Alzheimer's disease dynamic prediction method based on multi-modal causal inference

Abstract

The invention relates to the technical field of dynamic prediction of Alzheimer's disease. The invention relates to an Alzheimer's disease dynamic prediction method based on multi-modal causal inference. The method comprises the following steps: S1, collecting multi-source related data of a user, and performing unified quantization and feature extraction on the multi-source related data to form a factor feature set; s2, establishing a causal network of the Alzheimer's disease; according to the method, by constructing the structural causal model and the cross-modal causal network, the core defects of heavy correlation and light causal in the traditional technology are fundamentally solved. A PC algorithm is adopted to identify a real causal link among multi-modal data, a front door adjustment method is combined to quantify the causal contribution degree of a core driving factor, false correlation and confounding factor interference are effectively separated, the capturing ability of an early-stage AD pathological signal is remarkably improved, the accuracy of AD attack risk prediction and MCI conversion prediction is greatly improved, and the prediction accuracy of the early-stage AD pathological signal is greatly improved. A more reliable quantitative basis is provided for clinical early diagnosis, and diagnosis misjudgment caused by data deviation is avoided.

Description

Technical Field

[0001] This invention relates to the field of dynamic prediction technology for Alzheimer's disease, and more specifically, to a dynamic prediction method for Alzheimer's disease based on multimodal causal inference. Background Technology

[0002] Alzheimer's disease (AD), as a progressive neurodegenerative disease, has become a major health challenge for aging societies worldwide. Its core harm lies in the insidious pathological progression and irreversible cognitive decline. Therefore, early and accurate prediction and dynamic progression monitoring are of key significance for clinical intervention and prognosis improvement.

[0003] Existing technologies have significant shortcomings in practical applications. First, multimodal data fusion only involves simple feature splicing or correlation modeling, without considering causal relationships between variables. This makes the model susceptible to confounding factors such as age and comorbidities, unable to distinguish between spurious associations and real pathogenic drivers, thus masking weak early AD pathological signals and resulting in insufficient accuracy in early diagnosis. Second, existing technologies are mostly static prediction models, lacking the ability to model dynamic time series. They can only output risk assessment results at a single time point, failing to capture the dynamic evolution of AD pathological progression and making it difficult to accurately quantify the rate of disease progression and key inflection points. This limits their guiding value for the timing of clinical intervention. Therefore, this paper proposes a dynamic prediction method for Alzheimer's disease based on multimodal causal inference. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic prediction method for Alzheimer's disease based on multimodal causal inference, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, a dynamic prediction method for Alzheimer's disease based on multimodal causal inference is provided, including the following steps: S1. Collect multi-source related data from users, quantify and extract features from the multi-source related data to form a factor feature set; S2. Establish a causal network for Alzheimer's disease, input the factor feature set into the causal network to allocate causal link nodes, assign corresponding factor features to each causal link, and retain the factor features of unassigned causal links to establish a candidate causal set. S3. Obtain the time nodes corresponding to the factor features, combine the factor features of each causal link in a time-segmented manner to simulate the disease status score, obtain the disease status score of each time period, and extract the causal contribution between each factor feature. S4. Sort the symptom status scores by time period, perform change analysis and abnormal feature analysis on the factor characteristics corresponding to adjacent symptom status scores, and obtain factor change data and mark abnormal features based on the verification results. S5. Combine the abnormal features with the candidate causal feature set to conduct causal link supplementation analysis. Based on the analysis results, add the factor features of the candidate causal set into the causal link. Then, based on the factor features corresponding to the causal link in each time period, combine the factor change data and the disease status score to conduct feature influence analysis and obtain the contribution of each factor feature to the disease status score. S6. Establish a dynamic prediction model by combining the contribution scores and causal contribution scores of each factor characteristic with the latest disease status score. Input the simulated factor characteristics into the dynamic prediction model to output the disease status score of Alzheimer's disease in the future period.

[0006] As a further improvement to this technical solution, S1 collects multi-source related data on Alzheimer's disease related to the user. The multi-source related data includes gene testing results, physiological indicators, family medical history, vascular examination reports, and comorbidity diagnosis and treatment records, and is split into five dimensions: gene factors, physical factors, hereditary factors, vascular factors, and comorbidity factors.

