Method and device for assessing combined risk of chronic diseases and comorbidities, and computer equipment

Abstract

This application provides a method, apparatus, and computer device for joint risk assessment of chronic diseases and comorbidities, relating to the field of medical and health technology. The method includes: performing hierarchical preprocessing on health examination-related data from multiple subjects to generate a standardized dataset containing single-disease labels and comorbidity combination labels; inputting the target data into a single-framework multi-objective collaborative ensemble model to obtain single-disease and comorbidity prediction values; the single-disease and comorbidity prediction values ​​are obtained synchronously through a single inference by the single-framework multi-objective collaborative ensemble model; calculating a comorbidity risk gain value based on the single-disease and comorbidity prediction values, and generating a structured assessment result based on the comorbidity risk gain value. This method can improve the accuracy of joint risk assessment of chronic diseases and comorbidities.

Description

Technical Field

[0001] This application relates to the field of medical and health technology, specifically to a method, device, and computer equipment for assessing the combined risk of chronic diseases and comorbidities. Background Technology

[0002] Existing methods for assessing the risk of chronic diseases and comorbidities often suffer from low predictive accuracy. On one hand, most methods employ simplistic preprocessing of physical examination data, such as directly deleting missing records or using a single imputation method. This fails to effectively address the prevalent issues of high missing data rates, multi-source heterogeneity, and gender-specific characteristics in physical examination data, leading to noise from the input data being incorporated into subsequent prediction models. On the other hand, existing solutions often model single-disease prediction and comorbidity prediction as two independent tasks, or employ a method of first predicting single diseases and then simply superimposing them to determine comorbidities. This disrupts the information transfer between single-disease and comorbidity predictions, hindering effective feature sharing and collaborative error correction. Furthermore, multiple independent inferences or sequential calls to multiple models increase the chance of error accumulation, further reducing the reliability of the overall assessment.

[0003] Furthermore, existing technologies generally lack specialized modeling of synergistic effects between diseases when quantifying comorbidity risk. Most methods define comorbidity risk simply by determining whether a patient has multiple single diseases simultaneously, without considering the additional effects that may arise when two or more diseases coexist, exceeding the linear summation of independent risks. This simplistic summation approach ignores the pathophysiological connections and interactions between diseases, making it difficult for comorbidity risk assessment results to accurately reflect a patient's overall disease risk. Simultaneously, due to the lack of a collaborative optimization mechanism between single-disease prediction and comorbidity prediction, feature selection errors or biases in single-disease prediction cannot be corrected by the comorbidity prediction stage, and vice versa, limiting the model's predictive accuracy in complex disease scenarios. Therefore, existing solutions are insufficient in terms of accuracy in assessing the combined risk of chronic diseases and comorbidities. Summary of the Invention

[0004] This application provides a method, apparatus, and computer equipment for assessing the combined risk of chronic diseases and comorbidities, in order to improve the accuracy of such assessments.

[0005] In a first aspect, this application provides a method for joint risk assessment of chronic diseases and comorbidities, comprising: performing hierarchical preprocessing on physical examination-related data of multiple physical examination subjects to generate a standardized dataset containing single-disease labels and comorbidity combination labels; the physical examination-related data includes chronic disease diagnosis data and comorbidity diagnosis data; constructing a single-frame multi-objective collaborative integration model based on the standardized dataset; the single-frame multi-objective collaborative integration model is an integrated framework, the integrated framework including a single-disease prediction module and a comorbidity prediction module, the single-disease prediction module and the comorbidity prediction module reusing the feature representation layer and computation unit of the integrated framework, the single-disease prediction module and the comorbidity prediction module being collaboratively optimized through a global loss function; inputting target data into the single-frame multi-objective collaborative integration model to obtain single-disease prevalence prediction values ​​and comorbidity prevalence prediction values; the single-disease prevalence prediction values ​​and the comorbidity prevalence prediction values ​​are obtained synchronously by the single-frame multi-objective collaborative integration model through a single inference; calculating a comorbidity risk gain value based on the single-disease prevalence prediction value and the comorbidity prevalence prediction value, and generating a structured assessment result based on the comorbidity risk gain value.

[0006] Optionally, the step of constructing a single-frame multi-objective collaborative ensemble model based on the standardized dataset includes: processing the standardized dataset based on a single-head self-attention mechanism with prior constraints to calculate first association information, second association information, and third association information; wherein, the first association information is used to quantify the association strength of each key physical examination feature to a single chronic disease, the second association information is used to quantify the association strength of each key physical examination feature to a combination of common clinical comorbidities, and the third association information is used to quantify the risk synergistic effect between two or more chronic diseases; fusing the first association information, the second association information, and the third association information with the key physical examination feature to generate an enhanced feature set; and constructing the single-frame multi-objective collaborative ensemble model based on the enhanced feature set.

[0007] Optionally, the single-disease prediction module is used to adapt the risk prediction of each chronic disease under each preset time window, and dynamically adjust the feature weights according to the differences in biomarkers of different chronic diseases; the comorbidity prediction module is used to adapt the risk prediction of each comorbidity combination under each preset time window, and call the third association information to calculate the comorbidity risk.

[0008] Optionally, the output modes of the single-frame multi-objective collaborative integration model include a full mode and a lightweight mode. In the full mode, the single-frame multi-objective collaborative integration model performs a single inference based on all features of the enhanced feature set, and outputs single-disease prediction values ​​and comorbidity prediction values. In the lightweight mode, the single-frame multi-objective collaborative integration model performs a single inference based on a core feature subset, and outputs single-disease prediction values ​​and comorbidity prediction values. The core feature subset is determined by performing sensitivity analysis on the single-frame multi-objective collaborative integration model and optimizing the number of variables.

