Cognitive function screening system and method based on MMSE prediction model

Abstract

The invention discloses a cognitive function screening system and method based on an MMSE (Minimum Mean Square Error) prediction model, and relates to the technical field of data analysis, the method comprises the following steps: collecting physical examination index data, removing missing records, adopting a multiple interpolation method for interpolation, and obtaining multiple sets of complete data sets; a continuous prediction model is constructed, MMSE continuous prediction values are obtained, and a sensitivity analysis report is generated; constructing a first-stage classification model, adaptively dividing an optimal threshold combination, and dividing a sample into a high-confidence region, a to-be-discriminated region and a low-confidence region; if the sample size of the to-be-discriminated region is higher than a preset training threshold value, constructing an enhanced feature set, and constructing a second-stage classification model; when the prediction probability reaches the optimal re-discrimination threshold value, the classification result in the first stage is corrected, and otherwise, the classification result is maintained; if not, maintaining the classification result; and integrating the classification results to obtain a final classification result of all the samples.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, specifically to a cognitive function screening system and method based on the MMSE prediction model. Background Technology

[0002] Cognitive impairment is a significant health problem affecting the quality of life of middle-aged and elderly people, and early screening is crucial for delaying disease progression. The Mini-Mental State Examination (MMSE) is a commonly used clinical tool for assessing cognitive function, but it is highly specialized and time-consuming, making it difficult to conduct on a large scale during routine physical examinations.

[0003] Currently, predictive model research based on physical examination data faces challenges due to the widespread incompleteness of medical data, making it difficult to guarantee the medical validity of the data using traditional imputation methods. Existing models are mostly based on single algorithms, resulting in limited accuracy in identifying mild cognitive impairment, which falls into a "grey area." Furthermore, many models use fixed thresholds, lacking adaptability to different population characteristics. Finally, most predictive models lack interpretability, making it difficult for clinicians to understand the decision-making basis of the models, leading to insufficient trust. Although some research has attempted to apply machine learning to cognitive function prediction, a systematic solution that organically integrates continuous prediction, dynamic threshold optimization, multi-stage discrimination, and clinical interpretability has yet to be found. Summary of the Invention

[0004] The purpose of this invention is to provide a cognitive function screening system and method based on the MMSE prediction model to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a cognitive function screening method based on the MMSE prediction model, the method comprising: S100. Collect physical examination index data, remove missing records of MMSE (Simplified Mental State Examination), and impute the remaining missing values ​​using multiple imputation. Multiple imputation is based on Bayesian regression to establish the conditional distribution relationship between variables, and updates the missing values ​​through iterative prediction. Clinical safety range correction is also performed. The clinical safety range includes the range of electrolyte concentration, the physiological range of blood pressure, and the reasonable range of blood biochemical indicators, resulting in multiple complete datasets. S200. Using the multiple complete datasets as input, construct a continuous prediction model to obtain continuous MMSE prediction values; perform sensitivity analysis based on the continuous prediction model to generate a sensitivity analysis report containing the marginal impact of each feature on the prediction results. S300: Based on the feature variables of multiple complete datasets and the continuous MMSE prediction values ​​as input features, a first-stage classification model is constructed; and the optimal threshold combination is divided through an automatic parameter optimization algorithm; based on the class probabilities output by the first-stage classification model and the optimal threshold combination, the samples are divided into high-confidence regions, undiscriminated regions and low-confidence regions. S400. If the sample size of the region to be discriminated is higher than the preset training threshold, an enhanced feature set is constructed, and a second-stage classification model is built. The second-stage classification model adopts the gradient boosting decision tree algorithm, and a classifier is constructed based on histogram feature discretization and leaf node growth strategy. When the predicted probability of a sample in the region to be discriminated reaches the optimal re-discrimination threshold in the second-stage classification model, its classification result in the first stage is corrected; otherwise, it is maintained. If the sample size of the region to be discriminated is not higher than the preset training threshold, the first-stage classification result of all samples in the region to be discriminated is maintained. S500: Integrate the classification results of high-confidence regions with the classification results of regions to be discriminated to obtain the final classification results for all samples.

[0006] According to the above scheme, step S100 includes: S110. Collect physical examination indicator data, which includes demographic characteristics and laboratory test indicators. Demographic characteristics include age, gender, and body mass index; laboratory test indicators include blood pressure parameters, blood routine parameters, liver and kidney function parameters, electrolyte parameters, and metabolic parameters; based on a unique identifier, aggregate duplicate test records under the same identifier, and convert all indicators into a unified numerical data format to form a multi-source dataset. S120. Perform key variable integrity and data validity checks on the multi-source dataset, and remove records with missing MMSE, age, and gender fields; S130. For the remaining missing values ​​in the multi-source dataset, iterative imputation is performed using the multiple imputation method to construct a feature matrix containing the target imputation column and the auxiliary imputation column. The feature matrix is ​​then subjected to feature standardization, and multiple sets of initial complete data are generated through the iterative imputation algorithm. S140. Perform clinical medical safety range correction on the interpolated data to adjust the values ​​of each indicator to the preset physiological reasonable range. S150. Based on the corrected data, construct the multiple complete datasets.

