Construction method, dynamic prediction method and system of cognitive impairment dynamic prediction model

Abstract

The application discloses a kind of cognitive impairment dynamic prediction model construction method, dynamic prediction method and system.The construction method is by obtaining the longitudinal multidimensional clinical evaluation data of historical patient, constructs including clinical staging trajectory, decline rate trajectory and drive sub-dimension development trajectory in multidimensional representation;Based on multi-task neural network, with baseline data as input, with multidimensional representation as supervision target for joint training, form cognitive impairment dynamic prediction model.The model can receive current clinical data of patient, output future multiple time points' staging probability, decline rate and key drive sub-dimension prediction result, and generate personalized clinical report.The present application overcomes the limitations of traditional static prediction method, realizes the continuous dynamic capture to the development process of cognitive impairment, with the advantages of high prediction accuracy, low cost, easy popularization, result can be explained, suitable for early warning and individualized intervention.

Description

Technical Field

[0001] This invention relates to a method for constructing a dynamic prediction model for cognitive impairment, a dynamic prediction method using the dynamic prediction model for cognitive impairment, and a corresponding dynamic prediction system, belonging to the field of cognitive assessment technology. Background Technology

[0002] With the aging population, Alzheimer's disease and related dementias have become a serious public health challenge. These diseases not only cause progressive loss of cognitive function and inability to care for themselves, but also impose a heavy economic and care burden on families and society. Because the disease has a long prodromal period, accurate prediction and intervention before or at an early stage of cognitive impairment are key to slowing disease progression and reducing the social burden.

[0003] Existing technologies mostly rely on cross-sectional data to build statistical models to predict the probability and risk of developing the disease in the future. This approach typically uses logistic regression or Cox proportional hazards models, taking the patient's demographic information (such as age, sex, and education level) and clinical scores (such as MoCA score and MMSE score) at a single time point (baseline) as independent variables. The dependent variable of the model is a binary categorical outcome, usually defined as whether the disease will progress to mild cognitive impairment or Alzheimer's disease in the future (e.g., within 5 years).

[0004] For relatively complex model predictions, existing technologies primarily employ core biomarker systems for accurate prediction. Among these, cerebrospinal fluid (CSF) testing, obtained via lumbar puncture, directly measures the concentrations of Aβ42, total Tau protein, and phosphorylated Tau protein; abnormal patterns are considered one of the "gold standards" for AD diagnosis. For imaging, positron emission tomography (PET) is used to visualize the distribution and density of Aβ deposits or Tau protein neurofibrillary tangles across the entire brain in vivo using specific radioactive tracers. Structural magnetic resonance imaging (SMRI) is used to quantify the degree of atrophy in AD-related brain regions, such as measuring the volume of the hippocampus and entorhinal cortex, serving as an indirect marker of neuronal degeneration.

[0005] However, existing technologies are mostly based on a "static" approach at a single point in time, rather than capturing the crucial dynamic information of changes in cognitive function. Cognitive decline is a continuous evolutionary process, and its rate of decline and inflection points of acceleration are more predictive than absolute values ​​at a single point. Such models struggle to utilize this information, often resulting in insufficient predictive sensitivity and early warning capabilities.

[0006] For predictive biomarkers, cerebrospinal fluid testing is invasive, and amyloid PET imaging is extremely expensive; both require large medical centers and are performed by professionals. This limits the application of these predictive methods to community screening, primary healthcare, and large-scale population deployment, confining them to a few specialized hospitals and research settings. Summary of the Invention

[0007] The primary technical problem to be solved by this invention is to provide a method for constructing a dynamic prediction model for cognitive impairment.

[0008] Another technical problem to be solved by the present invention is to provide a dynamic prediction method for the development of cognitive impairment.

[0009] Another technical problem to be solved by the present invention is to provide a dynamic prediction system for the development of cognitive impairment.

