Two-channel feature screening method for cognitive impairment recognition modeling and modeling method thereof

Abstract

The invention belongs to the field of intelligent medical treatment. The invention provides a two-channel feature screening method for cognitive impairment recognition modeling and a modeling method thereof, and the method comprises the steps: taking multi-site cognitive impairment screening data as input data, and carrying out the preprocessing of the input data; constructing a traditional robust feature screening channel and an LLM knowledge enhancement screening channel, and performing feature screening on the preprocessed input data; integrating dual-channel feature screening results, performing bias perception joint score optimization, and obtaining a feature set; and performing semantic alignment and version normalization on the feature set to obtain a high-quality feature set. And applying the high-quality feature set to a classifier for cognitive impairment recognition modeling, and generating a clinical decision support result. According to the parallel feature screening method based on combination of a large language model and traditional feature screening, the problem of data heterogeneity bias in cognitive impairment recognition is systematically solved by constructing a dual-channel collaborative feature evaluation architecture, and organic unification of statistical robustness and clinical interpretability is achieved.

Description

Technical Field

[0001] This invention belongs to the field of smart healthcare, and in particular relates to a dual-channel feature screening method and modeling method for cognitive impairment identification and modeling. Background Technology

[0002] Cognitive impairment (including mild cognitive impairment and various types of dementia) has become one of the major global public health challenges. Cognitive impairment identification is a core technique in neuropsychological assessment. It assesses a patient's cognitive function by systematically analyzing their performance patterns on standardized cognitive tests, and is an important means of early diagnosis of neurodegenerative diseases such as Alzheimer's disease and mild cognitive impairment. Traditional cognitive impairment identification mainly relies on clinicians' comprehensive judgment of the results of multidimensional cognitive scales. This manual assessment method requires physicians to have extensive neuropsychological background knowledge and clinical experience. Furthermore, the assessment process is time-consuming and easily affected by factors such as the physician's personal experience level, fatigue, and subjective bias, leading to poor consistency in diagnostic results among different physicians. This is particularly true in primary healthcare institutions where there is a lack of specialized personnel, resulting in a high rate of misdiagnosis and missed diagnosis.

[0003] In recent years, automated cognitive impairment identification based on machine learning has gradually emerged. By learning discrimination patterns from large-scale cognitive test data, it is possible to reduce the influence of human subjectivity on diagnostic results, improve the objectivity, consistency, and efficiency of diagnosis, and show great potential in clinical decision support. However, existing automated cognitive impairment identification based on machine learning faces three core technical challenges in practical applications: First, the heterogeneity of data collection leads to serious domain adaptation problems. Different medical institutions use different versions of cognitive scales (e.g., different language versions of MMSE, different culturally adapted versions of MoCA), scoring standards and testing procedures are not uniform, and the distribution of patient demographic characteristics is uneven (age structure, education level, geographical differences), resulting in significant covariate bias and conceptual drift between training data and target application scenarios. The model experiences a sharp decline in performance when deployed across institutions and populations, limiting the widespread application of the algorithm. Second, traditional statistical feature selection methods lack guidance from medical domain knowledge. Existing methods primarily rely on purely data-driven statistical indicators such as correlation analysis, mutual information, and recursive feature elimination for feature selection, neglecting prior information from fields like neuropsychological theory, clinical diagnostic guidelines, and medical expert knowledge. This results in selected feature combinations that, while statistically significant, lack clinical interpretability and theoretical support, making it difficult to gain the trust and acceptance of clinicians. Third, the mechanisms for ensuring algorithmic fairness across different population groups are inadequate. Existing methods lack systematic bias detection and fairness constraint mechanisms when dealing with patient groups of different ages, genders, educational backgrounds, and cultural backgrounds, making them prone to algorithmic discrimination, particularly potentially exhibiting systematic bias towards disadvantaged groups, thus violating the principle of medical fairness.

[0004] In summary, existing automated cognitive impairment identification models based on machine learning suffer from unsatisfactory identification results due to the heterogeneity of data collection during the feature selection stage, the lack of medical knowledge to guide statistical feature screening methods, and the absence of systematic bias detection. Summary of the Invention

[0005] This invention addresses the problems in existing technologies where data collection heterogeneity during the feature selection stage, the lack of medical knowledge guidance in statistical feature selection methods, and the absence of systematic bias detection lead to unsatisfactory recognition results in current automated cognitive impairment recognition models based on machine learning. The invention provides a dual-channel feature selection method for cognitive impairment recognition modeling, comprising the following steps:

[0006] S1: Preprocess the multi-site cognitive impairment screening data as input data;

[0007] S2: Construct a traditional robust feature selection channel and an LLM knowledge-enhanced selection channel to perform feature selection on the preprocessed input data respectively;

[0008] S3: Integrate the dual-channel feature screening results, perform bias-aware joint scoring optimization, and obtain the feature set;

[0009] S4: Perform semantic alignment and version normalization on the feature set to obtain a high-quality feature set.