[0007] According to the above scheme, step S200 includes: S210. Establish a continuous prediction model, wherein the continuous prediction model uses age and at least one laboratory test indicator to model through a nonlinear smoothing function, and the remaining physical examination indicators are incorporated into the continuous prediction model in the form of linear terms. S220. Based on multiple complete datasets, train a continuous prediction model, use cross-validation to evaluate the prediction performance of the continuous prediction model, and record the root mean square error, mean absolute error and coefficient of determination. S230. Calculate the partial derivatives of each feature variable through sensitivity analysis, generate a feature sensitivity matrix, and quantify the change in MMSE prediction score for each input feature when the value increases by one unit. S240. Output a sensitivity analysis report containing smooth function curve parameters, linear term coefficients, and partial derivative matrices. The sensitivity analysis report is used to explain the marginal impact of each input feature on the cognitive function assessment results.

[0008] According to the above scheme, step S300 includes: S310. Based on age, gender, body mass index, blood pressure parameters, blood routine parameters, liver and kidney function parameters, electrolyte parameters, metabolic parameters and MMSE continuous predicted values ​​from multiple complete datasets as input features, a multi-class logistic regression model is constructed as the first-stage classification model. S320. Through an automatic parameter optimization algorithm, parameter optimization is performed in the preset high confidence threshold search space, the lower bound threshold search space of the region to be judged, and the optimal re-discrimination threshold search space to determine the optimal high confidence threshold, the optimal lower bound threshold of the region to be judged, and the optimal re-discrimination threshold. S330. Based on the probability vectors of the samples belonging to each category output by the first-stage classification model, extract the maximum category probability value; when the maximum category probability value is greater than or equal to the optimal high confidence threshold, classify the corresponding sample into the high confidence region; when the maximum category probability value is less than the optimal high confidence threshold and is simultaneously greater than or equal to the optimal lower bound threshold of the undiscriminated region, classify the corresponding sample into the undiscriminated region; when the maximum category probability value is less than the optimal lower bound threshold of the undiscriminated region, classify the corresponding sample into the low confidence region.

[0009] According to the above scheme, step S400 includes: S410. Determine whether the number of training samples in the region to be discriminated is higher than the preset training threshold; if not, maintain the first-stage classification result of all samples in the region to be discriminated. S420. If the value is higher, then for the sample in the region to be judged, an enhanced feature set is constructed. The enhanced feature set includes the probability of belonging to each category output by the first-stage classification model, interaction terms or ratio features constructed based on clinical indicators, and boundary correlation features constructed based on MMSE continuous prediction values. S430. Using the enhanced feature set as input features, construct a second-stage classification model based on gradient boosting decision tree, and perform a binary classification task of suspected mild cognitive impairment category and non-suspected mild cognitive impairment category. S440. For each sample in the region to be discriminated, if its predicted probability value in the second-stage classification model is greater than or equal to the optimal re-discrimination threshold, its final classification result is corrected to the suspected mild cognitive impairment category; otherwise, its classification result obtained in the first-stage classification model is maintained.

[0010] According to the above scheme, step S500 includes: S510. The classification results of the high-confidence region, the classification results of the low-confidence region, and the classification results of the region to be judged after being judged and corrected by the second-stage classification model are merged with the sample classification results that directly maintain the first-stage classification results to form a complete set of classification results for all samples. S520. Based on the complete set of classification results, output a comprehensive evaluation report. The comprehensive evaluation report includes cognitive function grading results, confidence level, MMSE continuous predictive value, feature sensitivity analysis summary, and classification confidence description.

[0011] A cognitive function screening system based on the MMSE prediction model includes: a data processing module, a continuous prediction module, a hierarchical decision-making module, an enhanced discrimination module, and an integrated output module. The data processing module is used to collect physical examination index data, remove missing MMSE records, impute the remaining missing values ​​and correct for clinical safety ranges to obtain multiple complete datasets. The continuous prediction module is used to construct a continuous prediction model to obtain continuous MMSE prediction values, perform sensitivity analysis based on the continuous prediction model, and generate a sensitivity analysis report that includes the marginal impact of each feature on the prediction results. The hierarchical decision module is used to build the first-stage classification model. It optimizes the algorithm to divide the optimal threshold combination and divides the samples into high-confidence regions, undiscriminated regions, and low-confidence regions based on the class probability. The enhanced discrimination module is used to construct an enhanced feature set and build a second-stage classification model when there are sufficient samples in the region to be discriminated, and to correct the classification results based on the predicted probability. The integrated output module is used to integrate the classification results from various regions and generate the final classification results and comprehensive evaluation report.