[0009] Optionally, the step of performing hierarchical preprocessing on the physical examination association data of multiple physical examination subjects to generate a standardized dataset containing single-disease labels and comorbidity combination labels includes: determining whether the physical examination subject is a chronic disease patient or a comorbidity patient based on the disease diagnosis results in the physical examination association data; if the physical examination subject is a single chronic disease patient, then the physical examination association data of the physical examination subject is marked as a single-disease positive sample, the time difference between the physical examination time and the disease diagnosis time of the physical examination subject is calculated, and a corresponding preset time window is matched for the single-disease positive sample based on the time difference; if the physical examination subject is a comorbidity patient, then the physical examination association data of the physical examination subject is marked as a comorbidity positive sample, and the time difference between the physical examination time and the initial diagnosis time is calculated. The time difference between the diagnosis time of comorbidity is used to match the corresponding preset time window for the comorbidity positive sample. If the examinee is a healthy person without a diagnosis record, the examinee's examination-related data is marked as a negative sample, and the time difference between the examination time and the virtual diagnosis time is set to a preset maximum value. The single-disease positive sample and the comorbidity positive sample are segmented according to their respective preset time windows, and the negative sample is allocated according to all preset time windows to obtain examination-related data under different time windows. All examination-related data are divided into male and female groups according to gender. Male-specific examination characteristics are retained in the male group, and female-specific examination characteristics are retained in the female group.

[0010] Optionally, the step of performing hierarchical preprocessing on the physical examination-related data of multiple physical examination subjects to generate a standardized dataset containing single-disease labels and comorbidity combination labels includes: performing structured processing on the physical examination-related data to obtain an initial dataset; the initial dataset includes single-disease labels and comorbidity combination labels, wherein each chronic disease corresponds to one single-disease label and each combination of chronic diseases corresponds to one comorbidity combination label; performing hierarchical cleaning processing on the initial dataset to obtain a clean dataset; performing feature reconstruction processing on the categorical variables in the clean dataset; extracting key physical examination features from the processed dataset, the key physical examination features including demographic features, routine examination indicators, specific examination indicators, and medical history-related features, and integrating the extracted data into the standardized dataset.

[0011] Optionally, the step of calculating the comorbidity risk gain value based on the single-disease morbidity prediction value and the comorbidity morbidity prediction value, and generating a structured assessment result based on the comorbidity risk gain value, includes: obtaining a first risk threshold set corresponding to each single disease under each preset time window, and a second risk threshold set corresponding to each comorbidity combination; comparing the single-disease morbidity prediction value with the first risk threshold set to determine the single-disease risk level to which the single-disease morbidity prediction value belongs; comparing the comorbidity morbidity prediction value with the second risk threshold set to determine the comorbidity risk level to which the comorbidity morbidity prediction value belongs; and calculating the comorbidity morbidity prediction value and all single-disease morbidity prediction values ​​constituting the comorbidity combination. The comorbidity risk gain value is calculated. If the comorbidity combination consists of two single diseases, the comorbidity risk gain value is equal to the predicted comorbidity incidence value minus the union probability of the predicted incidence values ​​of the two single diseases. If the comorbidity combination consists of three or more single diseases, the comorbidity risk gain value is calculated based on the difference between the predicted comorbidity incidence value and the union probability of the predicted incidence values ​​of all single diseases. A structured assessment result is output, in which each record includes at least a single disease type identifier, a preset time window identifier, the single disease risk level, the single disease predicted incidence value, and a comorbidity combination type identifier, the comorbidity risk level, the comorbidity predicted incidence value, and the comorbidity risk gain value.

[0012] Optionally, the method further includes:

[0013] Based on the first, second, and third association information, a first visual representation of the association strength between the key physical examination features and individual diseases, a second visual representation of the association strength between the key physical examination features and comorbidity combinations, and a ranking table of synergistic effects among diseases are generated. The first visual representation displays the association strength distribution of each key physical examination feature to each individual disease, the second visual representation displays the association strength distribution of each key physical examination feature to each comorbidity combination, and the ranking table displays the order of synergistic effects between different disease combinations. A preset algorithm is used to calculate the comprehensive contribution of each key physical examination feature to the predicted values ​​of the individual disease and the comorbidity, and a third visual representation of the contribution of each key physical examination feature to the prediction results is generated based on this comprehensive contribution. The third visual representation distinguishes between the positive and negative impacts of each key physical examination feature on the prediction results. The first, second, and ranking tables are integrated to output a report on the independent risk association degree of each disease and a report on the synergistic risk association degree of each key physical examination feature.

[0014] Secondly, this application provides a device for assessing the combined risk of chronic diseases and comorbidities, comprising: The first construction module is used to perform hierarchical preprocessing on the physical examination-related data of multiple physical examination subjects to generate a standardized dataset containing single disease labels and comorbidity combination labels; the physical examination-related data includes chronic disease diagnosis data and comorbidity diagnosis data; The second construction module is used to construct a single-framework multi-objective collaborative ensemble model based on the standardized dataset. The single-framework multi-objective collaborative ensemble model is an integrated ensemble framework, which includes a single-disease prediction module and a comorbidity prediction module. The single-disease prediction module and the comorbidity prediction module reuse the feature representation layer and computation unit of the integrated ensemble framework. The single-disease prediction module and the comorbidity prediction module are collaboratively optimized through a global loss function. The prediction module inputs target data into the single-frame multi-objective collaborative integration model to obtain single-disease and comorbidity prediction values; the single-disease and comorbidity prediction values ​​are obtained synchronously by the single-frame multi-objective collaborative integration model through a single inference. The results generation module is used to calculate the comorbidity risk gain value based on the single disease prevalence prediction value and the comorbidity prevalence prediction value, and generate a structured assessment result based on the comorbidity risk gain value.

[0015] Thirdly, this application provides a computer device, the computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the methods described above.

[0016] Compared with existing technologies, the beneficial effects of this application are as follows: By constructing a single-framework multi-objective collaborative integration model, the single-disease prediction module and the comorbidity prediction module reuse the same feature representation layer and computational unit, enabling single-disease prediction and comorbidity prediction to share the underlying feature space. Simultaneously, a global loss function is used to collaboratively optimize the two modules, allowing the feature weight adjustment for single-disease prediction and the mining of association information for comorbidity prediction to mutually enhance each other, thereby improving the model's ability to model complex disease relationships. The simultaneous output of single-disease and comorbidity prediction values ​​in a single inference reduces the accumulation of errors caused by multiple inferences. Calculating the comorbidity risk gain value quantifies the additional risk contributed by the synergistic effect of diseases, allowing the structured assessment results to more comprehensively reflect the true disease risk, thus improving the accuracy of joint risk assessment for chronic diseases and comorbidities. Attached Figure Description

[0017] Figure 1 A schematic diagram illustrating the steps of the combined risk assessment method for chronic diseases and comorbidities provided in the embodiments of this application.