[0010] In some embodiments, the multi-site cognitive impairment screening data in S1 includes different versions of the MMSE, MoCA, and ADAS-Cog standardized cognitive scales, demographic information, and clinical diagnostic labels.

[0011] In some embodiments, the preprocessing in S1 includes quality assessment and standardization; the standardization includes: storing data using a hierarchical data structure, which includes a four-level nested structure of site layer, patient layer, scale layer, and item layer.

[0012] In some embodiments, the preprocessing described in S1 further includes cross-site heterogeneity quantification detection to determine the data distribution difference index, covariate shift index, concept drift index, and potential bias source score among different sites.

[0013] In some embodiments, the distribution difference between different sites is calculated using a combination of multivariate Kolmogorov-Smirnov test and energy statistic to calculate the distribution difference index; for continuous variables, the modified KS statistic is calculated; for categorical variables, the chi-square goodness-of-fit test is used; covariate shift detection is based on the density ratio estimation method of discriminator network, training a binary classifier to distinguish samples from different sites, and the classification accuracy of the discriminator reflects the degree of covariate shift, thus obtaining the covariate shift index; by comparing the output distribution differences under the same input conditions between different sites, the conditional concept drift index is calculated.

[0014] In some embodiments, the traditional robust feature selection channel in S2 integrates at least one classical statistical feature selection algorithm for feature selection.

[0015] In some embodiments, when the traditional robust feature selection channel integrates multiple classic statistical feature selection algorithms, the feature selection adopts a weighted voting fusion mechanism to integrate the feature selection results of different algorithms.

[0016] In some embodiments, the classical statistical feature selection algorithm includes the LASSO path frequency algorithm and / or the maximum correlation minimum redundancy algorithm and / or the Boruta random forest algorithm and / or the permutation importance algorithm and / or the SHAP interpretive analysis algorithm and / or the stability selection algorithm and / or the recursive feature elimination algorithm and / or the univariate statistical test algorithm and / or the variance-based screening algorithm and / or the correlation analysis algorithm.

[0017] In some embodiments, the traditional robust feature selection channel employs hierarchical bootstrap stability selection and cross-site cross-validation to ensure the statistical robustness and cross-domain generalization ability of feature selection.

[0018] In some embodiments, the LLM knowledge enhancement channel utilizes a pre-trained medical large language model to perform feature filtering on the preprocessed input data.

[0019] In some embodiments, S3 includes: integrating the dual-channel feature screening results, obtaining a feature set, constructing a bias-aware joint feature scoring function, and achieving a balance between statistical performance and fairness through multi-objective optimization. The bias-aware joint feature scoring function is as follows:

[0020] FS(i)=σ(w T ·f(i)+b)-Bias penalty (i)

[0021] Where FS(i) represents the final joint score of the i-th feature; σ is the sigmoid activation function; w is the learnable weight vector; f(i) is a feature vector that includes at least the dimensions of statistical importance, stability, knowledge relevance, semantic equivalence, and confidence; b is the bias term; Bias penalty (i) represents the bias penalty term for the i-th feature.

[0022] In some embodiments, Bias penalty (i) The calculation formula is as follows:

[0023] Bias penalty (i)=λ1.DI i +λ2.CSI i +λ3.CDI i +λ4.Demobias(i)

[0024] Among them, DI i CSI is the distribution dissimilarity index for the i-th feature. i CDI is the covariate offset index for the i-th feature. i Let λ1, λ2, λ3, and λ4 be the conditional concept drift index for the i-th feature, and let λ1, λ2, λ3, and λ4 be the weight coefficients; Demobias(i) is the score for potential bias sources.

[0025] In some embodiments, S3 includes:

[0026] Construct multiple constraints to obtain the feature set:

[0027]

[0028] Where x is the feature selection indicator vector, F max τ is the upper limit of the number of features. min As the lower bound of performance, δ fair As a fairness threshold, The textAUC represents the total number of features selected. g (x) represents the model's AUC performance when using feature set x on population group g. This means that the condition holds true for all population groups g, where G is the set of population groups; This indicates the difference in AUC performance between different population groups.

[0029] In some embodiments, S4 includes using a multi-level semantic alignment algorithm to solve the semantic equivalence problem between different standardized cognitive scales. The multi-level semantic alignment algorithm is as follows:

[0030] d cluster (i,j)=1-(α·Vocab)sim (i,j)+β·Concept sim (i,j)+γ·Function equiv (i,j))

[0031] Where, d cluster (i,j) represents the cluster distance between items i and j; Vocabulary sim (i,j) represents word-level similarity; Concept sim (i,j) represents concept-level similarity; Function equiv (i,j) represents the functional level equivalence; α, β, and γ are the weight coefficients of the corresponding similarity, and α+β+γ=1.