[0012] According to the above scheme, the data processing module includes a data acquisition unit, a data cleaning unit, a data imputation unit, and a clinical correction unit. The data acquisition unit is used to collect physical examination indicator data and aggregate duplicate test records based on unique identifiers to form a multi-source dataset. The data cleaning unit is used to perform a key variable integrity check on the multi-source dataset and remove records with missing key fields. The data imputation unit is used to perform iterative imputation on the remaining missing values ​​using a multiple imputation method to generate multiple sets of initial complete data. The clinical correction unit is used to perform clinical medical safety range correction on the imputed data and calculate the albumin to globulin concentration ratio to obtain multiple sets of complete datasets.

[0013] According to the above scheme, the hierarchical decision module includes a model training unit, a threshold optimization unit, and a region partitioning unit. The model training unit is used to build a first-stage classification model based on the feature variables and continuous predicted values ​​of the complete dataset. The threshold optimization unit is used to determine the optimal threshold combination through an automatic parameter optimization algorithm. The region partitioning unit is used to divide the samples into regions with different confidence levels based on the class probabilities output by the first-stage classification model and the optimal threshold combination.

[0014] According to the above scheme, the enhanced discrimination module includes a feature construction unit and a specialized discrimination unit. The feature construction unit is used to construct an enhanced feature set containing the first-stage probability output, clinical indicator interaction terms, and boundary distance features when the sample size of the region to be discriminated is higher than a preset threshold. The specialized discrimination unit is used to construct a second-stage classification model based on the enhanced feature set and to correct the classification results according to the predicted probability.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention lowers the threshold for cognitive function screening by constructing a predictive model using routine physical examination indicators; 2. This invention employs a combination of continuous prediction and two-stage classification, and enhances discrimination through dynamic threshold division and region of discrimination, thereby improving the accuracy and recall rate of identification for the key population with mild cognitive impairment. 3. This invention ensures the integrity and rationality of input data and improves the reliability and applicability of the model through a linkage mechanism of multiple iterative interpolation and clinical safety threshold correction. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of the cognitive function screening method based on the MMSE prediction model of the present invention. Figure 2 This is a schematic diagram of the cognitive function screening system based on the MMSE prediction model of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example: Figures 1-2 As shown, this invention provides a technical solution: a cognitive function screening method based on the MMSE prediction model, the method comprising: S100. Collect physical examination index data, remove missing MMSE records, and impute the remaining missing values ​​using multiple imputation. The multiple imputation method establishes the conditional distribution relationship between variables based on Bayesian regression, updates the missing values ​​through iterative prediction, and performs clinical safety range correction. The clinical safety range includes the electrolyte concentration range, the physiological range of blood pressure, and the reasonable range of blood biochemical indicators, and obtains multiple complete datasets. Specifically, step S100 includes: S110. Collect physical examination indicator data, which includes demographic characteristics and laboratory test indicators. Demographic characteristics include age, gender, and body mass index; laboratory test indicators include blood pressure parameters, blood routine parameters, liver and kidney function parameters, electrolyte parameters, and metabolic parameters; based on a unique identifier, aggregate duplicate test records under the same identifier, and convert all indicators into a unified numerical data format to form a multi-source dataset. For example: Collecting physical examination data from 2876 cases at a physical examination center, including 33 physical examination indicators, and aggregating duplicate test records under the same identifier based on the individual's unique identifier, converting all indicators into a unified numerical format to form a multi-source dataset; this is just an example and is not a limitation; unique identifier aggregation avoids data bias caused by multiple tests on the same subject and ensures sample independence; S120. Perform key variable integrity and data validity checks on the multi-source dataset, and remove records with missing MMSE, age, and gender fields; S130. For the remaining missing values ​​in the multi-source dataset, iterative imputation is performed using the multiple imputation method to construct a feature matrix containing the target imputation column and the auxiliary imputation column. The feature matrix is ​​then subjected to feature standardization, and multiple sets of initial complete data are generated through the iterative imputation algorithm. For example, after aggregating physical examination IDs, 2592 independent samples were obtained. After removing missing MMSE records, 2308 samples remained for analysis. Target imputation columns (systolic blood pressure, diastolic blood pressure, sodium, potassium, chloride, blood glucose, glycated hemoglobin, albumin, total protein, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol) and auxiliary imputation columns (age, gender, BMI, weight, hemoglobin, red blood cells, hematocrit, AGR, total protein, albumin, triglycerides, gamma-glutamyl transferase, urea nitrogen, creatinine, uric acid, blood glucose) were constructed. After standardization, iterative imputation using multiple imputation methods was employed. For instance, a conditional distribution relationship was established based on Bayesian regression, iterated 20 times, and model convergence analysis was performed based on historical data testing to generate 5 sets of initial complete data. The optimal number of data sets balancing computational efficiency and robustness was determined through Bootstrap sampling error analysis. This is merely an example and not a limitation. S140. Perform clinical safety range correction on the imputed data, adjusting the values ​​of each indicator to the preset physiologically reasonable range. For example, perform clinical safety range correction to adjust sodium to 120-160 mmol / L, potassium to 2.5-6.5 mmol / L, chloride to 90-110 mmol / L, systolic blood pressure to 80-200 mmHg, and diastolic blood pressure to 50-120 mmHg. Correct the values ​​of 32 indicators that are outside the range to the corresponding intervals to ensure that all indicators meet the clinical safety standards. The clinical safety range correction refers to the clinical laboratory reference range and sets a physiologically reasonable range to avoid outliers interfering with model training. This is only an example and is not a limitation. S150. Based on the corrected data, construct the multiple complete datasets. In this embodiment, the albumin to globulin concentration ratio is calculated as a derived feature and incorporated into the dataset. The albumin to globulin concentration ratio is calculated using the formula: AGR = Alb / (TP - Alb), where AGR represents the albumin to globulin concentration ratio, Alb represents the albumin concentration, and TP represents the total protein concentration.