[0007] As a further improvement to this technical solution, in S1, gene factors, physical factors, genetic factors, vascular factors, and comorbidity factors are uniformly quantified by normalization method and one-hot encoding. At the same time, PCA cooling is used to extract factor features from the uniformly quantified gene factors, physical factors, genetic factors, vascular factors, and comorbidity factors. Then, the extracted factor features are summarized to form a factor feature set.

[0008] As a further improvement to this technical solution, in S2, a causal network for Alzheimer's disease is established based on a structural causal model. The preset causal link nodes are risk factors, symptom indicators, and cognitive decline. Each causal link node is interconnected. When a factor feature is added to a causal link node, each causal link node is synchronized in real time. The causal link nodes learn the conditional independence between factor features and causal network nodes through the PC algorithm, and assign features that satisfy causal association to the link where the corresponding node is located. The candidate causal set is the set of factor features that failed the conditional independence test.

[0009] As a further improvement to this technical solution, in step S3, the time nodes corresponding to the factor features are obtained, and then continuous time periods are divided according to the time nodes corresponding to the factor features, thereby generating multiple time periods; The factors and features contained in each causal link are matched according to the time period. Then, the causal links and factors and features corresponding to each time period are mapped and fused with the cognitive scale through the causal model to simulate the disease status score and obtain the disease status score of the user in each time period. Among them, the disease status score is a dimensionless score that combines the level of pathological markers, cognitive function score, and degree of vascular damage. The causal contribution of each factor is extracted using the front-door adjustment method. The causal contribution represents the strength of the correlation between the characteristics of each factor and its corresponding causal link.

[0010] As a further improvement to this technical solution, in step S4, the symptom status scores are sorted by time period to obtain the symptom status scores of users in different time periods. Then, the numerical change range and trend of the same factor feature in adjacent time periods are compared to generate factor change difference and change rate. By comparing the characteristics of the same factor in adjacent time periods using the trend deviation determination method, abnormal characteristics are determined and the abnormal factor characteristics are marked.

[0011] As a further improvement to this technical solution, in S5, the correlation coefficient between the abnormal feature and each feature in the candidate causal set is calculated, candidate features with a correlation coefficient ≥ 0.6 are screened, and then confounding variables are controlled by propensity score matching to verify the causal association between highly correlated candidate features and Alzheimer's disease. Significantly correlated candidate features are assigned to the most relevant causal link, and the disease status score is updated synchronously when the causal link is updated. The score contribution is calculated using SHAP, outputting the positive and negative contribution values ​​and contribution percentages of each factor feature to the disease status score.

[0012] As a further improvement to this technical solution, in step S6, a dynamic prediction model is established by combining the contribution degree and causal contribution degree of the corresponding scores of each factor characteristics and the score of the disease status in the latest time period. The dynamic prediction model is an LSTM-Transformer hybrid model. The model input includes the contribution of each factor feature score, causal contribution, and the latest time period disease status score. The model training uses multi-center longitudinal follow-up data. The input layer receives the contribution scores of each factor's characteristics, the causal contribution scores, and the latest time-period symptom status scores. The encoding layer strengthens the weights of high-contribution features through a causal attention mechanism, and then the LSTM layer captures short-term temporal dependencies. The output layer outputs the symptom status score and corresponding confidence interval for future time periods through a fully connected layer; Among them, the simulated factor characteristics include characteristic data under natural progression scenarios and characteristic data under intervention scenarios; The symptom status score ranges from 1 to 100, with different scores corresponding to different risk levels of Alzheimer's disease.

[0013] As a further improvement to this technical solution, the dynamic prediction model adopts an incremental learning mechanism. After each new user-related multi-source data is added, the latest score contribution and causal contribution obtained in step S5 are re-inputted to iteratively optimize the model parameters and improve the prediction accuracy of the disease status score in future periods.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This dynamic prediction method for Alzheimer's disease based on multimodal causal inference addresses the core deficiency of traditional techniques—emphasizing correlation while neglecting causality—by constructing a structural causal model and a cross-modal causal network. It employs a PC algorithm to identify genuine causal links between multimodal data and combines this with a front-door adjustment method to quantify the causal contribution of core driving factors. This effectively separates spurious associations and confounding factors, significantly improving the ability to capture early AD pathological signals. This greatly enhances the accuracy of predicting AD incidence risk and MCI conversion, providing more reliable quantitative evidence for early clinical diagnosis and avoiding diagnostic misjudgments caused by data bias.