[0032] In some embodiments, S4 includes constructing a multi-dimensional evidence chain for each item mapping relationship, providing traceable mapping basis for evidence strength assessment and confidence assessment, constructing a unified feature space, and achieving version normalization.

[0033] In some embodiments, the strength calculation of multi-dimensional evidence chains employs a Bayesian evidence fusion framework:

[0034]

[0035] Wherein, P(mapping) i→j |Evidence) represents the posterior probability that item i maps to item j given the evidence; P(Evidence|mapping i→j P(mapping) represents the likelihood probability of observing the current evidence given that the mapping relationship holds; i→j ) represents the prior probability of the mapping relationship; P(Evidence) represents the marginal probability of the evidence; Evidence represents a multi-dimensional set of evidence, which may include, but is not limited to, text similarity evidence, semantic embedding evidence, expert rule evidence, literature support evidence, and clinical validation evidence.

[0036] In some embodiments, a bootstrap resampling method is used to provide a confidence assessment for each item mapping relationship.

[0037] The present invention also provides a cognitive impairment identification and modeling method, which applies the high-quality feature set obtained by the dual-channel feature screening method for cognitive impairment identification and modeling based on any one of the above to cognitive impairment identification and modeling based on machine learning or cognitive impairment identification and modeling based on neural networks.

[0038] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the dual-channel feature screening method for cognitive impairment recognition modeling as described in any of the preceding claims.

[0039] The present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory that can be executed by the processor. When the computer program is executed by the processor, it implements the dual-channel feature screening method for cognitive impairment recognition modeling as described in any of the above claims.

[0040] The beneficial effects of this invention are as follows: By constructing a dual-channel collaborative feature evaluation architecture, this invention performs feature selection based on a large language model combined with a parallel channel of traditional feature selection, thereby obtaining a high-quality feature set. Applying this high-quality feature set to machine learning-based cognitive impairment identification modeling or neural network-based cognitive impairment identification modeling can effectively solve the data heterogeneity bias problem in cognitive impairment identification, achieving an organic unity of statistical robustness and clinical interpretability. Attached Figure Description

[0041] Figure 1 The flowchart of the dual-channel feature screening method used for machine learning cognitive impairment recognition modeling in this embodiment of the invention.

[0042] Figure 2 A flowchart illustrating the preprocessing of input data in an embodiment of the present invention.

[0043] Figure 3 The parallel architecture diagram of the traditional robust feature filtering channel and the LLM knowledge-enhanced filtering channel in the embodiments of the present invention.

[0044] Figure 4 The flowchart of the application of high-quality feature set classifier modeling in clinical practice in the embodiments of the present invention. Detailed Implementation

[0045] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0046] Example

[0047] This example will use a dual-channel feature selection method for machine learning-based cognitive impairment identification modeling to achieve clinical cognitive impairment identification, thus providing a detailed description of the invention. The method flow is as follows: Figure 1 As shown, it includes the following steps:

[0048] S1: Using multi-site cognitive impairment screening data as input, preprocess the data. This data includes different versions of the MMSE, MoCA, and ADAS-Cog standardized cognitive scales, demographic information, and clinical diagnostic labels. Other similar standardized neuropsychological assessment data, such as CDR, FAQ, and NPI scales, have similar item structures, scoring systems, and clinical application scenarios, and can be processed using the same methods.

[0049] Preprocessing methods such as Figure 2 As shown.

[0050] S101, performs quality assessment and standardization processing on multi-site cognitive impairment screening data.

[0051] Construct a unified data representation format. Data standardization processing must have the following characteristics: use a hierarchical data structure for storage, including a four-level nested structure of site layer, patient layer, scale layer, and item layer. Specific requirements include globally unique patient identifiers, standardized item coding system, unified missing value marking specifications, and a complete metadata recording system.

[0052] Furthermore, this example also includes S102, missing mechanism detection; S103, multi-dimensional detection of abnormal response patterns; and S104, demographic stratification labeling. These are commonly used techniques in existing technologies and will not be elaborated upon in this example.

[0053] S105, Cross-site Heterogeneity Quantification Detection. A multi-dimensional heterogeneity quantification index system is designed to systematically evaluate data distribution differences and potential sources of bias among different sites.

[0054] The distributional differences between sites were assessed using a combination of multivariate Kolmogorov-Smirnov tests and energy statistics to calculate the distributional difference index (DI). For continuous variables, a modified KS statistic was calculated. For categorical variables, a chi-square goodness-of-fit test was used. Covariate shift detection was performed using a density ratio estimation method based on a discriminator network. A binary classifier was trained to distinguish samples from different sites. The classification accuracy of the discriminator reflected the degree of covariate shift, resulting in the covariate shift index (CSI). CSI = 2 × (ACC) disc -0.5), where ACC disc This represents the classification accuracy of the discriminator. Concept drift detection calculates the Conditional Concept Drift Index (CDI) by comparing the differences in output distribution under the same input conditions at different sites. Where S is the set of sites and KL is the KL divergence.