[0010] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: According to a first aspect of the present invention, a method for constructing a dynamic prediction model for cognitive impairment is provided, comprising the following steps: S1: Obtain longitudinal multidimensional clinical assessment data of historical patients at multiple time points; wherein, the multiple time points include the baseline time point and at least one future follow-up time point, and the longitudinal multidimensional clinical assessment data includes demographic information, overall cognitive assessment data and cognitive assessment data of each sub-dimension; S2: Based on the aforementioned longitudinal multidimensional clinical assessment data, a multidimensional representation reflecting the dynamic development of cognitive impairment is constructed. The multidimensional representation includes at least: a clinical staging trajectory, obtained by mapping cognitive assessment data at each time point to a preset clinical staging judgment result of cognitive impairment; a decline rate trajectory, obtained by performing mixed-effects modeling on the changes in cognitive assessment scores of each sub-dimensional over time and extracting a patient-specific decline rate function; and a driving sub-dimensional development trajectory, obtained by analyzing the dynamic contribution of changes in each sub-dimensional to global cognitive decline. S3: Construct a multi-task neural network model, which includes a shared feature extraction layer and multiple parallel task-specific output layers; S4: Using the clinical assessment data of the baseline time point of the historical patients as the input of the feature extraction layer, and using each trajectory in the multi-dimensional representation of each patient as the supervision signal of multiple task-specific output layers, the multi-task neural network model is jointly trained to form an initial model. S5: The initial model is optimized by minimizing a joint loss function to form the final dynamic prediction model for cognitive impairment; wherein the joint loss function is a weighted sum of the classification task loss and the regression task loss.

[0011] Preferably, in step S3, the plurality of task-specific output layers specifically include: The classification output header uses either the sigmoid activation function or the softmax activation function to output the predicted results of clinical staging. The first regression output head uses a linear activation function to output the predicted decay rate of each sub-dimension; The second regression output head uses a linear activation function to predict the dynamic contribution of changes in each sub-dimension to global cognitive decline.

[0012] Preferably, the cognitive impairment dynamic prediction model is used to receive multi-dimensional clinical assessment data at a point in time and output multi-dimensional representation prediction results for multiple future time points from that point in time. The multidimensional characterization prediction results are composed of the clinical staging prediction results, the decline rate prediction results of each sub-dimension, and the contribution prediction results of each sub-dimension.

[0013] Preferably, in step S5, the classification task loss is cross-entropy loss, and the regression task loss is mean squared error loss or mean absolute error loss.

[0014] Preferably, in step S2, the mixed-effects model is a linear mixed-effects model that includes natural cubic spline basis functions to simulate nonlinear time effects, and the decay rate trajectory is the first derivative of the conditional mean with respect to time estimated by the model.

[0015] Preferably, in step S2, the development trajectory of the driving sub-dimension is obtained in the following way: Within a continuous fixed time window, calculate the annual rate of change of global cognition and each sub-dimension; Regression analysis was performed using the annual rate of change of global perception as the dependent variable and the annual rate of change of each sub-dimension as the independent variable to obtain the contribution coefficient of each sub-dimension. The contribution coefficients of each time window are arranged in chronological order to form the development trajectory of the driving sub-dimension.

[0016] According to a second aspect of the present invention, a method for dynamically predicting the development of cognitive impairment is provided, comprising the following steps: Obtain multi-dimensional clinical assessment data of the patient to be predicted at the current time point; The multidimensional clinical assessment data at the current time point is input into the pre-trained dynamic prediction model of cognitive impairment; wherein, the dynamic prediction model of cognitive impairment is constructed by the above method; Output the multi-dimensional dynamic prediction results predicted by the cognitive impairment dynamic prediction model; The multi-dimensional dynamic prediction results include at least: the probability of clinical staging of cognitive impairment at multiple future time points; the estimated rate of cognitive decline at multiple future time periods; and the key sub-dimension identifiers and their contributions driving cognitive decline at multiple future time periods.

[0017] Preferably, the dynamic prediction method further includes: Based on model attribution analysis, the risk factors and protective factors that contribute the most to the multi-dimensional dynamic prediction results are identified. For the aforementioned risk factors and protective factors, reversible and fixed factors are identified and marked; Generate a personalized report that includes the multi-dimensional dynamic prediction results, a list of key factors, and intervention priorities; wherein the list of key factors consists of tagged risk factors and protective factors.