[0015] 2. This dynamic prediction method for Alzheimer's disease based on multimodal causal inference upgrades from static risk assessment to dynamic progression tracking through temporal segmentation and dynamic modeling. Utilizing an LSTM-Transformer hybrid model to capture the short-term temporal dependence and long-term evolutionary trends of pathological progression, and incorporating new follow-up data in real-time through an incremental learning mechanism, it can accurately quantify the individual's AD pathological progression rate, cognitive decline slope, and key inflection points. This solves the problem that traditional static models cannot track dynamic disease changes, providing refined guidance for the selection of clinical intervention timing and helping to reduce the rate of disease progression and medical costs. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the dynamic prediction method for Alzheimer's disease based on multimodal causal inference of the present invention. Figure 2 This is a flowchart of S1 of the present invention; Figure 3 This is a flowchart of S2 of the present invention; Figure 4 This is a flowchart of S4 of the present invention; Figure 5 This is a flowchart of S6 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] Please see Figures 1-5 As shown, the purpose of this embodiment is to provide a dynamic prediction method for Alzheimer's disease based on multimodal causal inference, including the following steps: S1. Collect multi-source related data from users, quantify and extract features from the multi-source related data in a unified manner to form a factor feature set; transform the scattered and heterogeneous multi-source data into a standardized and modelable factor feature set to provide high-quality data support for subsequent causal analysis and dynamic prediction. S1 collects multi-source related data on Alzheimer's disease from users. This multi-source related data includes gene testing results (such as APOE genotyping, APP / PSEN1 gene mutation sites), physiological indicators (such as age, blood pressure, blood sugar, BMI), family medical history (such as the age of onset and disease status of first-degree relatives with AD), vascular examination reports (such as the degree of cerebral small vessel sclerosis, cerebral blood flow), and comorbidity diagnosis and treatment records (such as Parkinson's disease, history of depression). The data is then broken down into five dimensions: genetic factors, physical factors, hereditary factors, vascular factors, and comorbidity factors.

[0019] Genetic factors: Gene variations and expression levels related to AD pathology in gene testing results; Body factors: physiological indicators and related lifestyle data (such as metabolic status derived from blood pressure / blood sugar). Genetic factors: family medical history and data related to genetic susceptibility; Vascular factors: Structural / functional indicators and medical history of vascular diseases in the vascular examination report; Comorbid factors: Disease type, duration of diagnosis, and treatment details in the comorbid medical records.

[0020] In S1, gene factors, physical factors, genetic factors, vascular factors, and comorbidity factors are uniformly quantified by normalization method and one-hot coding. At the same time, PCA cooling is used to extract factor features from the uniformly quantified gene factors, physical factors, genetic factors, vascular factors, and comorbidity factors. Then, the extracted factor features are summarized to form a factor feature set.

[0021] For continuous data such as gene expression levels, blood pressure, blood glucose, and cerebral blood flow, Z-score normalization is used to eliminate dimensional differences and ensure that the data can be fused across dimensions. For categorized data such as genotype (e.g., APOE ε2 / ε3 / ε4), medical history (e.g., with / without hypertension), and comorbidity type, one-hot encoding is used to convert them into binary feature vectors that the model can recognize. All data quantified across the five dimensions are merged to construct an initial high-dimensional data matrix. PCA (Principal Component Analysis) is then used to reduce the dimensionality of the high-dimensional data matrix, retaining principal components (core features) with a cumulative variance contribution rate ≥90%, and eliminating redundant features to reduce the computational complexity of subsequent modeling. The core features extracted from the five dimensions by PCA are then summarized to form a standardized, low-redundancy factor feature set. The feature set must contain key information for each dimension, while ensuring that the data format is consistent and suitable for subsequent causal network modeling.

[0022] S2. Establish a causal network for Alzheimer's disease, input the factor feature set into the causal network to allocate causal link nodes, assign corresponding factor features to each causal link, and retain factor features of unassigned causal links to establish a candidate causal set; achieve accurate matching of features and causal links, while retaining potential causal features, and solve the defect of traditional prediction models that only focus on correlation and ignore causality. In S2, a causal network for Alzheimer's disease is established based on a structural causal model. The causal link nodes are preset to be risk factors, symptom indicators, and cognitive decline. Each causal link node is interconnected. When a factor feature is added to a causal link node, each causal link node is synchronized in real time. Risk factors → pathological indicators: Risk factors (such as gene mutations, hypertension) drive abnormalities in pathological indicators (such as Aβ deposition, tau phosphorylation); Pathological indicators → cognitive decline: Abnormal pathological indicators directly lead to cognitive decline (such as memory loss and executive function impairment). Risk factors → cognitive decline: Some risk factors (such as severe comorbidities) can directly affect cognitive function, forming an indirect correlation. A reserved node feature access interface is provided. When a new factor feature is added to a node, the parameters of the associated nodes are automatically updated (e.g., when the APOEε4 gene is added to the risk factor node, its influence weight on Aβ deposition in the pathological indicator node is updated synchronously), ensuring the consistency of the causal network.