[0055] S2: Construct a traditional robust feature selection channel and an LLM knowledge-enhanced selection channel to perform feature selection on the preprocessed input data respectively. For example... Figure 3 As shown, a parallel architecture is constructed for traditional robust feature screening channels and LLM knowledge-enhanced screening channels, which evaluate feature value from statistical and medical knowledge perspectives respectively, thus achieving an organic combination of data-driven and knowledge-driven approaches.

[0056] A. Traditional robust feature selection channel

[0057] Traditional feature selection channels integrate multiple classic statistical feature selection methods to construct a robust feature importance evaluation system, ensuring the statistical significance and cross-domain stability of the selection results. Multi-algorithm ensemble selection employs a weighted voting fusion mechanism to integrate the selection results of different algorithms.

[0058] This example uses LASSO regularized path analysis, which repeatedly performs regressions under a series of regularization parameters to calculate the stable selection probability by statistically analyzing the selection frequency of features at different sparsity levels. Ridge regression uses L2 regularization to control model complexity and assess the linear importance of features. Elastic Net combines the advantages of L1 and L2 regularization, taking into account feature selection and grouping effects when processing high-dimensional data. The combined maximum correlation minimum redundancy (mRMR) algorithm uses mutual information theory to evaluate the correlation between features and the target variable, as well as the redundancy between features. Boruta full feature space importance assessment constructs shadow features for relative importance comparison. Shadow features are generated by randomly permuting original features, preserving the statistical properties of the data but destroying the association with the target variable.

[0059] This can be combined with other traditional feature screening methods, including but not limited to: calculating the importance scores of all features (including shadow features) using the random forest algorithm and using statistical tests to determine whether the true features are significantly better than the best shadow features; Recursive Feature Elimination (RFE) employs a reverse greedy search strategy, starting from the full feature set and gradually removing the features with the lowest importance until a preset number of features or performance threshold is reached; Univariate statistical tests include chi-square tests, F-tests, and mutual information tests, which are applicable to different types of feature and target variable combinations; Variance-based feature screening removes near-constant features with excessively small variances by calculating the variance or coefficient of variation of the features; Correlation analysis calculates the Pearson correlation coefficient, Spearman rank correlation coefficient, or Kendall's tau correlation coefficient between the features and the target variable, selecting features with high correlation; Permutation importance assesses the contribution of features to model performance by randomly shuffling feature values, and calculates the performance decline as an indicator of feature importance; SHAP (SHapley Additive exPlanations), based on the Shapley value theory of game theory, assigns an importance score to each feature, reflecting the marginal contribution of the feature to the model's prediction results.

[0060] Stratified stability selection replicates the feature selection process across multiple demographic strata to assess the cross-group consistency of feature selection. For each feature, its selection frequency vector across different strata is calculated, and a stratified stability score is defined.

[0061]

[0062] Where p i Let be the selection frequency vector of feature i in each stratum, and ∈ be the numerical stability constant. The bias sensitivity score identifies potential sources of algorithmic bias by quantifying the performance differences of features across different population strata. A comprehensive bias sensitivity score system is constructed by calculating the differences in predictive performance, calibration offset, and fairness indices of features across different groups.

[0063] B.LLM Knowledge Enhancement Filtering Channel

[0064] The LLM knowledge enhancement pathway leverages the semantic understanding, knowledge reasoning, and multimodal fusion capabilities of large language models to assess the clinical relevance, theoretical support, and strength of evidence of features from a medical knowledge perspective.

[0065] In this example, semantic embedding extraction and alignment employ a pre-trained medical language model to perform deep semantic encoding on the cognitive test items in the input data. The semantic representation of each item not only includes the question text but also integrates multi-dimensional information such as option descriptions, scoring rules, and cognitive domain labels. To improve the accuracy of cross-language and cross-cultural semantic alignment, a contrastive learning mechanism is used to train the semantic embedding space, ensuring that semantically equivalent items have similar representations in the embedding space. The pre-trained medical language model incorporates prior knowledge from medical knowledge bases, clinical guidelines, and neuropsychological theories. Leveraging the semantic understanding and reasoning capabilities of a large language model, it evaluates feature value from dimensions such as clinical relevance, theoretical support, and strength of evidence, thereby improving the clinical interpretability of feature selection. Pre-trained medical language models are a common technique used by those skilled in the art; therefore, the modeling and training methods of the medical language model will not be elaborated upon in this example.