[0019] S200. Using the multiple complete datasets as input, construct a continuous prediction model. In the continuous prediction model, age and at least one laboratory test indicator are used to model the model through a nonlinear smoothing function to obtain the continuous prediction value of MMSE. Based on the continuous prediction model, perform sensitivity analysis and generate a sensitivity analysis report containing the marginal impact of each feature on the prediction result. Specifically, step S200 includes: S210. Establish a continuous prediction model, wherein the continuous prediction model uses age and at least one laboratory test indicator to model through a nonlinear smoothing function, and the remaining physical examination indicators are incorporated into the continuous prediction model in the form of linear terms. In this embodiment, the continuous prediction model is a generalized additive model, and the use of age and at least one key laboratory test indicator, namely age and albumin, is specifically expressed as follows: MMSE i =β0+f age (Age i )+f alb (Alb i )+Σ(β j ×x ij )+ε i ; Among them, MMSE i This represents the predicted score of the i-th sample, where i is the sample index, β0 represents the intercept term of the continuous prediction model, and f age () and f alb () represent the third-order spline smoothing functions for age and albumin, respectively. i And Alb i Let x represent the age and albumin concentration of the i-th sample, respectively. ij Let β represent the j-th linear feature value of the i-th sample. j ε represents the coefficient corresponding to the j-th linear feature. i Let represent the residual term of the i-th sample; S220. Based on multiple complete datasets, train a continuous prediction model, use cross-validation to evaluate the prediction performance of the continuous prediction model, and record the root mean square error, mean absolute error and coefficient of determination. S230. Calculate the partial derivatives of each feature variable through sensitivity analysis, generate a feature sensitivity matrix, and quantify the change in MMSE prediction score for each input feature when the value increases by one unit. S240. Output a sensitivity analysis report containing smooth function curve parameters, linear term coefficients, and partial derivative matrix. The sensitivity analysis report is used to explain the marginal impact of each input feature on the cognitive function assessment results. Age and albumin have been confirmed by multiple studies to have a non-linear relationship with cognitive function, so they are set as smoothing terms. The remaining indicators are used as linear terms to control the model complexity. For example, a continuous prediction model was trained using five complete datasets, with age and albumin as third-order spline smoothing terms and the remaining 31 indicators as linear terms. Evaluation of the continuous prediction model showed a mean root mean square error of 4.68 ± 0.26, a mean absolute error of 3.43 ± 0.12, and a coefficient of determination of 0.356 ± 0.025. Sensitivity analysis revealed that for every 5-year increase in age, the MMSE predicted value decreased by 0.96 points; for every 1 g / dL increase in albumin, the MMSE predicted value increased by 1.16 points. This is merely an example and is not intended to impose limitations.