[0023] Causal link nodes learn the conditional independence between factor features and causal network nodes through PC algorithm, and assign features that satisfy causal association to the link where the corresponding node is located. The factor feature set generated by S1 is associated with the variable set corresponding to the three core nodes to construct an initial variable pool (variable = factor feature + node core variable), and the significance level α = 0.05 (test threshold) is set. Then, all variable pairs (factor feature - node variable, factor feature - factor feature) are traversed. First, unconditional independence is tested: if the variable pair is independent, the potential association between the two is removed; if they are not independent, the association edge is retained to form the initial skeleton. Gradually increase the dimension of the condition set (starting from 1). For the related edges retained in the skeleton, test whether the variable pairs are independent under the given condition set: if they are independent, remove the related edge; if they are not independent, retain the related edge and proceed to the next dimension of the condition set for testing, until the dimension of the condition set reaches the preset upper limit (usually 1 / 2 of the total number of variables). Based on domain knowledge of AD pathological mechanisms (such as gene mutation preceding pathological abnormalities) and directional rules of PC algorithms (such as V-shaped structure determination), causal flow is labeled for the retained associated edges to ensure that the flow from factor features to corresponding nodes is logical (such as APOE gene → risk factor node). If a certain factor feature has a significant causal relationship with a variable of a certain core node (test P value < α = 0.05), then the feature is assigned to the causal link of the corresponding node.

[0024] The candidate causal set is the set of factor features that failed the conditional independence test.

[0025] Collect the features of factors that failed the conditional independence test (test p value ≥ α = 0.05), form a candidate causal set, and record the reasons for each feature's failure (such as insufficient sample size or the presence of uncontrolled confounding variables), so as to reserve a basis for the subsequent causal link supplementary analysis in S5.

[0026] S3. Obtain the time nodes corresponding to the factor features, combine the factor features of each causal link in a time-segmented manner to simulate the disease status score, obtain the disease status score of each time period, and extract the causal contribution between each factor feature; transform the static causal features into a dynamic disease progression trajectory, and quantify the correlation strength of features with the causal link to provide time-series data support for subsequent anomaly identification and dynamic prediction. In S3, the time nodes corresponding to the factor features are obtained, and then continuous time periods are divided according to the time nodes corresponding to the factor features, thereby generating multiple time periods; Extract the data acquisition timestamps of all features in the factor feature set (such as baseline detection time, 6-month follow-up time, 1-year follow-up time, 3-year follow-up time, etc.) to form a set of time nodes. Then, construct non-overlapping continuous time series windows with adjacent time nodes as boundaries to generate multiple time periods. If there is no factor feature data in a certain time period, merge it into the adjacent previous time period to ensure that each time period has effective data support. The factors and features contained in each causal link are matched according to the time period. Then, the causal links and factors and features corresponding to each time period are mapped and fused with the cognitive scale through the causal model to simulate the disease status score and obtain the disease status score of the user in each time period. Traverse the three established causal links, extract the factor features assigned to each link, obtain the rule of timestamp ∈ time period boundary according to the features, match the factor features of each link with the corresponding time period, and form a three-dimensional mapping relationship of time period-causal link-factor feature. The weights of the factors and features within each link are determined based on the causal model (the weights are positively correlated with the causal relationship of the features). The weighted sum of all feature values ​​of a link in each time period is then obtained to get the median score of that link in that time period. The intermediate scores of each link are mapped to the standard scoring range of clinical cognitive scales (such as MMSE and ADAS-Cog) to correct for scoring bias (the intermediate scores of the links are scaled proportionally to the 0-30 range of the cognitive scale). The weights for the three pathways are assigned based on the proportion of AD pathological contribution (preset: risk factor → pathological indicator pathway weight w1=0.4, pathological indicator → cognitive decline pathway weight w2=0.4, risk factor → cognitive decline pathway weight w3=0.2). The total symptom status score for this period is calculated. This calculation is repeated for all time periods to obtain the set of symptom status scores for each time period. The formula is as follows: ; in, Let be the intermediate score of a causal link in the i-th time period. This represents the number of factor features matched by the link in time period i. Let j be the quantized value of the j-th feature in the i-th time period. Let the j-th feature be the weight in this link; ; in, The score for the calibrated link. for, and These are the minimum and maximum standard scores for the Clinical Cognition Scale; ; in, The score represents the disease status in the i-th time period (1-100, with a lower score indicating a higher risk of AD). , The calibration scores for the three links in the i-th time period; Among them, the disease status score is a dimensionless score that combines the level of pathological markers, cognitive function score, and degree of vascular damage. The causal contribution of each factor was extracted using the front-door adjustment method. The causal contribution represents the strength of the association between each factor's characteristics and its corresponding causal link. The formula is as follows: ; in, To intervene in the average causal effect of X on mediator Z, This indicates the presence of intervention X (e.g., carrying the APOEε4 gene). The conditional expectation of the mediating variable Z after intervention in X (e.g., the mean of Aβ deposition). ; in, The average causal effect of mediator Z on outcome Y. and These represent outliers and normal values ​​of the mediating variable Z (such as the pathological threshold values ​​for Aβ deposition). To represent the conditional expectation of outcome variable Y (such as the mean score of pathological indicators) after intervention Z; ; in, Causal contribution degree, causal contribution degree The larger the absolute value, the stronger the correlation between the factor feature and the causal link; a positive value indicates that it promotes AD progression, while a negative value indicates that it inhibits AD progression.