[0066] Cross-version semantic equivalence calculation is based on a fusion method of multiple similarity measures, comprehensively considering text similarity, semantic embedding similarity, and cognitive domain matching degree. An uncertainty quantification mechanism is introduced, and the confidence interval of semantic equivalence is calculated through bootstrap resampling to provide a credibility assessment for the mapping relationship.

[0067] The Knowledge Relevance Rating (KRS) employs a multi-turn dialogue and chain-reasoning prompting engineering strategy to design a structured medical knowledge assessment framework. First, a medical knowledge graph retrieval system is constructed to retrieve relevant medical concepts, clinical guidelines, and research literature based on the semantic vectors of the items. Multi-dimensional assessment prompt templates are designed, allowing LLM (Medical Knowledge Management) researchers to evaluate the clinical value of items from dimensions such as cognitive domain specificity, pathological sensitivity, clinical applicability, literature support, and cross-cultural applicability. A temperature-controlled multiple sampling strategy is used to generate a score distribution, and the expected value and variance of the knowledge relevance scores are calculated.

[0068] The bias risk assessment employs a structured, multi-dimensional analytical framework, enabling LLM to identify and quantify potential sources of bias from various perspectives, including cultural, educational, age, and linguistic biases. It generates structured bias risk annotations, including risk levels, impact mechanisms, and mitigation strategies. A self-consistency calibration mechanism enhances the reliability of LLM assessments through multi-agent voting and uncertainty modeling. A judge LLM is designed to perform consistency checks on multiple assessment results, calculating consistency metrics and confidence scores among the assessment outcomes.

[0069] S3: Integrate the dual-channel feature screening results, perform bias-aware joint scoring optimization, and obtain the feature set.

[0070] This example integrates the evaluation results of the traditional channel and the LLM channel to construct a bias-aware joint feature scoring function, achieving a balance between statistical performance and fairness through multi-objective optimization.

[0071] A. Joint scoring function based on adaptive weight learning

[0072] An adaptive weight learning mechanism is designed to dynamically adjust the weights of different scoring dimensions based on data features and task requirements. The joint scoring function employs a non-linear fusion strategy.

[0073] FS(i)=σ(w T ·f(i)+b)-Bias penalty (i)

[0074] Where FS(i) represents the final joint score of the i-th feature; σ is the sigmoid activation function; w is the learnable weight vector; f(i) is a feature vector that includes at least the dimensions of statistical importance, stability, knowledge relevance, semantic equivalence, and confidence; b is the bias term; Bias penalty (i) represents the bias penalty term for the i-th feature, calculated as follows:

[0075] Bias penalty (i)=λ1.DI i +λ2.CSI i +λ3.CDI i+λ4.Demobias(i)

[0076] Among them, DI i CSI is the distribution dissimilarity index for the i-th feature. i CDI is the covariate offset index for the i-th feature. i Let λ1, λ2, λ3, and λ4 be the conditional concept drift index for the i-th feature, and let λ1, λ2, λ3, and λ4 be the weighting coefficients. Demobias(i) is the potential bias source score, which can be a bias score based on demographic factors such as age, gender, education, site, and culture.

[0077] The statistical importance refers to the feature importance score calculated using traditional statistical methods, including the LASSO coefficient, mRMR score, Boruta importance, etc.; the stability refers to the consistency of feature selection across different data stratifications and cross-validation; the knowledge relevance refers to the feature clinical relevance score assessed using a large language model; the semantic equivalence refers to the degree of semantic similarity between items in different versions of the scale; and the confidence score refers to the credibility score of the assessment results.

[0078] B. Constrained Optimization Feature Selection

[0079] Feature selection is modeled as a multi-constraint integer programming problem, minimizing the number of features while ensuring performance and fairness. The optimization problem includes multiple constraints, such as the number of features, performance lower bound constraints, fairness constraints, and cognitive domain coverage constraints.

[0080]

[0081] Where x is the feature selection indicator vector, F max τ is the upper limit of the number of features. min As the lower bound of performance, δ fair As a fairness threshold, The textAUC represents the total number of features selected. g (x) represents the model's AUC performance when using feature set x on population group g. This means that the condition holds true for all population groups g, where G is the set of population groups; This indicates the difference in AUC performance between different population groups.

[0082] S4: Perform semantic alignment and version normalization on the feature set to obtain a high-quality feature set.

[0083] A cross-version item mapping system is constructed based on medical language models and clinical ontology. A multi-dimensional evidence chain is built for each item mapping relationship, providing traceable evidence strength assessment and confidence assessment of the mapping basis. This solves the semantic equivalence problem between different scale versions and constructs a unified feature space to achieve version normalization.