[0018] According to a third aspect of the present invention, a dynamic prediction system for the development of cognitive impairment is provided, comprising: The data acquisition module is used to acquire multi-dimensional clinical assessment data representing historical patients at multiple time points; The characterization construction module is used to construct a multi-dimensional characterization of cognitive impairment development based on the longitudinal multi-dimensional clinical assessment data. The multi-dimensional characterization includes clinical staging trajectory, decline rate trajectory, and driving sub-dimensional development trajectory. The model training module is used to train a multi-task neural network model with the clinical assessment data of the patient's baseline time point as input and the multi-dimensional representation as the supervision target, so as to construct the above-mentioned dynamic prediction model of cognitive impairment. The prediction output module is used to input the current multi-dimensional clinical assessment data of the new patient into the cognitive impairment dynamic prediction model, thereby outputting the prediction results of the patient's future cognitive impairment development.

[0019] Preferably, the dynamic prediction system further includes: The explanation and analysis module is used to identify key risk factors and protective factors based on model attribution analysis and to label them as reversible factors or fixed factors. The report generation module is used to generate a personalized clinical report containing the cognitive impairment development prediction results, a list of key factors, and intervention recommendations; wherein the list of key factors consists of labeled risk factors and protective factors.

[0020] Compared with the prior art, the present invention has the following technical effects: (1) The embodiments of the present invention, by constructing multi-dimensional representations (clinical staging trajectory, decline rate trajectory, and driving sub-dimensional trajectory), break through the limitations of traditional single-point static prediction, can capture the continuous dynamic process of cognitive impairment development, significantly improve early warning capabilities, and have strong dynamic prediction capabilities.

[0021] (2) The embodiments of the present invention rely only on routine clinical assessment data (such as MoCA, MMSE, etc.), without the need for invasive cerebrospinal fluid testing or expensive imaging examinations, and are low in cost; moreover, they are suitable for community screening, primary healthcare and large-scale population promotion and application.

[0022] (3) The embodiments of the present invention employ a multi-task neural network to simultaneously predict stage, rate, and driving dimension, sharing feature representations to improve the model's generalization ability and prediction consistency, and avoid overfitting of a single task. Furthermore, the output results are rich and interpretable, not only outputting the probability of future stage, but also providing the decay rate, key driving sub-dimensions and their contributions, and supporting attribution analysis such as SHAP, providing a quantitative basis for individualized intervention. Attached Figure Description

[0023] Figure 1 A flowchart illustrating a method for constructing a dynamic prediction model for cognitive impairment, as provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the construction principle of the dynamic prediction model for cognitive impairment in the first embodiment of the present invention. Figure 3 A flowchart illustrating a dynamic prediction method for the development of cognitive impairment provided in the second embodiment of the present invention; Figure 4 This is a schematic diagram of the output results of the dynamic prediction model for cognitive impairment in the second embodiment of the present invention; Figure 5 This is a structural diagram of a dynamic prediction system for the development of cognitive impairment provided in the third embodiment of the present invention. Detailed Implementation

[0024] The technical content of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0025] The core technical concept of this invention is as follows: First, a multi-dimensional representation system that can comprehensively depict the dynamic nature of cognitive impairment development is actively constructed from longitudinal clinical data including baseline and multiple follow-ups; then, a specialized multi-task neural network is designed to jointly learn the mapping relationship of future multi-dimensional development trajectory, starting from baseline data; finally, the trained model is applied to individuals to achieve synchronous, multi-dimensional and prospective prediction of their cognitive decline process.

[0026] Specifically, this invention first defines and constructs three types of dynamic development goals with clear clinical significance: clinical staging trajectories reflecting the macroscopic stage of the disease, decline rate trajectories quantifying the rate of functional change, and driving sub-dimensional contribution trajectories revealing the underlying driving mechanisms. It is understood that these goals collectively constitute a more refined and richer supervisory signal system than a single diagnostic label or risk score. Subsequently, this invention employs a multi-task neural network sharing underlying feature representations, using an individual's clinical assessment data at baseline as input, and jointly training the model with the aforementioned three types of dynamic trajectories as parallel supervisory goals. This training method forces the model to learn deep feature representations that can simultaneously explain disease staging, rate of change, and underlying driving factors, thereby significantly improving the model's generalization ability and prediction accuracy. Finally, the trained model can receive current clinical data from a new individual and output multi-dimensional prediction results, including staging probability, decline rate estimation, and key driving sub-dimensions, at multiple future time points, providing unprecedented quantitative tools and decision support for early clinical warning, personalized intervention, and disease progression management.