[0020] S300: Based on feature variables from multiple complete datasets and continuous MMSE prediction values ​​as input features, a first-stage classification model is constructed; and an optimal threshold combination is determined using an automatic parameter optimization algorithm; based on the class probabilities output by the first-stage classification model and the optimal threshold combination, the samples are divided into high-confidence regions, undecided regions, and low-confidence regions; samples in low-confidence regions will maintain their first-stage classification results and be marked as low-confidence, indicating that further clinical evaluation is needed; Specifically, step S300 includes: S310. Based on age, gender, body mass index, blood pressure parameters, blood routine parameters, liver and kidney function parameters, electrolyte parameters, metabolic parameters and MMSE continuous predicted values ​​from multiple complete datasets as input features, a multi-class logistic regression model is constructed as the first-stage classification model. For example, using 33 physical examination indicators and continuous MMSE prediction values ​​from 5 complete datasets as input features, a multi-class logistic regression model is constructed to output the probability that each sample belongs to normal cognition (MMSE ≥ 27.3 points), suspected mild cognitive impairment (22.1-27.2 points, suspected mild cognitive impairment), or high risk (≤ 22.0 points). S320. Through an automatic parameter optimization algorithm, parameters are optimized within a preset high-confidence threshold search space, a lower bound threshold search space for the region to be discriminated, and an optimal re-discrimination threshold search space to determine the optimal high-confidence threshold, the optimal lower bound threshold for the region to be discriminated, and the optimal re-discrimination threshold. The search space is set based on the observation of the probability distribution in the pre-experiment to ensure coverage of common confidence levels. The joint optimization strategy takes into account both global search and efficiency. In this embodiment, the automatic parameter optimization algorithm is a grid search combined with a Bayesian optimization algorithm. The search space is set based on the observation of the probability distribution in the pre-experiment to ensure coverage of common confidence levels. The joint optimization strategy takes into account both global search and efficiency. Through grid search, the search space for the high confidence threshold is 0.65-0.75, the search space for the lower bound threshold of the region to be discriminated is 0.35-0.45, and the search space for the optimal re-discrimination threshold is 0.50-0.60. Through joint Bayesian optimization, the optimal threshold combination is determined as follows: optimal high confidence threshold = 0.72, optimal lower bound threshold of the region to be discriminated is 0.43, and optimal re-discrimination threshold = 0.56. The search space for each threshold is determined by analyzing the distribution characteristics of classification probabilities in historical data: the high confidence region corresponds to the upper quartile of the probability distribution, the region to be discriminated corresponds to the middle 40% probability interval, and the optimal re-discrimination threshold is determined by maximizing the F1-score on the validation set. This is only an example and is not a limitation. S330. Based on the probability vectors of samples belonging to each category output by the first-stage classification model, extract the maximum category probability value. When the maximum category probability value is greater than or equal to the optimal high-confidence threshold, the corresponding sample is classified into the high-confidence region. When the maximum category probability value is less than the optimal high-confidence threshold and is simultaneously greater than or equal to the optimal lower bound threshold of the undiscriminated region, the corresponding sample is classified into the undiscriminated region. When the maximum category probability value is less than the optimal lower bound threshold of the undiscriminated region, the corresponding sample is classified into the low-confidence region. The three-region division strategy aims to separate high-deterministic samples from samples that need further discrimination, thereby improving the accuracy of identifying boundary cases. For example: Extracting the maximum class probability value of the samples and dividing them into regions: The 2308 samples were divided into three regions: a high-confidence region (392 cases), a region to be determined (1265 cases), and a low-confidence region (651 cases). The three-region division strategy is designed based on the classification error rate analysis on the historical validation set. The error rate of the high-confidence region is <5%, the error rate of the region to be determined is 5%-20%, and the error rate of the low-confidence region is >20%, ensuring effective identification of boundary cases. This is only an example and is not a limitation.