[0018] Figure 2 This is a schematic diagram of the architecture of the joint risk assessment device for chronic diseases and comorbidities provided in the embodiments of this application. Detailed Implementation

[0019] The present application will now be described in further detail with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the subject matter of the present application to the following embodiments. All technologies implemented based on the content of the present application fall within the scope of protection of the present application.

[0020] In the description of the embodiments of this application, technical terms such as "first" and "second" only distinguish one entity or operation from another, and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary or secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] Please refer to Figure 1 , Figure 1 This diagram illustrates the steps of the combined risk assessment method for chronic diseases and comorbidities provided in this application. The combined risk assessment method for chronic diseases and comorbidities may include the following steps: S1. Perform hierarchical preprocessing on the physical examination association data of multiple physical examination subjects to generate a standardized dataset containing single disease labels and comorbidity combination labels.

[0023] Among them, physical examination-related data refers to a multi-dimensional data set related to the physical examination subject obtained from the information system of the physical examination center or hospital, including chronic disease diagnosis data and comorbidity diagnosis data. Specifically, it can include physical examination characteristics, physical examination time, chronic disease diagnosis results, comorbidity diagnosis results, and the time corresponding to each diagnosis result.

[0024] S2. Construct a single-framework multi-objective collaborative integration model based on standardized datasets.

[0025] Among them, the single-framework multi-objective collaborative integration model is an integrated framework. The integrated framework includes a single-disease prediction module and a comorbidity prediction module. The single-disease prediction module and the comorbidity prediction module reuse the feature representation layer and computing unit of the integrated framework. The single-disease prediction module and the comorbidity prediction module are collaboratively optimized through a global loss function.

[0026] In this embodiment, the single-framework multi-objective collaborative ensemble model refers to a single, indivisible ensemble machine learning model. This model is constructed as an ensemble framework with a unified structure, i.e., an integrated ensemble framework. Within this ensemble framework, there are two functional modules: a single-disease prediction module and a comorbidity prediction module. The single-disease prediction module outputs a predicted risk value for a single chronic disease, while the comorbidity prediction module outputs a predicted risk value for a combination of multiple diseases. The single-disease prediction module and the comorbidity prediction module are not independent models but share the same feature representation layer and computational units within the same ensemble framework. The feature representation layer is a network layer or computational component in the model used to extract and express feature information from input data, while the computational unit refers to the processing unit in the model that performs basic mathematical operations such as matrix operations and activation function operations.

[0027] The reuse of these components by the two modules indicates that during model operation, the single-disease prediction module and the co-disease prediction module share the same set of underlying parameters and computational resources, and there are no independent feature extraction paths. To enable the two modules to optimize their respective prediction objectives while sharing the underlying structure, a global loss function is used during model training to collaboratively optimize the single-disease prediction module and the co-disease prediction module. The global loss function is a single loss function constructed by adding the prediction errors of the single-disease prediction module and the co-disease prediction module according to preset weights.

[0028] The collaborative optimization of the single-disease prediction module and the comorbidity prediction module through a global loss function aims to enable both modules to optimize their respective prediction objectives while sharing the underlying structure. The global loss function is a single loss function constructed by adding the prediction errors of the single-disease prediction module and the comorbidity prediction module according to preset weights. During each parameter update, the model simultaneously reduces both the single-disease and comorbidity prediction errors. This allows the adjustment of feature weights in the single-disease prediction module to facilitate the mining of association information in the comorbidity prediction module, and conversely, the disease association analysis in the comorbidity prediction module optimizes the feature selection in the single-disease prediction module.

[0029] S3. Input the target data into the single-framework multi-objective collaborative integration model to obtain the predicted values ​​of single disease incidence and comorbidity incidence.

[0030] After the model is built and trained stably, the target data is input into the single-frame multi-objective ensemble model for prediction. Target data refers to standardized data obtained after hierarchical preprocessing of the physical examination correlation data of a subject to be evaluated. Upon receiving the target data, the single-frame multi-objective ensemble model performs a single forward propagation calculation, i.e., a single inference, simultaneously outputting the predicted values ​​for single-disease and comorbidities. "Synchronous output in a single inference" means that the prediction results for all single diseases and all comorbidities are obtained simultaneously within one computational cycle of the same model.

[0031] For example, the single disease prediction value is a probability value between 0 and 1, representing the likelihood that the examinee has a certain chronic disease within a preset time window; the comorbidity prediction value is also a probability value between 0 and 1, representing the likelihood that the examinee has multiple chronic diseases simultaneously within a preset time window.

[0032] S4. Calculate the comorbidity risk gain value based on the single disease prevalence prediction value and the comorbidity prevalence prediction value, and generate a structured assessment result based on the comorbidity risk gain value.

[0033] After obtaining the predicted values ​​for individual diseases and comorbidities, the model calculates a quantitative indicator called the comorbidity risk gain value based on these two values. The comorbidity risk gain value quantifies the additional risk caused by the synergistic effect of diseases. The specific method for calculating the comorbidity risk gain value is as follows: obtain the predicted values ​​for the individual diseases constituting a certain comorbidity combination, then calculate the union probability of these individual disease prediction values, and finally subtract this union probability from the predicted comorbidity value. The difference obtained is the comorbidity risk gain value. For example, for the comorbidity combination of hypertension and diabetes, the comorbidity risk gain value equals the predicted comorbidity value minus (the predicted hypertension value plus the predicted diabetes value minus their product). Based on the calculated comorbidity risk gain value, the model generates a structured assessment result. The structured assessment result refers to output information organized according to a preset field format. Each result record may include data such as single disease type identifier, time window identifier, single disease predicted value, comorbidity combination type identifier, comorbidity predicted value, and comorbidity risk gain value. This structured assessment result can be directly used for subsequent risk stratification, visualization, or clinical decision support.