[0027] It should be noted that, in this embodiment of the invention, "baseline" refers to the time point at which the subject first undergoes a full set of multidimensional clinical assessments, denoted as the starting point of the study. All information collected at this time point, including demographic information, cognitive and emotional assessment scores, personality and behavioral scale scores, etc., are collectively referred to as "baseline data" or "baseline characteristics." All longitudinal time information is calculated with this baseline time point as the origin. For example, "follow-up for 1 year" refers to the assessment at year 1 from the baseline time point; "rate of decline over the next 3 years" refers to the average rate of change over the next 3 years from the baseline time point.

[0028] During the model training phase, the baseline time point data of the subjects are used as input features, and the multi-dimensional trajectory calculated based on the subject's baseline and subsequent follow-up data is used as the supervision target, thereby establishing a predictive mapping of "current state → future process".

[0029] First Embodiment like Figure 1 As shown, the first embodiment of the present invention provides a method for constructing a dynamic prediction model for cognitive impairment, specifically including the following steps: S1: Obtain longitudinal, multi-dimensional clinical assessment data of historical patients at multiple time points.

[0030] In this embodiment of the invention, multiple time points include a baseline time point and at least one future follow-up time point (e.g., follow-up for 1 to 5 years from the baseline time point). Multidimensional clinical assessment data includes demographic information (e.g., age, gender, years of education, center / batch), cognitive and emotional assessment scores (e.g., MoCA, MMSE, CDR, Hamilton Anxiety Rating Scale, and Hamilton Depression Rating Scale), clinically commonly used personality / behavioral scales (Neuropsychiatric scale NPI), overall cognitive assessment data, and cognitive assessment data for each sub-dimension.

[0031] S2: Based on longitudinal multidimensional clinical assessment data, construct a multidimensional representation to reflect the dynamic development of cognitive impairment.

[0032] Specifically, it includes the following steps: S21: Obtain clinical staging trajectory.

[0033] In this embodiment of the invention, based on preset thresholds of existing cognitive assessment scales, subjects are classified as cognitively normal, mild cognitive impairment (MCI), or dementia at each follow-up visit, primarily referring to standardized tools such as CDR, MoCA, and MMSE. Thus, by mapping cognitive assessment data at each time point to preset clinical staging results of cognitive impairment, the clinical staging trajectory of the subjects is obtained.

[0034] Furthermore, in this embodiment of the invention, in addition to clinical staging, the degree of impairment at each time point during a 1-5 year follow-up period is quantified based on corresponding test scores across cognitive sub-dimensions such as memory, executive function, language, attention / processing speed, and visuospatial function, and individualized longitudinal change trajectories are constructed. Moreover, to ensure the consistency and comparability of different scales, the score direction is standardized before trajectory modeling, ensuring that both high and low scores signify the same thing: lower scores indicate poorer function.

[0035] S22: Obtain the decay rate trajectory.

[0036] In this embodiment of the invention, the decline rate trajectory is obtained by modeling the changes in cognitive scores of each sub-dimension over time using mixed-effects modeling, and extracting a personalized decline rate function for each patient.

[0037] Specifically, for both global and sub-domain cognitive levels, a linear mixed-effects model (LMM) with spline time terms was used to fit all longitudinal observation data from 1 to 5 years. Using each cognitive test score (including global cognition and sub-dimension tests) as the target variable Y, a linear mixed-effects model with random intercepts and random slopes was fitted for each target Y: ; in, The number of years from the baseline; For time smoothing function, For natural cubic spline basis functions, For the coefficients that need to be estimated, The number of basis functions; To remove confounding covariates (including baseline age, gender, years of education, center / batch, assessment wave or equipment version, etc.); and These represent the individual random intercept and random slope, respectively; This represents the residual term. The variance components are estimated using REML. Based on this model, individual... The time-varying decay rate trajectory is the first derivative of the conditional mean with respect to time: ; In this model, after directional unification, negative values ​​represent cognitive decline. The final output shows the individual's decline rate trajectory and its 95% confidence band across the global and sub-dimension dimensions.

[0038] S23: Obtain the development trajectory of the driving sub-dimension.

[0039] In this embodiment of the invention, the development trajectory of the driving sub-dimension is obtained by analyzing the dynamic contribution of changes in each sub-dimension to global cognitive decline.