[0021] S400. If the sample size of the region to be discriminated is higher than the preset training threshold, an enhanced feature set is constructed, and a second-stage classification model is built. The second-stage classification model adopts the gradient boosting decision tree algorithm, and a classifier is constructed based on histogram feature discretization and leaf node growth strategy. When the predicted probability of a sample in the region to be discriminated reaches the optimal re-discrimination threshold in the second-stage classification model, its classification result in the first stage is corrected; otherwise, it is maintained. If the sample size of the region to be discriminated is not higher than the preset training threshold, the first-stage classification result of all samples in the region to be discriminated is maintained. Specifically, step S400 includes: S410. Determine whether the number of training samples in the region to be discriminated is higher than the preset training threshold; if not, maintain the first-stage classification results of all samples in the region to be discriminated; for example, the training threshold is set to 500 cases, which is determined based on the learning curve analysis. When the number of samples is lower than this value, the model performance is unstable. S420. If the value is higher, then for the sample in the region to be judged, an enhanced feature set is constructed. The enhanced feature set includes the probability of belonging to each category output by the first-stage classification model, interaction terms or ratio features constructed based on clinical indicators, and boundary correlation features constructed based on MMSE continuous predicted values. For example, in this embodiment, the interaction terms or ratio features constructed based on clinical indicators specifically include the ratio of systolic blood pressure to diastolic blood pressure, and the interaction term of albumin and age. The boundary correlation features constructed based on MMSE continuous predicted values ​​are specifically the absolute value of the difference between the MMSE continuous predicted values ​​and the classification threshold. The interaction terms and boundary distance features are selected from 30 candidate features through feature importance analysis and mutual information calculation. These features contribute the most to the improvement of classification performance on the validation set. S430. Using the enhanced feature set as input features, construct a second-stage classification model based on gradient boosting decision tree, and perform a binary classification task of suspected mild cognitive impairment category and non-suspected mild cognitive impairment category. S440. For each sample in the region to be discriminated, if its predicted probability value in the second-stage classification model is greater than or equal to the optimal re-discrimination threshold, its final classification result is corrected to the suspected mild cognitive impairment category; otherwise, its classification result obtained in the first-stage classification model is maintained. For example: In the 1265 samples to be classified, 885 training samples exceeded the preset threshold, so the enhanced classification process was executed; an enhanced feature set was constructed, which included the first-stage classification probability, blood pressure ratio, albumin-age interaction term, and boundary distance features; a gradient boosting decision tree model was trained, with a maximum tree depth of 6, a learning rate of 0.1, and 100 trees; after reclassification, 428 samples were corrected to suspected mild cognitive impairment, and 837 samples maintained their original classification; this is only an example and is not a limitation.

[0022] S500: Integrate the classification results of high-confidence regions with the classification results of regions to be discriminated to obtain the final classification results for all samples; Specifically, step S500 includes: S510. The classification results of the high-confidence region, the classification results of the low-confidence region, and the classification results of the region to be judged after being judged and corrected by the second-stage classification model are merged with the sample classification results that directly maintain the first-stage classification results to form a complete set of classification results for all samples. S520. Based on the complete set of classification results, output a comprehensive evaluation report. The comprehensive evaluation report includes cognitive function grading results, confidence level, MMSE continuous predictive value, feature sensitivity analysis summary, and classification confidence description.

[0023] For example, by integrating the classification results from various regions, the final classification is obtained: 1256 cases with normal cognition, 683 cases with suspected mild cognitive impairment, and 369 cases with high risk; the output includes a comprehensive evaluation report containing confidence distribution, predicted value statistics, sensitivity analysis, and model performance.

[0024] This invention provides another technical solution: a cognitive function screening system based on the MMSE prediction model. The system includes: a data processing module, a continuous prediction module, a hierarchical decision-making module, an enhanced discrimination module, and an integrated output module. The data processing module collects physical examination indicator data, removes missing MMSE records, imputes and corrects for clinical safety ranges for remaining missing values, and obtains multiple complete datasets. The continuous prediction module constructs a continuous prediction model to obtain continuous MMSE prediction values, performs sensitivity analysis based on the continuous prediction model, and generates a sensitivity analysis report including the marginal impact of each feature on the prediction results. The hierarchical decision module constructs a first-stage classification model, optimizes the algorithm to divide the optimal threshold combination, and classifies samples into high-confidence regions, undiscriminated regions, and low-confidence regions based on class probabilities. The enhanced discrimination module constructs an enhanced feature set and a second-stage classification model when there are sufficient samples in the undiscriminated regions, and corrects the classification results based on the prediction probabilities. The integrated output module integrates the classification results of each region to generate the final classification results and a comprehensive evaluation report.

[0025] Furthermore, the data processing module includes a data acquisition unit, a data cleaning unit, a data imputation unit, and a clinical correction unit. The data acquisition unit is used to collect physical examination indicator data and aggregate duplicate test records based on unique identifiers to form a multi-source dataset. The data cleaning unit is used to perform integrity checks on key variables in the multi-source dataset and remove records with missing key fields. The data imputation unit is used to perform iterative imputation on the remaining missing values ​​using multiple imputation methods to generate multiple sets of initial complete data. The clinical correction unit is used to perform clinical medical safety range correction on the imputed data and calculate the albumin to globulin concentration ratio to obtain multiple sets of complete datasets.

[0026] Furthermore, the hierarchical decision-making module includes a model training unit, a threshold optimization unit, and a region partitioning unit; the model training unit is used to construct a first-stage classification model based on the feature variables and continuous predicted values ​​of the complete dataset; the threshold optimization unit is used to determine the optimal threshold combination through an automatic parameter optimization algorithm; the region partitioning unit is used to divide the samples into regions with different confidence levels based on the class probabilities output by the first-stage classification model and the optimal threshold combination.