[0034] In some optional embodiments, the step of performing hierarchical preprocessing on the health examination association data of multiple health examination subjects to generate a standardized dataset containing single disease labels and comorbidity combination labels may specifically include: Based on the disease diagnosis results in the physical examination association data, determine whether the physical examination subject is a chronic disease patient or a comorbid patient; if the physical examination subject is a single chronic disease patient, then mark the physical examination association data of the physical examination subject as a single disease positive sample, calculate the time difference between the physical examination time and the disease diagnosis time of the physical examination subject, and match the corresponding preset time window for the single disease positive sample based on the time difference; If the examinee is a person with comorbidities, the examinee's examination-related data is marked as a comorbidity-positive sample, the time difference between the examination time and the first diagnosis time of the comorbidity is calculated, and a corresponding preset time window is matched for the comorbidity-positive sample based on the time difference; if the examinee is a healthy person without a diagnosis record, the examinee's examination-related data is marked as a negative sample, and the time difference between the examination time and the virtual diagnosis time is set to a preset maximum value. The single-disease positive samples and the comorbid-disease positive samples are segmented according to their respective preset time windows, and the negative samples are allocated according to all preset time windows to obtain physical examination associated data under different time windows; all physical examination associated data are divided into male group and female group according to gender, male-specific physical examination characteristics are retained in the male group, and female-specific physical examination characteristics are retained in the female group.

[0035] For example, in the tiered preprocessing, it is necessary to determine whether each examinee belongs to the diseased population or the healthy population based on the disease diagnosis results in the physical examination-related data. The diseased population is further divided into individuals with a single chronic disease and individuals with comorbidities. Individuals with a single chronic disease refer to those who have only one chronic disease, such as those with only hypertension without diabetes or other diseases; individuals with comorbidities refer to those who have two or more chronic diseases simultaneously, such as those with both hypertension and diabetes.

[0036] For the different population groups mentioned above, this application embodiment employs a differentiated sample labeling and time window matching strategy. Specifically, the physical examination-related dataset can be represented as follows: ,in, For the set of physical examination features ( (This refers to the total number of features, including routine examination indicators, features shared by multiple diseases, and disease-specific biomarkers, etc.) For single-disease diagnostic label matrix ( For the total number of single diseases, such as (Corresponding to hypertension, diabetes, depression, and anxiety) Each Divided into two categories: For the first A single positive sample for a specific disease; For the first A single negative sample for a disease. Comorbidity tag matrix ( The total number of comorbidity combinations, such as (Corresponding to hypertension + diabetes, depression + hyperlipidemia, hypertension + coronary heart disease), each Divided into two categories: For the first Positive samples from comorbidity combinations; For the first Negative samples from comorbidity combinations. The examination time and diagnosis time can be in the format YYYY-MM-DD; For gender tags ( Male; (For women).

[0037] The preset time window set is defined as follows (Corresponding to 1 year, 3 years, 5 years, and 10 years, unit: year); The physical examination time for each examinee is... For the kth single disease, the diagnosis time of the disease is extracted from the clinical diagnosis record. For the l-th comorbidity combination, the time of the first comorbidity diagnosis is extracted as follows: .

[0038] If the individuals undergoing physical examinations are individuals with a single chronic disease, then their examination-related data are marked as single-disease positive samples. In this case, the time difference between the examination date and the corresponding disease diagnosis date is calculated. .like Falling within a certain preset time window Inside( If so, then the positive sample for that single disease is matched with that time window. For example, if a patient with hypertension is diagnosed 2.5 years later than the time of their physical examination, then... The timeframe is 2.5 years, and this sample is matched with a 3-year time window.

[0039] If the examinee is a person with comorbidities, then the examinee's examination-related data is marked as a positive comorbidity sample. In this case, the time difference between the examination date and the date of the first diagnosis of the comorbidity is calculated. .like Falling within a certain preset time window Within that time window, the positive sample from the comorbid disease is matched with that time window. For example, if a patient is first diagnosed with both hypertension and diabetes 0.8 years later than the time of their physical examination, then... The timeframe is 0.8 years, and this sample is matched with a 1-year time window.

[0040] If the examinee is a healthy individual with no prior medical diagnosis, their examination-related data will be marked as a negative sample. To ensure consistent processing, a virtual diagnosis time will be established. The time difference is 99 years. This value is much larger than the upper limit of all preset time windows, matching all time windows of the corresponding single disease / comorbidity combination. Therefore, negative samples were assigned to corresponding single disease and comorbidity combinations under all preset time windows (1 year, 3 years, 5 years, 10 years), indicating that the subject was considered disease-free within any time window.

[0041] After completing sample labeling and time window matching, single-disease positive samples and comorbid-disease positive samples are segmented according to their respective pre-defined time windows. That is, data from different time windows of the same physical examination subject are considered as independent training samples. Negative samples are allocated according to all pre-defined time windows, meaning that a negative sample is generated for the same healthy physical examination subject in each time window. After the above processing, physical examination correlation data under different time windows are obtained.

[0042] Then, all physical examination-related data are divided into male and female groups based on gender. Specifically, this can be done based on gender labels. Split: For the male group, For the female group. Based on this, it is further subdivided into more granular subsets according to the single-disease-comorbidity dimension. ,in They represent male or female respectively. Traverse all single diseases, The study iterates through all comorbidity combinations. Male-specific physical examination characteristics (e.g., prostate-related indicators) are retained in the male group, and female-specific physical examination characteristics (e.g., breast-related indicators) are retained in the female group. Simultaneously, each group retains its corresponding single-disease-specific characteristics and comorbidity association characteristics. This eliminates the interference of gender factors on disease risk prediction, especially for chronic diseases with significant gender differences such as depression, thereby improving prediction accuracy through gender-specific modeling.

[0043] In some optional embodiments, the step of performing hierarchical preprocessing on the health examination association data of multiple health examination subjects to generate a standardized dataset containing single disease labels and comorbidity combination labels may specifically include: The physical examination-related data is structured to obtain an initial dataset, which includes single-disease labels and comorbidity combination labels. Each chronic disease corresponds to a single-disease label, and each combination of chronic diseases corresponds to a comorbidity combination label. The initial dataset is then subjected to hierarchical cleaning to obtain a clean dataset. The categorical variables in the clean dataset undergo feature reconstruction. Key physical examination features are extracted from the processed dataset, including demographic features, routine examination indicators, specific examination indicators, and medical history-related features. The extracted data is then integrated into the standardized dataset.