[0040] Specifically, to identify key driving sub-dimensions of cognitive impairment over time, the dynamic contribution of each cognitive sub-dimension to global cognitive decline was estimated within consecutive fixed one-year windows (0-1, 1-2, 2-3, 3-4, 4-5 years). This was done for the subjects... Calculate the annual rate of change of global cognition and each sub-dimension within window [a, b], defined as: ; in, and , and These represent the scores and time points of the two tests closest to the start and end of the window, respectively, where negative values ​​indicate a decline in cognitive function. Within each time window, a linear regression is fitted using the annualized change in global cognition as the dependent variable and the annualized changes in each cognitive sub-dimension as the independent variable: ; Among them, when Smaller negative values ​​indicate a stronger driving force of that sub-dimension on global cognitive decline within the corresponding time period. The estimated values ​​from each time window... Arranged chronologically, we obtain the "sub-dimensional driving trajectory," thereby revealing the key driving sub-dimensions of cognitive impairment at different stages.

[0041] S3: Construct a multi-task neural network model.

[0042] In this embodiment of the invention, the multi-task neural network model includes a shared feature extraction layer and multiple parallel task-specific output layers. The feature extraction layer takes clinical assessment data from the baseline time point of historical patients as input, and the multiple task-specific output layers sequentially include: The classification output head (corresponding to the classification task) uses the sigmoid activation function or the softmax activation function to output the prediction results of clinical staging; The first regression output head (corresponding to the regression task) uses a linear activation function to output the predicted results of the decay rate of each sub-dimension; The second regression output head (corresponding to the regression task) uses a linear activation function to output the predicted results of the dynamic contribution of changes in each sub-dimension to global cognitive decline.

[0043] S4: Jointly train the multi-task neural network model to form the initial model.

[0044] In this embodiment of the invention, clinical assessment data at the baseline time of historical patients are used as input to the feature extraction layer, and each trajectory in the multi-dimensional representation of each patient is used as the supervision signal for multiple task-specific output layers, thereby jointly training the multi-task neural network model to form an initial model.

[0045] Specifically, the input side includes basic demographic information of the patient, baseline and early follow-up cognitive and emotional assessment scores, and clinically used personality and lifestyle assessment scores. The output side simultaneously supervises three dependent variables related to the developmental trajectory of cognitive impairment: (i) The time series of clinical staging and the continuous trajectory of cognitive sub-dimension decline; (ii) Global and sub-dimension decay rate trajectories; (iii) Annual sub-dimension driving coefficients.

[0046] Therefore, within the same framework, the system simultaneously outputs clinical diagnostic results and cognitive test scores at future time points, as well as individualized cognitive decline rates and key driving sub-dimensions. The dataset is divided into 90% training and 10% independent testing to evaluate sensitivity, specificity, and accuracy, and different models are compared within a unified evaluation framework.

[0047] In this embodiment of the invention, the specific training process of the multi-task neural network model is as follows: (1) A stratified 10-fold cross-validation was used to conduct 10 training and validation cycles. Patients used for training were randomly divided into 10 folds, and 9 folds were used for training in sequence, with the remaining 1 fold used for validation and evaluation.

[0048] (2) The basic information at the baseline time point, the total score and sub-scores of each scale are used as the input layer; the specific task objectives are used as the output layer, including: cognitive impairment stage, rate of decline, and sub-dimensional contribution trajectory. ReLU or tanh activation is used in the hidden layer, and activation is differentiated according to the task in the output layer: the sigmoid function is used for the classification head, and the linear activation function is used for continuous output. The activation function is expressed as follows: in, For the output of the neural node, Represents the weighted sum of all input nodes. Refers to a specific neural node.

[0049] S5: Parameter optimization to form the final dynamic prediction model for cognitive impairment.

[0050] In this embodiment of the invention, the parameters of the initial model are optimized by minimizing a joint loss function to form the final dynamic prediction model for cognitive impairment. The joint loss function is a weighted sum of the classification task loss and the regression task loss. The classification task loss is the cross-entropy loss, and the regression task loss is the mean squared error loss or the mean absolute error loss.