[0084] A. Multi-level semantic alignment algorithm

[0085] A multi-layered semantic alignment strategy is designed, mapping cognitive table items at three levels: lexical, conceptual, and functional. Lexical alignment involves text preprocessing based on a medical dictionary and thesaurus to construct a standardized medical vocabulary mapping table. Conceptual alignment utilizes a medical knowledge graph for semantic reasoning, calculating the semantic distance between concepts. Functional alignment is evaluated based on the equivalence of cognitive function measurements, and a method for calculating cognitive domain correspondence is designed.

[0086] The item clustering uses a hierarchical clustering algorithm, and the cluster distance is defined as a weighted combination of the alignment scores across multiple levels:

[0087] d cluster (i,j)=1-(α·Vocab) sim (i,j)+β·Concept sim (i,j)+γ·Function equiv (i,j))

[0088] Where, d cluster (i,j) represents the cluster distance between items i and j; Vocabulary sim (i,j) represents word-level similarity; Concept sim (i,j) represents concept-level similarity; Function equiv (i,j) represents the functional level equivalence; α, β, and γ are the weight coefficients of the corresponding similarity, and α+β+γ=1.

[0089] The clustering stopping criterion is based on a dual determination of the silhouette coefficient and domain knowledge constraints, ensuring that the clustering results conform to neuropsychological theories.

[0090] B. Evidence Chain Generation and Credibility Assessment

[0091] A multi-dimensional chain of evidence is constructed for each item mapping relationship, providing traceable mapping basis and confidence assessment. Evidence collection employs a multi-source information fusion strategy, with evidence types including text similarity evidence, semantic embedding evidence, expert rule evidence, literature support evidence, and clinical validation evidence. Evidence strength calculation utilizes a Bayesian evidence fusion framework.

[0092]

[0093] Wherein, P(mapping) i→j |Evidence) represents the posterior probability that item i maps to item j given the evidence; P(Evidence|mapping i→j P(mapping) represents the likelihood probability of observing the current evidence given that the mapping relationship holds; i→j ) represents the prior probability of the mapping relationship; P(Evidence) represents the marginal probability of the evidence; Evidence represents a multi-dimensional set of evidence, which may include, but is not limited to, text similarity evidence, semantic embedding evidence, expert rule evidence, literature support evidence, and clinical validation evidence.

[0094] The prior probabilities are based on expert knowledge, and the likelihood function is calculated using evidence reliability weights. The confidence interval estimation employs a bootstrap resampling method to provide a statistical confidence assessment for each item mapping.

[0095] S5: This example applies a high-quality feature set to a classifier for cognitive impairment identification modeling, generating clinical decision support results. For example... Figure 4 As shown.

[0096] Commonly used classifiers include, but are not limited to, those based on traditional statistics, those based on machine learning, and those based on neural networks. Traditional classifiers include logistic regression. Classifiers based on machine learning algorithms include: Support Vector Machine (SVM) classifiers, which use radial basis function (RBF) kernels and optimize hyperparameters C and gamma through grid search to find the optimal separating hyperplane in a high-dimensional feature space; Random Forest classifiers, which construct multiple decision trees and use voting to make decisions, exhibit good generalization ability and feature importance interpretation capabilities; and Gradient Boosting Decision Tree (GBDT) classifiers, which progressively correct prediction errors, perform excellently when handling complex nonlinear relationships.

[0097] Alternatively, a classifier based on deep neural networks could be considered. Deep neural networks employ a multilayer perceptron architecture, incorporating dropout and batch normalization techniques to prevent overfitting, and learn complex feature interaction patterns through backpropagation. Logistic regression, as a linear baseline model, provides highly interpretable diagnostic criteria. The Naive Bayes classifier, based on the conditional independence assumption, maintains stable performance even with small sample sizes.

[0098] In addition, a multi-classifier ensemble strategy can be introduced, employing a soft voting mechanism to combine the predicted probabilities of each classifier:

[0099]

[0100] P ensemble(y = 1|x) represents the probability that the ensemble model predicts sample x as cognitive impairment (category 1), K is the number of classifiers, and w k The weights of the k-th classifier are determined by the performance on the validation set, P k This represents the probability that the k-th classifier predicts sample x as having cognitive impairment.

[0101] Ultimately, the constructed cognitive impairment identification model will be applied to the diagnostic decision-making system.

[0102] This example constructs a multi-level diagnostic decision system that provides personalized screening strategies based on the severity of cognitive impairment and clinical needs. Level 1 screening employs a high-sensitivity model to prioritize the detection of potential cognitive impairment patients, reducing the false negative rate. A low threshold is set to ensure that patients with early cognitive decline are identified promptly. Level 2 refined diagnosis uses a high-specificity model to further differentiate the type and severity of cognitive impairment, reducing the false positive rate. Level 3 dynamic monitoring establishes a longitudinal tracking model to assess the trend of cognitive function changes by comparing historical data.

[0103] The risk scoring system generates an individualized cognitive impairment risk score, which is then adjusted for covariates such as age, education, and genetic factors.