[0027] S4. Sort the disease status scores by time period, perform change analysis and abnormal feature analysis on the factor characteristics corresponding to adjacent disease status scores, obtain factor change data and mark abnormal features based on the verification results; identify key change nodes and abnormal driving features in AD progression through time series comparison, provide targeted basis for subsequent causal link supplementation and optimization, and solve the problem of incomplete initial links in the causal network. In S4, the symptom status scores are sorted by time period to obtain the symptom status scores of users in different time periods. Then, the magnitude and trend of the numerical changes of the same factor feature in adjacent time periods are compared to generate factor change difference and change rate. Extract the symptom status scores for each time period. Sort the scores in ascending order according to the time sequence (e.g., baseline → 6 months → 1 year → 3 years) to form a temporally ordered sequence of symptom status scores. Based on the three-dimensional mapping relationship of time period-causal link-factor feature, extract the quantified (normalized) value of the same factor feature between two adjacent time periods to form feature time pairs. Then, for each feature's adjacent time pairs, calculate the numerical difference between the later and earlier time periods to reflect the absolute change magnitude of the feature. Combined with the time interval between two adjacent time periods (e.g., 6 months, 1 year), convert the change difference into the rate of change per unit time, eliminating the influence of different time intervals on the judgment of the degree of change. The formula is as follows:

[0028] in, Let be the difference in change of the j-th factor feature between adjacent time periods i and i+1. This is the quantized value of the j-th feature in the (i+1)th time period (the next time period). Let be the quantized value of the j-th feature in the i-th time period (previous time period); ; in, Let the rate of change of the j-th factor feature be the rate of change between the i-th and i+1-th time periods. Let be the time interval between the i-th and i+1-th time periods. The timestamps are for the (i+1)th and i-th time periods; By comparing the characteristics of the same factor in adjacent time periods using the trend deviation determination method, abnormal characteristics are determined and the abnormal factor characteristics are marked.

[0029] Based on historical longitudinal data of healthy individuals or AD clinical guidelines, establish normal temporal trend ranges for various factors (such as tau protein deposition rate and blood pressure change trends in healthy individuals), fit the normal trend slope, and for the same feature in adjacent time periods, fit the actual trend through linear regression to obtain the actual trend slope. Then, quantify the degree of deviation between the actual trend slope and the normal trend slope, and set a deviation threshold (based on the clinically normal fluctuation range, such as 50%). If the deviation exceeds the threshold, the feature is determined to be an abnormal feature.