[0051] Specifically, in this embodiment of the invention, mini-batch gradient descent and its adaptive variant (Adam) are used for fitting, and corresponding losses are defined for different tasks: cross-entropy is used for classification, and MSE / MAE is used for continuous trajectories. The update formulas for the model parameters in each iteration are as follows: in, The weight parameter representing the input node i and output node j of data point n. The learning rate represents the network's learning rate, thus ensuring that the parameters respond quickly to changes in the results. This represents batch or sample-level loss.

[0052] Plot the ROC (Receiver Operating Characteristic curve) on the retained validation fold and the final independent test set, and calculate the AUC (Area Under Curve) to estimate the generalization performance, as shown in the following formula: in, If The value is 1 if it is 1, otherwise it is 0. Represents the negative sample set; This represents the positive sample set.

[0053] The optimal model is selected based on prediction accuracy, sensitivity, specificity, AIC / BIC, and comparison results of Bayesian models. This is the final trained dynamic prediction model for cognitive impairment. (Refer to...) Figure 2 As shown, the cognitive impairment dynamic prediction model finally trained in this embodiment of the invention is used to receive multi-dimensional clinical assessment data at a point in time and output multi-dimensional representation prediction results for multiple future time points (e.g., the next 1 year, the next 3 years, and the next 5 years) from that point in time. The multi-dimensional representation prediction results are composed of clinical staging prediction results, prediction results of the decline rate of each sub-dimension, and prediction results of the contribution of each sub-dimension.

[0054] Second Embodiment like Figure 3 As shown, based on the first embodiment described above, the second embodiment of the present invention also provides a dynamic prediction method for the development of cognitive impairment, specifically including the following steps: S10: Obtain multi-dimensional clinical assessment data of the patient to be predicted at the current time point.

[0055] S20: Input the multi-dimensional clinical assessment data at the current time point into the pre-trained dynamic prediction model for cognitive impairment.

[0056] The cognitive impairment dynamic prediction model is constructed using the method provided in the first embodiment described above.

[0057] S30: Output the multi-dimensional dynamic prediction results of the cognitive impairment dynamic prediction model.

[0058] Specifically, refer to Figure 4 As shown in this embodiment of the invention, the multi-dimensional dynamic prediction result includes at least the following: (1) Probability of clinical staging of cognitive impairment at multiple future time points; for example: probability of clinical staging of cognitive impairment in the next 1 year, 3 years and 5 years.

[0059] (2) Estimated rates of cognitive decline over multiple time periods in the future; for example: estimated rates of cognitive decline over the next 1 year, 3 years and 5 years.

[0060] (3) Key sub-dimensional identifiers and their contributions that drive cognitive decline in multiple time periods in the future; for example: key sub-dimensional identifiers and their contributions in the next 1 year, 3 years and 5 years.

[0061] S40: Generate personalized reports.

[0062] In this embodiment of the invention, based on model attribution analysis, risk factors and protective factors that contribute most to the multi-dimensional dynamic prediction results are identified. Specifically, for the multi-dimensional dynamic prediction results output by the cognitive impairment dynamic prediction model, the contribution-based interpretable method (SHAP) is used to analyze the key features: those with a positive average contribution are defined as risk factors, and those with a negative average contribution are defined as protective factors.

[0063] Then, for risk factors and protective factors, reversible and fixed factors are identified and labeled. Specifically, for the risk factor and protective factor, features with large fluctuations over time are labeled as reversible (changeable) factors, and those with smaller fluctuations are labeled as fixed factors, thereby supporting the prioritization of interventions.

[0064] Finally, a personalized report is generated that includes multi-dimensional dynamic prediction results, a list of key factors, and intervention priorities; the list of key factors consists of tagged risk factors and protective factors.

[0065] Third Embodiment like Figure 5 As shown, based on the first embodiment described above, the third embodiment of the present invention provides a dynamic prediction system for the development of cognitive impairment, comprising at least: a data acquisition module 1, a representation construction module 2, a model training module 3, a prediction output module 4, an interpretation and analysis module 5, and a report generation module 6. Wherein, Data acquisition module 1 serves as the system entry point, responsible for collecting and storing raw assessment data from historical and new patients. Data acquisition includes baseline and follow-up multivariate information (demographics, cognitive scores, emotion scales, etc.), thus providing a data foundation for downstream representation construction module 2 and predictive output module 4.