[0044] In raw storage, physical examination-related data typically uses a vertical record structure, where each examination result for each examinee is stored as a separate record. For example, when a person undergoes multiple physical examinations, the multiple indicators from each examination are stored in multiple rows. Structured processing transforms this vertical record structure into a horizontal record structure, aggregating all features, labels, and time information from each examinee's single examination into a single record, thus obtaining the initial dataset. In this initial dataset, each chronic disease corresponds to a single-disease label, indicating the presence of that disease; each clinically common combination of chronic diseases corresponds to a comorbidity label, indicating the presence of that comorbidity state. Taking hypertension and diabetes as examples, the single-disease labels include hypertension and diabetes labels, and the comorbidity label includes hypertension combined with diabetes.

[0045] After obtaining the initial dataset, a tiered cleaning process is performed on it to address the high missing value rate commonly found in physical examination data. Tiered cleaning involves applying different missing value handling strategies based on the severity of the missing value rate. For example, suppose the initial dataset is... (After structural transformation, including single disease / comorbidity labels), for each physical examination feature, set the feature missing rate. If the missing rate of a certain feature is... If so, then delete the feature column directly. If the missing rate of a certain feature If so, then the feature is retained and filled in.

[0046] The imputation operation is performed according to priority: the first priority is imputation of adjacent data, that is, if the same physical examination subject has a previous physical examination record, the corresponding feature value from the previous physical examination is used to imput the current missing value, represented as... ,in, For the first Characteristic values ​​from the second physical examination For the first The feature values ​​of the next physical examination (if they exist). If there are no adjacent physical examination data, the KNN algorithm is used for imputation, that is, finding the k most similar samples to the current sample and using the mean of these samples on the same feature to imput the data, denoted as: ,in, For the most similar The feature values ​​of each sample are used to obtain a clean dataset after hierarchical cleaning. .

[0047] Feature reconstruction is performed on the categorical variables in the clean dataset to improve the accuracy of the interpretability analysis. Categorical variables are variables that take values ​​in a finite number of discrete categories, such as textual descriptions like "positive," "negative," and "weakly positive" in physical examination results. Feature reconstruction mainly involves converting categorical variables into numerical variables, for example, by using binarization mapping.

[0048] If a categorical variable takes a positive value (such as "positive" or "weak positive," indicating an abnormal state), it is mapped to 1; if it takes a negative value (such as "negative" or "normal," indicating a normal state), it is mapped to 0. For multi-category variables, one-hot encoding or other methods can be used to expand them into multiple binary features. Numerical variables need to be normalized to eliminate the influence of different units on model training.

[0049] Normalization can be achieved using the Min-Max standardization method, which linearly scales the value of each numerical variable to the [0,1] interval. The specific calculation formula is as follows:

[0050] in, Features The set of values ​​in all samples and They are respectively The maximum and minimum values.

[0051] After completing the above processing, key physical examination features can be extracted from the processed dataset. Key physical examination features refer to a subset of features that have significant value in predicting chronic diseases, specifically including demographic characteristics (such as age and gender), routine examination indicators (such as heart rate, blood pressure, and blood glucose), specific examination indicators (such as tumor markers and imaging features for specific diseases), and medical history-related features (such as past medical history and family history). After extracting these features, they are integrated with corresponding single-disease labels, comorbidity combination labels, time window information, and gender information to obtain a standardized dataset. This standardized dataset is split according to the "gender-single-comorbidity" dimension, and can be represented as follows: Where g represents gender grouping, and k represents the kth single disease, Indicates the first Comorbidity combination, number The set of specific biomarkers for a single disease is , No. The set of association biomarkers for comorbidity combinations is The standardized dataset after splitting can be adapted to the differentiated processing needs of different genders, different diseases, and different comorbidity combinations in the subsequent two-stage modeling.

[0052] In some alternative embodiments, the process of constructing a single-framework multi-objective collaborative ensemble model based on this standardized dataset may include: The standardized dataset is processed using a single-head self-attention mechanism with prior constraints to calculate first, second, and third association information. This first, second, and third association information are then fused with the key physical examination features to generate an enhanced feature set. Based on this enhanced feature set, a single-framework multi-objective collaborative ensemble model is constructed. In this embodiment, the single-head self-attention mechanism with prior constraints refers to an attention calculation method that, based on standard self-attention calculation, introduces additional clinical prior knowledge to regularize the attention weights. The calculation formula for the standard self-attention mechanism is:

[0053] In the formula, Q represents the feature query matrix, used to query key features of the physical examination; K represents the key matrix of disease or comorbidity tags, used to match with the query; and V represents the feature value matrix, used to output the weighted feature representation. The feature dimension is used to scale the dot product result to stabilize the gradient.

[0054] The above calculations yield an attention weight matrix. Based on this, prior clinical knowledge is introduced to regularize the weight matrix. For example, if the comorbidity correlation coefficient between hypertension and diabetes is known to be no less than 0.7, then the association weights between the corresponding features of these two diseases are forcibly preserved or enhanced, while spurious associations that do not conform to medical logic (such as feature pairs between blood pressure and mental illnesses with no clear pathological link). After the above constrained self-attention calculations, three types of association information are obtained: the first association information, i.e., feature-single-disease association information, is used to quantify the association strength of each key physical examination feature with a single chronic disease, i.e., the degree of association between a certain physiological indicator or demographic characteristic and a specific single disease; the second association information, i.e., feature-comorbidity combination association information, is used to quantify the association strength of each key physical examination feature with common clinical comorbidity combinations, i.e., the degree of association between a certain feature and the simultaneous occurrence of two or more diseases; and the third association information, i.e., feature-comorbidity combination association information, is used to quantify the risk synergy effect between two or more chronic diseases, i.e., the mutual enhancement or weakening relationship between diseases in terms of disease incidence risk.

[0055] The calculated first, second, and third association information are fused with the original key physical examination features to generate an enhanced feature set. In this embodiment, fusion refers to concatenating the original feature matrix with the above three types of association information matrices column-wise or adding them by weight to form a new, higher-dimensional feature representation. For example, the enhanced feature set can be represented as:

[0056] in, This is the original physical examination key feature matrix. This is a feature-single-disease association information matrix. This is a feature-comorbidity combination association information matrix. This is the disease-disease collaborative weight matrix. The above three types of related information together constitute the attention-related information set. .

[0057] The calculated first, second, and third association information are fused with the original key physical examination features to generate an enhanced feature set. The fusion operation involves concatenating the original feature matrix with the three association information matrices column-by-column to form a new high-dimensional feature representation. Specifically, the enhanced feature set... The number of key physical examination features is m, the number of single disease types is K, the number of comorbidity combinations is L, and the total number of samples is N.