[0030] S5. Combine the abnormal features with the candidate causal feature set to conduct causal link supplementation analysis. Based on the analysis results, add the factor features of the candidate causal set into the causal link. Then, based on the factor features corresponding to the causal link in each time period, combine the factor change data and the disease status score to conduct feature impact analysis, obtain the contribution of each factor feature to the disease status score; optimize and improve the causal network, quantify the direct impact of each feature on the disease status score, realize the upgrade of causal network → precise weight, and improve the interpretability and accuracy of the subsequent prediction model. In S5, the correlation coefficient between the abnormal feature and each feature in the candidate causal set is calculated, and candidate features with a correlation coefficient ≥ 0.6 are screened. Then, propensity score matching is used to control confounding variables, and the causal association between highly correlated candidate features and Alzheimer's disease is verified. Significantly correlated candidate features are assigned to the most relevant causal links. When the causal links are updated, the disease status score is updated synchronously. Extract all anomalous features marked in step S4 and all features in the candidate causal set. Calculate the correlation coefficient between each anomalous feature and each candidate feature (the best linear correlation measurement method, which is highly interpretable and computationally efficient). Construct a correlation matrix and then retain candidate features with a correlation coefficient ≥ 0.6 with any anomalous feature to form a subset of highly correlated candidate features. Variables that are associated with both highly correlated candidate features and Alzheimer's disease (AD) outcomes (such as age, gender, education level, and history of underlying diseases) are selected as the set of confounding variables to be controlled. Using the presence or absence of a highly correlated candidate feature cq as the dependent variable and the set of confounding variables as the independent variable, a logistic regression model was used to calculate the propensity score (i.e., the probability of carrying the feature) for each sample. Then, a 1:1 nearest neighbor matching method (the optimal matching method with good balancing effect and high computational efficiency) was used to match one non-confounding sample with the closest propensity score for each sample carrying the feature. After matching, a t-test was used to verify the distribution difference of the two groups of confounding variables to ensure that the confounding variables between the groups were balanced after matching (P>0.05). Compare the differences in AD outcomes (such as symptom status score and AD onset status) between the two matched groups, and use the Mann-Whitney U test (or t test) to determine the significance of the difference. If P < 0.05, then the highly associated candidate feature has a significant causal relationship with AD. Candidate features that pass the causal effect test will be assigned to the most relevant causal links based on their associated AD pathological mechanisms (e.g., if related to vascular damage, they will be assigned to the vascular factors → pathological indicators link). At the same time, due to the new features added to the causal links, the intermediate scores of the single links and the total disease status scores for each time period will be recalculated. Based on the original weighted sum of features, the quantized value of the new feature multiplied by the corresponding weight term is added to ensure that the score is consistent with the updated causal link.

[0031] The score contribution is calculated using SHAP, outputting the positive and negative contribution values ​​and contribution percentages of each factor feature to the disease status score.

[0032] Organize the quantitative values ​​of factor features (including newly added features) and corresponding disease status scores for each time period, construct a feature-score dataset, with disease status score as the dependent variable and factor features as the independent variable, and call the SHAP interpreter based on the constructed dynamic prediction model; The SHAP interpreter calculates the SHAP value of each factor feature for the disease status score at each time period, which directly represents the marginal contribution of the feature to the score. The SHAP value of each feature is converted into a contribution percentage (the proportion of the absolute value of the SHAP value of a single feature to the sum of the absolute values ​​of all feature SHAP values), and positive contributions (SHAP value > 0, increasing the disease score and reducing the risk of AD) or negative contributions (SHAP value < 0, decreasing the disease score and increasing the risk of AD) are marked according to the positive or negative SHAP value.