[0066] The representation construction module 2 is the core technology preprocessing unit. This module receives historical patient longitudinal data from the data acquisition module 1, performs medical statistical modeling (such as LMM and window regression), and calculates three core multi-dimensional representations: clinical staging trajectory, decline rate trajectory, and driving sub-dimensional development trajectory. Therefore, the output of this representation construction module 2 (multi-dimensional representations) is directly used as the supervised target for the model training module. This module transforms the raw data into dynamic labels with clear clinical significance that can be learned by the machine learning model.

[0067] Model training module 3 is the model manufacturing center. It trains a multi-task neural network model by receiving historical patient baseline data from data acquisition module 1 (as input features) and multi-dimensional representations from representation construction module 2 (as supervised targets). This training forms a dynamic predictive model of cognitive impairment, which serves as the core asset of the entire system and is then delivered to prediction output module 4.

[0068] The prediction output module 4 serves as the model application interface. By loading the cognitive impairment dynamic prediction model generated by the model training module 3 and receiving new patient current data from the data acquisition module 1, it performs forward inference and outputs multi-dimensional dynamic prediction results (such as future stage, rate of decline, and key driver sub-dimensions). Thus, the offline-trained model is applied to online prediction, and its structured prediction output is the direct input to the interpretation and analysis module 5 and the report generation module 6 for downstream processing.

[0069] The interpretation and analysis module 5 is used to analyze the prediction results. By receiving the prediction results from the prediction output module 4, it calls upon the attribution information (such as SHAP values) within the model for in-depth analysis, identifying the risk factors and protective factors that contribute the most, and classifying them into interventionable factors and fixed factors. This provides interpretability to the prediction results, transforming the model's "black box" output into human-understandable and actionable insights. Its analytical results are a key input for generating personalized reports.

[0070] The report generation module 6 serves as the delivery terminal for the system's output. By integrating the prediction results from the prediction output module 4 and the factor analysis results from the interpretation and analysis module 5, it automatically generates a structured, personalized clinical report that includes quantitative predictions, risk interpretations, and intervention recommendations, following a preset template. This achieves the final realization of the technology's value, creating a decision support document that can be directly used by doctors or patients.

[0071] It is understood that the functions and synergistic relationships of the above-mentioned modules are only a specific application example of the cognitive impairment dynamic prediction model constructed based on the first embodiment. In other embodiments, the functions and synergistic relationships of the modules can be adaptively adjusted as needed, and no specific limitations are made here.

[0072] It should be noted that the above embodiments are merely illustrative examples. The technical solutions of each embodiment can be combined, and all are within the protection scope of this invention.

[0073] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0074] The foregoing has provided a detailed description of the construction method, dynamic prediction method, and system for the cognitive impairment dynamic prediction model provided by this invention. Any obvious modifications made by those skilled in the art without departing from the essential content of this invention will constitute an infringement of the patent rights of this invention and will incur corresponding legal liability.

Claims

1. A method for constructing a dynamic prediction model for cognitive impairment, characterized in that... Includes the following steps: S1: Obtain longitudinal multidimensional clinical assessment data of historical patients at multiple time points; wherein, the multiple time points include the baseline time point and at least one future follow-up time point, and the longitudinal multidimensional clinical assessment data includes demographic information, overall cognitive assessment data and cognitive assessment data of each sub-dimension; S2: Based on the aforementioned longitudinal multidimensional clinical assessment data, a multidimensional representation reflecting the dynamic development of cognitive impairment is constructed. The multidimensional representation includes at least: a clinical staging trajectory, obtained by mapping cognitive assessment data at each time point to a preset clinical staging judgment result of cognitive impairment; a decline rate trajectory, obtained by performing mixed-effects modeling on the changes in cognitive assessment scores of each sub-dimensional over time and extracting a patient-specific decline rate function; and a driving sub-dimensional development trajectory, obtained by analyzing the dynamic contribution of changes in each sub-dimensional to global cognitive decline. S3: Construct a multi-task neural network model, which includes a shared feature extraction layer and multiple parallel task-specific output layers; S4: Using the clinical assessment data of the baseline time point of the historical patients as the input of the feature extraction layer, and using each trajectory in the multi-dimensional representation of each patient as the supervision signal of multiple task-specific output layers, the multi-task neural network model is jointly trained to form an initial model. S5: The initial model is optimized by minimizing a joint loss function to form the final dynamic prediction model for cognitive impairment; wherein the joint loss function is a weighted sum of the classification task loss and the regression task loss.