[0058] After generating an enhanced feature set through attention association mining, a single-frame multi-objective collaborative ensemble model is constructed based on this enhanced feature set, and its sensitivity is systematically analyzed and optimized. The model has a built-in dual prediction dimension. ,in Represents "the Gender - Number Single disease - No. Comorbidity combination - number The integrated prediction model of "time window", all They share the same underlying feature representation layer and computation unit, rather than being trained independently.

[0059] During training, an integrated training strategy is adopted:

[0060] in, To accommodate a hybrid algorithm set with two prediction dimensions, this set includes five basic learners. The corresponding algorithms are LightGBM, Multilayer Perceptron, Random Forest, Stochastic Gradient Boosting, and Logistic Regression, which are dynamically selected or combined based on the different characteristics of single-disease prediction and comorbidity prediction. During training, a global loss function is used to collaboratively optimize the single-disease prediction module and the comorbidity prediction module.

[0061] In this embodiment, the single-framework multi-objective collaborative ensemble model is a single, indivisible ensemble machine learning model constructed as an ensemble framework with a unified structure. This framework includes a single-disease prediction module and a comorbidity prediction module, both reusing the feature representation layer and computational units within the same ensemble framework. The single-disease prediction module outputs the risk probability of each chronic disease within each preset time window, while the comorbidity prediction module outputs the risk probability of each comorbidity combination within each preset time window. The two modules are collaboratively optimized using a global loss function, ensuring that the feature weights learned during single-disease prediction aid in the association mining for comorbidity prediction, while the disease synergistic effects identified during comorbidity prediction, in turn, optimize the feature selection for single-disease prediction. After model training, it can receive preprocessed data from any test subject and simultaneously output the predicted disease values ​​for all single diseases and all comorbidities within different time windows through a single forward propagation.

[0062] Specifically, the single-disease prediction dimension and the comorbidity prediction dimension are embedded as built-in modules within the same integrated framework. Firstly, the single-disease prediction module and the comorbidity prediction module share the enhanced feature space; that is, both use the enhanced feature set. As input, this enhanced feature set integrates original physical examination features, feature-single disease association information, feature-comorbidity combination association information, and disease-disease synergy information. Secondly, the single-disease prediction module and the comorbidity prediction module share the core computation layer; that is, the feature representation layer, attention computation unit, and intermediate hidden layers of the basic learner in the model are all reused by both modules, and there are no independent computation paths. Thirdly, the single-disease prediction module and the comorbidity prediction module share parameter optimization logic; that is, the model parameters of both modules are updated during training using the same optimizer and the same learning rate scheduling strategy. Only at the output layer, specific mapping is used to adapt to different prediction targets. For example, a sigmoid activation function is used to output K probability values ​​for the single-disease output dimension, and another set of sigmoid functions is used to output L probability values ​​for the comorbidity output dimension. This allows the model to obtain the prediction results of all single diseases and all comorbidities simultaneously through a single forward propagation during inference.

[0063] To address the different functional requirements of the two prediction dimensions, this application employs a multi-algorithm integration strategy within a unified collaborative integration architecture. The single-disease prediction dimension focuses on "biomarker sensitivity," requiring accurate capture of the independent impact of a single physiological indicator or demographic characteristic on a specific chronic disease. Therefore, the LightGBM algorithm is chosen as the base learner, as it excels in handling high-dimensional tabular data and categorical features. The comorbidity prediction dimension focuses on "synergistic effect capture," requiring the quantification of the risk superposition or synergistic enhancement effect when two or more diseases occur simultaneously. Therefore, a fusion algorithm combining LightGBM and stochastic gradient boosting is chosen as the base learner. This complementary approach of two fusion learners with different principles enhances the ability to model complex interactive relationships. All base learners are embedded within the same unified collaborative integration architecture and do not operate independently but are dynamically correlated through integration weights. These integration weights are adaptively adjusted during training. For example, the LightGBM output of the single-disease prediction module and the fusion algorithm output of the comorbidity prediction module are weighted and fused using a learnable weight matrix, allowing information from both modules to flow between them.

[0064] Furthermore, to achieve mutual benefit between single-disease prediction and comorbidity prediction, this application designs a global loss function for collaborative training of the two modules. This global loss function is:

[0065] In the formula, This is a balance coefficient, with a value ranging from 0.5 to 0.6. This is the sum of the cross-entropy losses for all single-disease prediction dimensions. This represents the sum of the cross-entropy losses across all comorbidity prediction dimensions. During training, the model optimizes the prediction errors for both single-disease and comorbidity predictions simultaneously through backpropagation. Specifically, the optimized feature weights of the single-disease prediction dimension are fed back to the comorbidity prediction dimension via the shared underlying feature space and core computation layer, helping it to more accurately uncover synergistic information between diseases. Conversely, the synergistic effect weights identified by the comorbidity prediction dimension through disease association analysis are also passed to the single-disease prediction dimension via shared parameters, optimizing the single-disease prediction module's feature selection for different disease-specific biomarkers. This bidirectional feedback mechanism enables mutual gains between the two dimensions, making the model overall superior to the scheme of training single-disease and comorbidity models separately and then simply stacking them.

[0066] For the aforementioned collaborative ensemble model as a whole, this application performs sensitivity analysis on the collaborative ensemble model, rather than analyzing each individual module separately. The sensitivity analysis includes adjusting the input feature dimensions (e.g., increasing or decreasing the number of features in the enhanced feature set by ±20%), adjusting the ensemble weights (e.g., changing the fusion weights between the base learners by ±15%), and simulating different data distribution scenarios (e.g., introducing 10% random noise into the training data). Through these adjustments, the model's stability is evaluated across three dimensions: single-disease prediction accuracy (using AUC as an indicator), comorbidity prediction accuracy (using AUC as an indicator), and overall inference efficiency (using average time or throughput per inference as an indicator). Only model parameter configurations with performance changes ≤5% are retained.

[0067] After model training is complete, sensitivity analysis is performed to evaluate the model's robustness and provide adapted outputs for different application scenarios. In this embodiment, sensitivity analysis is performed by adjusting the amount of change in the number of input variables. Feature weight adjustment and data distribution fluctuation To test the stability of the model's performance.