[0104]

[0105] Where, x i For the selected cognitive features, covariate j For demographic covariates, β i and γ j β0 represents the corresponding regression coefficient, where Risk Score represents the individualized cognitive impairment risk score and β0 represents the baseline risk intercept.

[0106] The diagnostic decision system provides physicians with standardized tools for identifying cognitive impairments and diagnostic recommendations.

[0107] The diagnostic report can be generated into modules including basic patient information, cognitive test results, risk score, explanations of key features, and clinical recommendations. Feature contribution analysis uses the SHAP value method to provide a quantitative explanation of the contribution of each key feature, helping physicians understand the diagnostic basis.

[0108] In practice, quality control mechanisms can be set up, including model performance monitoring, prediction uncertainty assessment, and anomaly detection. When the model's prediction confidence is low, the system will remind physicians to conduct manual review. A feedback learning mechanism is established to collect physicians' diagnostic results and opinions, continuously improving model performance. The clinical validation process includes prospective cohort studies, multi-center clinical trials, and real-world data validation to ensure the algorithm's effectiveness and safety in real-world clinical environments. Standardized usage guidelines and training programs are developed to support the algorithm's widespread application in different medical institutions.

[0109] Through the above classifier adaptation and clinical application design, the feature selection method in this example can be seamlessly integrated with various machine learning algorithms, providing technical support for the accurate identification and personalized diagnosis and treatment of cognitive impairment, and has broad clinical application prospects and practical value.

[0110] In summary, during the feature selection stage, this invention designs a parallel architecture combining a traditional robust feature selection channel and an LLM knowledge-enhanced selection channel. The former integrates various classic methods such as LASSO path frequency, maximum relevance and minimum redundancy, Boruta random forest, permutation importance, SHAP interpretability analysis, stability selection, recursive feature elimination, univariate statistical tests, variance-based selection, and correlation analysis, combined with hierarchical bootstrap stability selection and cross-site cross-validation, ensuring the statistical robustness and cross-domain generalization ability of feature selection. The latter, in the bias mitigation stage, proposes a bias-sensitive quantification method based on multi-dimensional hierarchical analysis and a joint scoring algorithm based on multi-objective constraint optimization. By explicitly modeling the impact of sensitive attributes such as age, gender, education, and region on feature importance, it constructs a bias-aware feature scoring function, minimizing cross-group performance differences while ensuring overall performance. In the semantic alignment stage, a cross-version item mapping algorithm based on a medical language model and clinical ontology is designed. Through multiple mechanisms such as semantic embedding similarity calculation, medical concept alignment, and expert rule verification, it solves the semantic equivalence problem between different scale versions and constructs a unified feature space representation. This systematically addresses the data heterogeneity bias problem in the identification of cognitive impairment, achieving an organic unity between statistical robustness and clinical interpretability.

[0111] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0112] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that flowchart illustrations and / or block diagrams of computer program products can be used to describe them. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0113] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0114] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0115] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. Memory may include non-persistent memory in computer-readable media, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media. Computer-readable media includes both permanent and non-persistent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information that can be accessed by the computing device. As defined in this article, computer-readable media do not include transient media, such as modulated data signals and carrier waves.

[0118] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0119] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A dual-channel feature screening method for cognitive impairment identification modeling, characterized in that, The method comprises the following steps: S1: taking multi-site cognitive impairment screening data as input data, and preprocessing the same; S2: constructing a traditional robust feature screening channel and an LLM knowledge enhanced screening channel, and screening the preprocessed input data respectively; S3: integrating the screening results of the two channels, performing bias perception joint scoring optimization, and obtaining a feature set; S4: performing semantic alignment and version normalization on the feature set to obtain a high-quality feature set.

2. The dual channel feature screening method for cognitive impairment identification modeling according to claim 1, wherein, The multi-site cognitive impairment screening data in S1 includes MMSE, MoCA, ADAS-Cog standardized cognitive scale data of different versions, demographic information and clinical diagnosis labels.

3. The cognitive impairment identification modeling dual-channel feature screening method according to claim 2, characterized in that, The preprocessing in S1 includes quality evaluation and standardization processing; The standardization processing includes: storing by using a hierarchical data structure, and containing a four-level nested structure of a site layer, a patient layer, a scale layer and an item layer; the preprocessing in S1 further includes cross-site heterogeneity quantification detection, and determining data distribution difference indexes, covariate offset indexes, concept drift indexes and potential bias source scores between different sites; the data distribution difference between different sites is calculated by using a combination of multivariate Kolmogorov-Smirnov test and energy statistics; for continuous variables, a modified K-S statistic is calculated; for categorical variables, a chi-square goodness-of-fit test is used; the covariate offset detection is based on a density ratio estimation method of a discriminator network, a binary classifier is trained to distinguish samples of different sites, the classification accuracy of the discriminator reflects the degree of covariate offset, and a covariate offset index is obtained; the conditional concept drift index is calculated by comparing the output distribution difference under the same input condition between different sites.