[0033] S6. A dynamic prediction model is established by combining the contribution scores and causal contributions of each factor characteristic with the latest disease status score. The model is input into the simulated factor characteristics to output the disease status score for future periods of Alzheimer's disease. By integrating causal contribution, score contribution, and time-series data, a dynamic prediction model is constructed to achieve accurate prediction of future AD progression and continuous model optimization, while also considering the clinical needs of natural progression and intervention scenarios. In S6, a dynamic prediction model is established by combining the contribution of scores and causal contribution of each factor characteristic with the latest disease status score. The dynamic prediction model is an LSTM-Transformer hybrid model. The model input includes the contribution of each factor feature score, causal contribution, and the latest time period disease status score. The model training uses multi-center longitudinal follow-up data. The input layer receives the contribution scores of each factor feature, the causal contribution score, and the latest time period symptom status score; it also receives the input feature vector and maps the feature dimensions to a unified embedding dimension (e.g., 256 dimensions) of the model through a fully connected layer to obtain the feature embedding matrix. The encoding layer strengthens the weights of high-contribution features through a causal attention mechanism, and then the LSTM layer captures short-term temporal dependencies. Calculate the comprehensive contribution weight of each feature (integrating score contribution and causal contribution), and weight the feature embedding matrix to highlight the influence of high-contribution features (such as APOE gene and hippocampal atrophy rate). The system receives the weighted feature embedding, captures short-term temporal dependencies (such as the feature change trend of the last three time periods) through LSTM units, and outputs temporal feature encoding. The output layer outputs the symptom status score and corresponding confidence interval for future time periods through a fully connected layer; The temporal feature encoding is mapped by two fully connected layers. The first layer outputs the predicted mean of the score for the future time period, and the second layer outputs the predicted variance. Finally, the mean and variance are combined to calculate the 95% confidence interval and output the symptom status score and confidence interval for the future time period. Among them, the simulated factor characteristics include characteristic data under natural progression scenarios and characteristic data under intervention scenarios; The mean squared error (MSE) between the predicted score and the true score is used as the loss function. The Adam optimizer (the best optimizer, which converges quickly and is highly stable) is used to train the model. The iteration continues until the loss on the validation set no longer decreases. In a natural progression scenario, based on historical feature change trends (such as the rate of change in step S4), simulated feature data for future periods (without intervention) is generated and input into the model to obtain a predicted symptom score under natural progression. The intervention scenario simulates the characteristic changes corresponding to clinical intervention measures, generates simulated characteristic data after the intervention, and inputs it into the model to obtain the predicted symptom score under the intervention scenario.

[0034] The disease status score is 1-100 (e.g., ≥80 points is low risk, 60-79 points is medium risk, and <60 points is high risk), with different scores corresponding to different risk levels of Alzheimer's disease.

[0035] The dynamic prediction model adopts an incremental learning mechanism. After each new user-related multi-source data is added, the latest score contribution and causal contribution obtained in step S5 are re-inputted to iteratively optimize the model parameters and improve the prediction accuracy of the symptom status score in future periods.

[0036] For each new batch of user multi-source data (e.g., ≥300 cases), repeat step S5 to calculate the updated score contribution and causal contribution. Incremental learning is performed using the fine-tuning method—the weights of the model's encoding layer (causal attention + LSTM) are frozen (preserving the learned temporal and causal features), and only the parameters of the input and output layers are updated; the incremental dataset constructed from the newly added data is used as the training data, and the model is fine-tuned with a small learning rate (such as 1 / 10 of the initial learning rate) to minimize the prediction loss of the new data. Then, the prediction accuracy is verified by testing the new data after fine-tuning, ensuring that the mean absolute error (MAE) of the updated model is ≤0.3 points (the disease score range is 1-100), and the iterative optimization is completed.

[0037] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic prediction method for Alzheimer's disease based on multimodal causal inference, characterized in that: Includes the following steps: S1. Collect multi-source related data from users, quantify and extract features from the multi-source related data to form a factor feature set; S2. Establish a causal network for Alzheimer's disease, input the factor feature set into the causal network to allocate causal link nodes, assign corresponding factor features to each causal link, and retain the factor features of unassigned causal links to establish a candidate causal set. S3. Obtain the time nodes corresponding to the factor features, combine the factor features of each causal link in a time-segmented manner to simulate the disease status score, obtain the disease status score of each time period, and extract the causal contribution between each factor feature. S4. Sort the symptom status scores by time period, perform change analysis and abnormal feature analysis on the factor characteristics corresponding to adjacent symptom status scores, and obtain factor change data and mark abnormal features based on the verification results. S5. Combine the abnormal features with the candidate causal feature set to conduct causal link supplementation analysis. Based on the analysis results, add the factor features of the candidate causal set into the causal link. Then, based on the factor features corresponding to the causal link in each time period, combine the factor change data and the disease status score to conduct feature influence analysis and obtain the contribution of each factor feature to the disease status score. S6. Establish a dynamic prediction model by combining the contribution scores and causal contribution scores of each factor characteristic with the latest disease status score. Input the simulated factor characteristics into the dynamic prediction model to output the disease status score of Alzheimer's disease in the future period.

2. The method for dynamic prediction of Alzheimer's disease based on multimodal causal inference according to claim 1, characterized in that: The S1 collects multi-source related data on Alzheimer's disease for users. The multi-source related data includes gene testing results, physiological indicators, family medical history, vascular examination reports, and comorbidity diagnosis and treatment records, and is split into five dimensions: gene factors, physical factors, hereditary factors, vascular factors, and comorbidity factors.