[0068] Let the first The prediction scenario sensitivity analysis function is The model performance was evaluated using the area under the curve (AUC), and the stability threshold was set to... (Performance changes ≤5% are considered stable), the sensitivity analysis logic formula includes: Impact of the number of variables: ; Impact of feature weights: ; Impact of data distribution: ; in, This refers to the performance changes of the model under different scenarios. Through the above sensitivity analysis, a subset of core features can be selected for each prediction scenario. This refers to the set of key features that have the greatest impact on model performance when the input variables change.

[0069] Based on the sensitivity analysis results, the model supports a dual-track output design, with both tracks sharing the same integration framework. The dual-track output includes a full-track mode and a lightweight mode. In full-track mode, the single-framework multi-objective collaborative integration model performs a single inference based on all features of the enhanced feature set, outputting single-disease prediction values ​​and comorbidity prediction values. That is, the model calls upon the complete enhanced feature set. Perform predictions and output high-precision single-disease prediction probabilities. Comorbidity prediction probability It can be applied to scenarios with high precision requirements, such as top-tier hospitals.

[0070] In the lightweight mode, the single-framework multi-objective collaborative ensemble model performs a single inference based on a subset of core features, outputting single-disease and comorbidity prediction values. That is, the model only calls upon a subset of core features. Make predictions and output the single-disease prediction probability for rapid response. Comorbidity prediction probability It can be applied to scenarios with limited computing power or requiring rapid response, such as community hospitals and health management institutions.

[0071] The core feature subset was determined by performing sensitivity analysis on the single-frame multi-objective collaborative integration model and optimizing the result based on the number of variables.

[0072] The two modes can be switched based on application scenario requirements without retraining or reconstructing the model. Both output modes share the same integrated architecture, switching only through the "feature input dimension"—the full mode calls all enhanced features (e.g., 158 dimensions), while the lightweight mode calls the core feature subset selected by sensitivity analysis (e.g., 30 dimensions), without needing to reconstruct the model structure, adapting to different computing power scenarios.

[0073] In some alternative embodiments, based on the enhanced feature set and integrated framework, the embodiments of this application have made differentiated functional designs for the single disease prediction module and the comorbidity prediction module to adapt to the different prediction characteristics of chronic diseases and comorbidities.

[0074] The single-disease prediction module is used to adapt risk prediction for each chronic disease within each preset time window. For example, for an assessment scenario of "K single diseases × 4 preset time windows (1 year, 3 years, 5 years, 10 years)," this module incorporates single-disease-specific prediction logic. Because the pathophysiological mechanisms of different chronic diseases vary significantly, their corresponding physical examination biomarkers also differ—for example, the core predictive biomarkers for diabetes are fasting blood glucose and glycated hemoglobin, while the prediction of depression relies more on neurotransmitter-related indicators or psychological scale scores. To address these differences, the single-disease prediction module, based on reusing the underlying feature representation layer, outputs the first association information (i.e., the feature-single-disease association information matrix) through an attention mechanism. The model dynamically assigns feature weights to different diseases. During model training and inference, when predicting diabetes, the model automatically increases the weight coefficients of metabolic-related features such as fasting blood glucose and postprandial blood glucose, while relatively decreasing the weights of other irrelevant features. When predicting depression, the model dynamically adjusts the weights to highlight mental health-related features such as sleep quality and anxiety scores. This dynamic adjustment mechanism allows for accurate prediction of different chronic diseases without splitting the same model framework into multiple independent models.

[0075] The comorbidity prediction module is used to adapt the risk prediction of each comorbidity combination under each preset time window, and directly calls the third association information to calculate the comorbidity risk. Unlike the traditional technical solution of "first predicting the risk of each individual disease separately, and then simply superimposing them to determine whether comorbidity exists," the comorbidity prediction module in this embodiment incorporates disease synergy effect quantification logic. Specifically, when constructing the single-framework multi-objective synergistic integration model, the third association information (i.e., the disease-disease synergistic weight matrix) has already been calculated in the first stage through an attention mechanism with clinical prior constraints. Each element in this matrix quantifies the risk synergy between two chronic diseases; for example, the synergy weight between hypertension and diabetes is 0.75, and the synergy weight between depression and hyperlipidemia is 0.68. When performing forward computation, the comorbidity prediction module directly calls this third association information from the enhanced feature set as part of the input features, enabling the model to directly learn the rule that "when disease A and disease B coexist, their combined risk is not a simple linear addition of their individual risks, but rather exhibits a synergistic enhancement effect." Taking the comorbidity combination of "hypertension + diabetes" as an example, the output probability of the comorbidity prediction module depends not only on the original physical examination features such as fasting blood glucose and blood pressure, but also directly on the synergistic weight value of hypertension and diabetes in the third association information, thereby achieving a quantitative assessment of the comorbidity risk.

[0076] In some optional embodiments, the step of calculating a comorbidity risk gain value based on the single-disease morbidity prediction value and the comorbidity morbidity prediction value, and generating a structured assessment result based on the comorbidity risk gain value, may specifically include: Obtain the first risk threshold set for each single disease under each preset time window, and the second risk threshold set for each comorbidity combination.

[0077] The single disease incidence prediction value is compared with the first risk threshold set to determine the single disease risk level to which the single disease incidence prediction value belongs; the comorbidity incidence prediction value is compared with the second risk threshold set to determine the comorbidity risk level to which the comorbidity incidence prediction value belongs. Based on the predicted comorbidity values ​​and the predicted individual disease values ​​constituting the comorbidity combination, a comorbidity risk gain value is calculated; wherein, if the comorbidity combination consists of two individual diseases, the comorbidity risk gain value is equal to the predicted comorbidity value minus the union probability of the predicted individual disease values; if the comorbidity combination consists of three or more individual diseases, the comorbidity risk gain value is calculated based on the difference between the predicted comorbidity value and the union probability of the predicted individual disease values. The output is a structured assessment result. Each record in the structured assessment result contains at least a single disease type identifier, a preset time window identifier, the single disease risk level, the single disease incidence prediction value, and a comorbidity combination type identifier, the comorbidity risk level, the comorbidity incidence prediction value, and the comorbidity risk gain value.