2. The construction method as described in claim 1, characterized in that... In step S3, the plurality of task-specific output layers include: The classification output header uses either the sigmoid activation function or the softmax activation function to output the predicted results of clinical staging. The first regression output head uses a linear activation function to output the predicted decay rate of each sub-dimension; The second regression output head uses a linear activation function to predict the dynamic contribution of changes in each sub-dimension to global cognitive decline.

3. The construction method as described in claim 2, characterized in that: The cognitive impairment dynamic prediction model is used to receive multi-dimensional clinical assessment data at a point in time and output multi-dimensional representation prediction results for multiple future time points starting from that point in time. The multidimensional characterization prediction results are composed of the clinical staging prediction results, the decline rate prediction results of each sub-dimension, and the contribution prediction results of each sub-dimension.

4. The construction method as described in claim 1, characterized in that... In step S5, the classification task loss is cross-entropy loss, and the regression task loss is mean squared error loss or mean absolute error loss.

5. The construction method as described in claim 1, characterized in that... In step S2, the mixed-effects model is a linear mixed-effects model that includes natural cubic spline basis functions to simulate nonlinear time effects, and the decay rate trajectory is the first derivative of the conditional mean with respect to time estimated by the model.

6. The construction method as described in claim 1, characterized in that... In step S2, the development trajectory of the driving sub-dimension is obtained in the following way: Within a continuous fixed time window, calculate the annual rate of change of global cognition and each sub-dimension; Regression analysis was performed using the annual rate of change of global perception as the dependent variable and the annual rate of change of each sub-dimension as the independent variable to obtain the contribution coefficient of each sub-dimension. The contribution coefficients of each time window are arranged in chronological order to form the development trajectory of the driving sub-dimension.

7. A dynamic prediction method for the development of cognitive impairment, characterized in that... Includes the following steps: Obtain multi-dimensional clinical assessment data of the patient to be predicted at the current time point; The multidimensional clinical assessment data at the current time point is input into the pre-trained dynamic prediction model of cognitive impairment; wherein the dynamic prediction model of cognitive impairment is constructed by the method described in any one of claims 1 to 6; Output the multi-dimensional dynamic prediction results predicted by the cognitive impairment dynamic prediction model; The multi-dimensional dynamic prediction results include at least: the probability of clinical staging of cognitive impairment at multiple future time points; the estimated rate of cognitive decline at multiple future time periods; and the key sub-dimension identifiers and their contributions driving cognitive decline at multiple future time periods.

8. The dynamic prediction method as described in claim 7, characterized in that... Also includes: Based on model attribution analysis, the risk factors and protective factors that contribute the most to the multi-dimensional dynamic prediction results are identified. For the aforementioned risk factors and protective factors, reversible and fixed factors are identified and marked; Generate a personalized report that includes the multi-dimensional dynamic prediction results, a list of key factors, and intervention priorities; wherein the list of key factors consists of tagged risk factors and protective factors.

9. A dynamic prediction system for the development of cognitive impairment, characterized in that... include: The data acquisition module is used to acquire multi-dimensional clinical assessment data representing historical patients at multiple time points; The characterization construction module is used to construct a multi-dimensional characterization of cognitive impairment development based on the longitudinal multi-dimensional clinical assessment data. The multi-dimensional characterization includes clinical staging trajectory, decline rate trajectory, and driving sub-dimensional development trajectory. The model training module is used to train a multi-task neural network model with the clinical assessment data of the patient's baseline time point as input and the multi-dimensional representation as the supervision target, so as to construct the cognitive impairment dynamic prediction model as described in any one of claims 1 to 6. The prediction output module is used to input the current multi-dimensional clinical assessment data of the new patient into the cognitive impairment dynamic prediction model, thereby outputting the prediction results of the patient's future cognitive impairment development.

10. The dynamic prediction system as described in claim 9, characterized in that... Also includes: The explanation and analysis module is used to identify key risk factors and protective factors based on model attribution analysis and to label them as reversible factors or fixed factors. The report generation module is used to generate a personalized clinical report that includes the predicted results of the cognitive impairment development, a list of key factors, and intervention recommendations; wherein the list of key factors consists of labeled risk factors and protective factors.

Images