3. The method for dynamic prediction of Alzheimer's disease based on multimodal causal inference according to claim 2, characterized in that: In S1, gene factors, physical factors, genetic factors, vascular factors, and comorbidity factors are uniformly quantified by normalization method and one-heat coding. At the same time, PCA cooling is used to extract factor features from the uniformly quantified gene factors, physical factors, genetic factors, vascular factors, and comorbidity factors. Then, the extracted factor features are summarized to form a factor feature set.

4. The method for dynamic prediction of Alzheimer's disease based on multimodal causal inference according to claim 1, characterized in that: In S2, a causal network for Alzheimer's disease is established based on a structural causal model. The preset causal link nodes are risk factors, symptom indicators, and cognitive decline. Each causal link node is interconnected. When a factor feature is added to a causal link node, each causal link node is synchronized in real time. The causal link nodes learn the conditional independence between factor features and causal network nodes through the PC algorithm, and assign features that satisfy causal association to the link where the corresponding node is located. The candidate causal set is the set of factor features that failed the conditional independence test.

5. The method for dynamic prediction of Alzheimer's disease based on multimodal causal inference according to claim 1, characterized in that: In step S3, the time nodes corresponding to the factor features are obtained, and then continuous time periods are divided according to the time nodes corresponding to the factor features, thereby generating multiple time periods; The factors and features contained in each causal link are matched according to the time period. Then, the causal links and factors and features corresponding to each time period are mapped and fused with the cognitive scale through the causal model to simulate the disease status score and obtain the disease status score of the user in each time period. Among them, the disease status score is a dimensionless score that combines the level of pathological markers, cognitive function score, and degree of vascular damage. The causal contribution of each factor is extracted using the front-door adjustment method. The causal contribution represents the strength of the correlation between the characteristics of each factor and its corresponding causal link.

6. The method for dynamic prediction of Alzheimer's disease based on multimodal causal inference according to claim 1, characterized in that: In S4, the symptom status scores are sorted by time period to obtain the symptom status scores of users in different time periods. Then, the magnitude and trend of the numerical changes of the same factor feature in adjacent time periods are compared to generate factor change difference and change rate. By comparing the characteristics of the same factor in adjacent time periods using the trend deviation determination method, abnormal characteristics are determined and the abnormal factor characteristics are marked.

7. The method for dynamic prediction of Alzheimer's disease based on multimodal causal inference according to claim 1, characterized in that: In S5, the correlation coefficient between the abnormal feature and each feature in the candidate causal set is calculated, and candidate features with a correlation coefficient ≥ 0.6 are screened. Then, confounding variables are controlled by propensity score matching to verify the causal association between highly correlated candidate features and Alzheimer's disease. Significantly correlated candidate features are assigned to the most relevant causal links. When the causal links are updated, the disease status score is updated synchronously. The score contribution is calculated using SHAP, outputting the positive and negative contribution values ​​and contribution percentages of each factor feature to the disease status score.

8. The method for dynamic prediction of Alzheimer's disease based on multimodal causal inference according to claim 1, characterized in that: In S6, a dynamic prediction model is established by combining the contribution scores and causal contribution scores of each factor characteristic with the latest disease status score. The dynamic prediction model is an LSTM-Transformer hybrid model. The model input includes the contribution of each factor feature score, causal contribution, and the latest time period disease status score. The model training uses multi-center longitudinal follow-up data. The input layer receives the contribution scores of each factor's characteristics, the causal contribution scores, and the latest time-period symptom status scores. The encoding layer strengthens the weights of high-contribution features through a causal attention mechanism, and then the LSTM layer captures short-term temporal dependencies. The output layer outputs the symptom status score and corresponding confidence interval for future time periods through a fully connected layer; Among them, the simulated factor characteristics include characteristic data under natural progression scenarios and characteristic data under intervention scenarios; The symptom status score ranges from 1 to 100, with different scores corresponding to different risk levels of Alzheimer's disease.

9. The method for dynamic prediction of Alzheimer's disease based on multimodal causal inference according to claim 8, characterized in that: The dynamic prediction model adopts an incremental learning mechanism. After each new user-related multi-source data is added, the latest score contribution and causal contribution obtained in step S5 are re-inputted to iteratively optimize the model parameters and improve the prediction accuracy of the disease status score in future periods.

Images