationship based on multi-kernel learning: in For the first The weights of the kernel functions satisfy... , The three kernel functions are: the rational quadratic kernel, used to capture multi-scale global nonlinear correlations; the wavelet kernel, used to detect risk mutations near the uric acid threshold; and the chi-square kernel, used to process histogram distribution features (such as the distribution of inflammatory factor levels and the distribution of cognitive scale scores). Rational quadratic kernel: , Among them, scale parameter Take the median of the Euclidean distance between samples, shape parameter (Default value), parameter Can be set to The value in Can be set to ( The values in the range (distance from standard deviation) are selected through cross-validation. Wavelet kernel: , in The feature dimension, For the first 3D eigenvalues For scale parameters; The standard deviation of the Euclidean distance between samples is taken as . If a significant threshold effect is detected (such as a U-shaped association between uric acid levels and cognitive risk), further five-fold cross-validation can be used. Intra-range optimization; Chi-square core: , in For the sample In the The frequency of each histogram interval (obtained by binning the eigenvalues). For interval numbers, To prevent division by zero for small constants; S32. Based on the fuzzy similarity relationship obtained in step S31, calculate the fuzzy lower approximation and upper approximation of the sample belonging to the decision class; The following approximate definition is: The upper approximation is defined as: S33. Calculate the quality of the rough set approximation of the fuzzy neighborhood of the feature subset with respect to the decision: in Based on feature subsets Calculation of fuzzy similarity relationships For decision-making Balance weights ( For the first (Number of samples per class) to eliminate the effects of class imbalance.
5. The method for reducing key features of cognitive risk in hyperuricemia based on the fusion of self-information and rough set theory according to claim 1, characterized in that, Step S4, which introduces self-information entropy to measure the ability of each feature to distinguish uncertainty in cognitive risk decision-making, includes the following steps: S41. Define the self-information entropy of a feature: To calculate the self-information entropy of continuous features, the feature values need to be discretized. This paper uses an equal-frequency discretization method to divide the value range of each feature into... Each interval contains approximately the same number of samples to preserve as much information as possible about the original distribution. Number of discretized intervals. Based on sample size Adaptive determination, for example ; in Features The set of value intervals obtained after the above discretization For sample values to fall within the interval The probability of it; S42. Define the conditional self-information entropy of a feature relative to a decision: S43. Calculate the self-information gain of the feature: in For decision-making prior entropy. The larger the value, the more characteristic it indicates. The stronger the ability to distinguish between perceived risks.
6. The method for reducing key features of cognitive risk of hyperuricemia based on self-information and rough set fusion as described in claim 1, characterized in that, Step S5, which involves constructing a self-information-rough set fusion importance index and screening an initial key feature candidate set, includes the following steps: S51. Define SRFI for each feature: in For balancing parameters, For a single feature The approximate quality of the rough set in the fuzzy neighborhood; S52. Based on the SRFI values of each feature obtained in step S51, sort the features from largest to smallest, select the number of features whose cumulative SRFI is not lower than a preset threshold, and obtain the initial key feature candidate set.
7. The method for reducing key features of hyperuricemia cognitive risk based on self-information and rough set fusion according to claim 1, characterized in that, Step S6 describes using an adaptive optimization algorithm combined with class-balanced cross-entropy loss to optimize feature weights, which includes the following steps: S61. Define weighted fuzzy similarity relations; Define feature weight vector The weighted fuzzy similarity relations are as follows: in The multi-kernel combination in step S31 is adopted (each feature independently selects the optimal kernel or a fixed kernel); S62, construct the class-balanced cross-entropy loss function: in For class-balanced weights, For soft classification probabilities based on weighted fuzzy similarity relations: S63. Construct the objective function: in and These are hyperparameters that control the strength of L1 sparse regularization and L2 weight decay regularization, respectively. The L1 norm of the feature weights, The L2 norm squared. This objective function achieves sparsity and stability control of feature weights by simultaneously minimizing the classification loss and the regularization term, where... This causes the weights of unimportant features to approach zero. Prevent overfitting; S64. Update feature weights using an adaptive optimization algorithm, such as the RAdam optimizer; The update rule for RAdam is defined as follows: First, calculate the gradient. First-order moment estimation Second-order moment estimation Deviation correction item , and calculate ,in ;when Adaptive learning rate update is used: Otherwise, non-adaptive updates are used (degenerate to SGD): Among them, learning rate , , , RAdam uses dynamic variance correction to automatically degenerate into SGD in the early stages of training, and then activates adaptive momentum after the variance stabilizes, effectively mitigating the risk of overfitting under small, imbalanced data. S65. Feature weight convergence determination and key feature subset selection; When the overall objective function The optimal feature weights are obtained when convergence or when the maximum number of iterations is reached. Remove those with weights below the threshold. (like The features of ) are used to obtain the final key feature subset. .
8. The method for reducing key features of hyperuricemia cognitive risk based on self-information and rough set fusion according to claim 1, characterized in that, The construction of an interpretable cognitive risk classification rule set in step S7 includes the following steps: S71. Based on the final key feature subset and weighted fuzzy similarity relationship, extract the fuzzy decision rules for each decision class: The interval Through the Class samples in features The value distribution on the above is obtained by interval estimation (e.g., mean ± standard deviation); S72, Rule Selection and Quality Assessment; Calculate the support for each rule ( ) and confidence level ( ), Filter support 0.05 and confidence level Rule 0.7 serves as the final interpretation rule; S73. For new samples, output the cognitive risk level according to the matching rules, and give the key risk factors (i.e. the features that appear in the rules and their value ranges).
9. A key feature reduction system for cognitive risk of hyperuricemia based on the fusion of self-information and rough set theory, characterized in that, include: The data acquisition and preprocessing module is used to collect multi-source heterogeneous data from people with hyperuricemia and perform missing value imputation, outlier correction, Min-Max normalization and nonlinear transformation. The multi-kernel fuzzy neighborhood rough set module, whose input is connected to the data acquisition and preprocessing module, is used to calculate the fuzzy similarity relationship between samples based on the combination of multiple kernel functions (such as rational quadratic kernel, wavelet kernel and chi-square kernel), and to calculate the approximate quality of each feature subset; The self-information entropy and feature importance fusion module has its input terminals connected to the multi-kernel fuzzy neighborhood rough set module, which is used to calculate the self-information gain of each feature and fuse the approximate quality to obtain the feature importance score, and to screen the initial key feature candidate set. An adaptive feature weight optimization module, whose input is connected to the self-information entropy and feature importance fusion module, is used to iteratively optimize the feature weights using an adaptive optimization algorithm (such as RAdam) combined with class-balanced cross-entropy loss to obtain the final key feature subset. The interpretable rule generation and risk prediction module has its input connected to the adaptive feature weight optimization module. It is used to extract fuzzy decision rules based on the final key feature subset and output the cognitive risk level and explanation of key risk factors. The visualization and clinical decision support module, whose input is connected to the interpretable rule generation and risk prediction module, is used to display feature reduction results, risk level distribution, importance ranking of key risk factors, and generate individualized risk assessment reports. The model update and incremental learning module, whose input is connected to the data acquisition and preprocessing module, is used to quickly and incrementally update the feature weights and rule sets based on an adaptive optimization algorithm when new data is added.