4. The dual-channel feature screening method for cognitive impairment recognition modeling according to any one of claims 3, characterized in that, The traditional robust feature screening channel in S2 integrates at least one classical statistical feature selection algorithm for feature screening; when the traditional robust feature screening channel integrates multiple classical statistical feature selection algorithms, a weighted voting fusion mechanism is used to integrate the feature screening results of different algorithms.

5. The cognitive impairment identification modeling dual-channel feature screening method according to claim 4, characterized in that, The classical statistical feature selection algorithm includes a LASSO path frequency algorithm and / or a maximum relevance minimum redundancy algorithm and / or a Boruta random forest algorithm and / or a permutation importance algorithm and / or a SHAP interpretability analysis algorithm and / or a stability selection algorithm and / or a recursive feature elimination algorithm and / or a univariate statistical test algorithm and / or a variance-based screening algorithm and / or a correlation analysis algorithm; the LLM knowledge enhanced channel uses a pre-trained medical large language model to screen the preprocessed input data.

6. The cognitive impairment identification modeling dual-channel feature screening method according to claim 5, characterized in that, S3 includes: integrating the screening results of the two channels, obtaining a feature set, constructing a bias perception joint feature scoring function, balancing statistical performance and fairness through multi-objective optimization, and the bias perception joint feature scoring function is as follows: FS(i) = σ(w T • f(i) + b) - Bias penalty (i) wherein FS(i) represents a final joint score of the i-th feature; σ is a sigmoid activation function; w is a learnable weight vector; f(i) is a feature vector comprising at least a statistical importance, a stability, a knowledge relevance, a semantic equivalency, a confidence dimension; b is a bias term; Bias penalty (i) represents a bias penalty term of the i-th feature.

7. The cognitive impairment identification modeling dual-channel feature screening method according to claim 6, wherein Bi as penalty (i) The calculation formula is as follows: Bias penalty (i) = λ1.DI i + λ2.CSI i + λ3.CDI i + λ4.Demobias(i) where DI i is the distribution difference index of the i-th feature, CSI i is the covariate drift index of the i-th feature, CDI i is the conditional concept drift indicator of the i-th feature, λ1, λ2, λ3, λ4are weight coefficients; Demobias(i) is the potential bias source score.

8. The cognitive impairment identification modeling dual-channel feature screening method according to claim 7, characterized in that, S3 includes: Constructing multiple constraint conditions to obtain a feature set: where x is the feature selection indicator vector, F max is the upper bound of the number of features, τ min is the lower bound of the performance, δ fair is the fairness threshold, denotes the total number of selected features, textAUC g (x) denotes the model AUC performance on population group g using the feature set x, denotes that it holds for all population groups g, where G is the set of population groups; denotes the difference in AUC performance across different population groups.

9. The cognitive impairment identification modeling dual-channel feature screening method according to claim 8, characterized in that, The multi-level semantic alignment algorithm in S4 solves the semantic equivalence problem between different standardized cognitive scales, and the multi-level semantic alignment algorithm is as follows: d cluster (i,j) = 1 - (a - Vocab sim (i,j) + β - Concept sim (i,j) + γ - Function equiv (i,j)) where d cluster (i,j) represents the clustering distance between items i and j; Vocab sim (i,j) represents the word-level similarity; Concept sim (i,j) represents the concept-level similarity; Function equiv (i,j) represents the function-level equivalence; α, β, γ are the weight coefficients of the corresponding similarity, and α+β+γ=1; The S4 comprises a mapping relationship for each question item to construct a multi-dimensional evidence chain, provides evidence strength evaluation and confidence evaluation of traceable mapping basis, constructs a unified feature space, and realizes version normalization; the evidence strength calculation of the multi-dimensional evidence chain adopts a Bayesian evidence fusion framework: where P (Evidence | mapping i→j i→j) denotes the posterior probability of item i mapping to item j given the evidence; P (Evidence | mapping i→j i→j) denotes the likelihood of observing the current evidence given the mapping relationship; P (mapping i→j i→j) denotes the prior probability of the mapping relationship; P (Evidence) denotes the marginal probability of the evidence; Evidence denotes the multi-dimensional evidence set, which can include but is not limited to text similarity evidence, semantic embedding evidence, expert rule evidence, literature support evidence, and clinical verification evidence; bootstrap resampling method is used to provide confidence evaluation for each item mapping relationship.

10. A cognitive impairment identification modeling method, characterized by, The high-quality feature set obtained by the dual-channel feature screening method for cognitive impairment identification modeling according to any one of claims 1-9 is applied to machine learning-based cognitive impairment identification modeling or neural network-based cognitive impairment identification modeling.

Images