[0027] Furthermore, the enhanced discrimination module includes a feature construction unit and a specialized discrimination unit. The feature construction unit is used to construct an enhanced feature set containing the first-stage probability output, clinical indicator interaction terms, and boundary distance features when the sample size of the region to be discriminated exceeds a preset threshold. The specialized discrimination unit is used to construct a second-stage classification model based on the enhanced feature set and to correct the classification results according to the predicted probability.

[0028] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A cognitive function screening method based on the MMSE prediction model, characterized in that: The method includes: S100. Collect physical examination index data, remove missing MMSE records, impute the remaining missing values ​​using multiple imputation methods, and perform clinical safety range correction to obtain multiple complete datasets. S200. Using the multiple complete datasets as input, construct a continuous prediction model to obtain continuous MMSE prediction values; perform sensitivity analysis based on the continuous prediction model to generate a sensitivity analysis report containing the marginal impact of each feature on the prediction results. S300. Based on the feature variables of the multiple complete datasets and the continuous MMSE prediction values, a first-stage classification model is constructed; and the optimal threshold combination is determined through an automatic parameter optimization algorithm; based on the class probabilities output by the first-stage classification model, the samples are divided into high-confidence regions, undiscriminated regions, and low-confidence regions. S400. If the sample size of the region to be discriminated is higher than the preset training threshold, an enhanced feature set is constructed, and a second-stage classification model is constructed. When the predicted probability of a sample in the region to be discriminated in the second-stage classification model reaches the optimal re-discrimination threshold, its classification result in the first stage is corrected; otherwise, it is maintained. If the sample size of the region to be discriminated is not higher than the preset training threshold, the first-stage classification result of all samples in the region to be discriminated is maintained. S500. Integrate the classification results of the high-confidence region and the region to be discriminated to obtain the final classification result of all samples.

2. The cognitive function screening method based on the MMSE prediction model according to claim 1, characterized in that: Step S100 includes: S110. Collect physical examination indicator data, which includes demographic characteristics and laboratory test indicators. The demographic characteristics include age, gender, and body mass index. The laboratory test indicators include blood pressure parameters, blood routine parameters, liver and kidney function parameters, electrolyte parameters, and metabolic parameters. Based on a unique identifier, aggregate duplicate test records under the same identifier and convert all indicators into a unified numerical data format to form a multi-source dataset. S120. Perform key variable integrity and data validity checks on the multi-source dataset, and remove records with missing MMSE, age, and gender fields; S130. For the remaining missing values ​​in the multi-source dataset, iterative interpolation is performed using the multiple interpolation method to construct a feature matrix containing a target interpolation column and an auxiliary interpolation column. The feature matrix is ​​then subjected to feature standardization processing, and multiple sets of initial complete data are generated through the iterative interpolation algorithm. S140. Perform clinical medical safety range correction on the interpolated data to adjust the values ​​of each indicator to the preset physiological reasonable range. S150. Based on the corrected data, construct the multiple complete datasets.

3. The cognitive function screening method based on the MMSE prediction model according to claim 1, characterized in that: Step S200 includes: S210. Establish a continuous prediction model, wherein the continuous prediction model uses age and at least one laboratory test indicator to model through a nonlinear smoothing function, and the remaining physical examination indicators are incorporated into the continuous prediction model in the form of linear terms. S220. Based on the multiple complete datasets, train a continuous prediction model, evaluate the prediction performance of the continuous prediction model using cross-validation, and record the root mean square error, mean absolute error, and coefficient of determination. S230. Calculate the partial derivatives of each feature variable through sensitivity analysis, generate a feature sensitivity matrix, and quantify the change in MMSE prediction score for each input feature when the value increases by one unit. S240. Output a sensitivity analysis report containing smooth function curve parameters, linear term coefficients, and partial derivative matrices. The sensitivity analysis report is used to explain the marginal impact of each input feature on the cognitive function assessment results.

4. The cognitive function screening method based on the MMSE prediction model according to claim 1, characterized in that: Step S300 includes: S310. Based on the age, gender, body mass index, blood pressure parameters, blood routine parameters, liver and kidney function parameters, electrolyte parameters, metabolic parameters and the continuous predicted value of MMSE in the multiple complete datasets as input features, a multi-class logistic regression model is constructed as the first-stage classification model. S320. Through an automatic parameter optimization algorithm, parameter optimization is performed in the preset high confidence threshold search space, the lower bound threshold search space of the region to be judged, and the optimal re-discrimination threshold search space to determine the optimal high confidence threshold, the optimal lower bound threshold of the region to be judged, and the optimal re-discrimination threshold. S330. Based on the probability vectors of the samples belonging to each category output by the classification model in the first stage, extract the maximum category probability value; when the maximum category probability value is greater than or equal to the optimal high confidence threshold, classify the corresponding sample into the high confidence region; when the maximum category probability value is less than the optimal high confidence threshold and is simultaneously greater than or equal to the optimal lower bound threshold of the undetermined region, classify the corresponding sample into the undetermined region; when the maximum category probability value is less than the optimal lower bound threshold of the undetermined region, classify the corresponding sample into the low confidence region.

5. The cognitive function screening method based on the MMSE prediction model according to claim 4, characterized in that: Step S400 includes: S410. Determine whether the number of training samples in the region to be discriminated is higher than a preset training threshold; if not, maintain the first-stage classification result of all samples in the region to be discriminated. S420. If the value is higher, then for the sample in the region to be judged, an enhanced feature set is constructed. The enhanced feature set includes the probability of belonging to each category output by the first-stage classification model, interaction terms or ratio features constructed based on clinical indicators, and boundary correlation features constructed based on MMSE continuous prediction values. S430. Using the enhanced feature set as input features, construct a second-stage classification model based on gradient boosting decision tree, and perform a binary classification task of suspected mild cognitive impairment category and non-suspected mild cognitive impairment category. S440. For each sample in the region to be discriminated, if its predicted probability value in the second-stage classification model is greater than or equal to the optimal re-discrimination threshold, its final classification result is corrected to the suspected mild cognitive impairment category; otherwise, its classification result obtained in the first-stage classification model is maintained.

6. The cognitive function screening method based on the MMSE prediction model according to claim 1, characterized in that: Step S500 includes: S510. The classification results of the high-confidence region, the classification results of the low-confidence region, and the classification results of the region to be judged after being judged and corrected by the second-stage classification model are merged with the sample classification results that directly maintain the first-stage classification results to form a complete set of classification results for all samples. S520. Based on the complete set of classification results, output a comprehensive evaluation report; the comprehensive evaluation report includes cognitive function grading results, confidence level, MMSE continuous predicted value, feature sensitivity analysis summary and classification confidence description.

7. A cognitive function screening system based on the MMSE prediction model, characterized in that: The system includes: a data processing module, a continuous prediction module, a hierarchical decision-making module, an enhanced discrimination module, and an integrated output module; The data processing module is used to collect physical examination index data, remove missing MMSE records, imputate the remaining missing values ​​and correct for clinical safety range, and obtain multiple complete datasets. The continuous prediction module is used to construct a continuous prediction model to obtain continuous MMSE prediction values, perform sensitivity analysis based on the continuous prediction model, and generate a sensitivity analysis report containing the marginal impact of each feature on the prediction results. The hierarchical decision module is used to construct a first-stage classification model, optimize the algorithm to divide the optimal threshold combination, and divide the samples into high-confidence regions, undiscriminated regions, and low-confidence regions based on the class probability. The enhanced discrimination module is used to construct an enhanced feature set and a second-stage classification model when there are sufficient samples in the region to be discriminated, and to correct the classification result according to the predicted probability. The integrated output module is used to integrate the classification results of each region and generate the final classification result and comprehensive evaluation report.

8. The cognitive function screening system based on the MMSE prediction model according to claim 7, characterized in that: The data processing module includes a data acquisition unit, a data cleaning unit, a data interpolation unit, and a clinical correction unit; The data acquisition unit is used to collect physical examination indicator data, aggregate duplicate detection records based on unique identifiers, and form a multi-source dataset; the data cleaning unit is used to perform a key variable integrity check on the multi-source dataset and remove records with missing key fields. The data imputation unit is used to iteratively imput the remaining missing values ​​using a multiple imputation method to generate multiple sets of initial complete data; the clinical correction unit is used to perform clinical medical safety range correction on the imputed data and calculate the albumin to globulin concentration ratio to obtain the multiple sets of complete datasets.

9. The cognitive function screening system based on the MMSE prediction model according to claim 7, characterized in that: The hierarchical decision-making module includes a model training unit, a threshold optimization unit, and a region partitioning unit. The model training unit is used to construct a first-stage classification model based on the feature variables and continuous predicted values ​​of the complete dataset. The threshold optimization unit is used to determine the optimal threshold combination through an automatic parameter optimization algorithm; the region division unit is used to divide the samples into regions with different confidence levels based on the class probabilities output by the first-stage classification model and the optimal threshold combination.

10. The cognitive function screening system based on the MMSE prediction model according to claim 7, characterized in that: The enhanced discrimination module includes a feature construction unit and a specialized discrimination unit; The feature construction unit is used to construct an enhanced feature set containing the first-stage probability output, clinical indicator interaction terms, and boundary distance features when the sample size of the region to be judged is higher than a preset threshold; the specialized discrimination unit is used to construct a second-stage classification model based on the enhanced feature set and correct the classification results according to the predicted probability.

Images