[0078] Suppose we are considering a specific prediction scenario, which consists of gender g (e.g., 1 represents male, 2 represents female), the kth single disease, the lth comorbidity combination, and the i-th preset time window. This is jointly determined. In this scenario, the single-disease incidence prediction value output by the single-frame multi-objective collaborative ensemble model is... This value is a probability value between 0 and 1, representing the likelihood of an individual developing the disease within a given gender, specific disease type, and specific time window. The output also includes a predicted comorbidity rate. It is also a probability value between 0 and 1, representing the probability that a person undergoing a physical examination will have multiple diseases at the same time under a given gender, a specific combination of comorbidities, and a specific time window.

[0079] To convert the aforementioned continuous probability values ​​into clinically relevant risk levels, it is necessary to pre-define a set of risk thresholds for each single disease and each comorbidity combination within each time window. For example, for the k-th single disease in the... The set of risk thresholds under a time window is defined as follows ;in, The threshold for low risk This is the threshold for dividing medium risk. This is the high-risk threshold. All of them. This constitutes the first risk threshold set.

[0080] Similarly, for the lth comorbidity combination in the first... The set of risk thresholds under a time window is defined as follows ,in The threshold for low risk This is the threshold for dividing medium risk. This is the high-risk threshold. All of them. This constitutes the second set of risk thresholds. These thresholds can be pre-set based on clinical guidelines, historical data distribution, or expert experience.

[0081] The risk level is classified as follows:

[0082]

[0083] In the above formula, the predicted incidence value of a single disease is compared with... Compare with the corresponding risk threshold set, if If it is low risk (R=0), then it is considered low risk; if If so, it is classified as medium risk (R=1); if If the value is 2, it is considered high risk (R=2).

[0084] Similarly, the predicted comorbidity rate will be... The risk level is compared with the corresponding risk threshold set and classified as low risk, medium risk, and high risk.

[0085] Furthermore, in addition to the risk levels of single diseases and comorbidities, this application also introduces a quantitative indicator of comorbidity risk gain value. This is used to characterize the additional risk resulting from the synergistic effect of diseases. If the comorbidity combination consists of two single diseases, such as hypertension combined with diabetes, let these two single diseases be... and The formula for calculating the comorbidity risk gain value is:

[0086] The expression within parentheses represents the union probability of the predicted values ​​for the two individual diseases, i.e., the theoretical probability of having at least one of the diseases. Subtracting this union probability from the predicted value for comorbidity gives the difference that exceeds the linear superposition of independent risks when both diseases are present. The larger this difference, the more significant the synergistic enhancement effect between the diseases. If the comorbidity combination consists of three or more individual diseases, the union probability of all individual disease predicted values ​​is calculated, and then this union probability is subtracted from the predicted value for comorbidity; the difference is the comorbidity risk gain value.

[0087] After completing risk grading and gain calculation, the model outputs structured evaluation results:

[0088] Each result record includes the following fields: single disease type identifier (i.e., which specific chronic disease it is), preset time window identifier (1 year, 3 years, 5 years or 10 years), single disease risk level, single disease prevalence prediction value, and comorbidity combination type identifier (i.e., which two or more diseases are combined), comorbidity risk level, comorbidity prevalence prediction value, and comorbidity risk gain value.

[0089] For example, a specific output could be:

[0090] The results indicate that for female participants (g=2), regarding diabetes (the second single disease) and the comorbid combination of hypertension and diabetes (the first comorbid combination), the predicted probability of diabetes within a 3-year time window is 45%, which is classified as medium risk (level 1); the predicted probability of comorbid hypertension and diabetes is 68%, which is classified as high risk (level 2); the comorbidity risk gain is 13%, which means that the additional synergistic risk of the two diseases coexisting is 13 percentage points.

[0091] In some optional embodiments, the method provided in this application may further include the step of generating an evaluation image to assist in interpreting the evaluation results of the model.

[0092] The method used in this embodiment combines attention-related information with model interpretability algorithms to generate assessment images of key features of the physical examination, and outputs a risk correlation report between single diseases and comorbidities to assist in interpreting the joint risk assessment results. The steps may include: Based on the first, second, and third association information, a first visual representation of the association strength between the key physical examination features and individual diseases, a second visual representation of the association strength between the key physical examination features and comorbidity combinations, and a ranking table of synergistic effects among diseases are generated. The first visual representation displays the association strength distribution of each key physical examination feature to each individual disease, the second visual representation displays the association strength distribution of each key physical examination feature to each comorbidity combination, and the ranking table displays the order of synergistic effects between different disease combinations. A preset algorithm is used to calculate the comprehensive contribution of each key physical examination feature to the predicted values ​​of the individual disease and the comorbidity, and a third visual representation of the contribution of each key physical examination feature to the prediction results is generated based on this comprehensive contribution. The third visual representation distinguishes between the positive and negative impacts of each key physical examination feature on the prediction results. The first, second, and ranking tables are integrated to output a report on the independent risk association degree of each disease and a report on the synergistic risk association degree of each key physical examination feature.

[0093] Based on the first, second, and third association information calculated using the attention mechanism, three types of visualizations are generated. The first association information is used to quantify the association strength of each key health check feature with a single chronic disease. Based on this, a first visualization representation of the association strength between the key health check feature and the single disease is generated. This representation is used to show the distribution of the association strength of each key health check feature with each single disease. For example, the first visualization representation can be a heatmap, with the horizontal axis representing different single diseases (hypertension, diabetes, depression, etc.) and the vertical axis representing various key health check features (fasting blood glucose, blood pressure, heart rate, etc.). The color intensity of each cell represents the strength of the association between that feature and the corresponding single disease.

[0094] The second association information is used to quantify the association strength of each key physical examination feature with common clinical comorbidity combinations. Based on this, a second visual representation of the association strength between the key physical examination feature and the comorbidity combination is generated. This representation is used to show the distribution of the association strength of each key physical examination feature with each comorbidity combination. For example, a heatmap is used to show that the association strength between fasting blood glucose and the comorbidity combination of "hypertension + diabetes" is 0.88, and the association strength between fasting blood glucose and the comorbidity combination of "depression + hyperlipidemia" is 0.32.

[0095] The third association information is used to quantify the risk synergy between two or more chronic diseases. Based on this, a ranking table of the synergy between diseases is generated. This ranking table is used to show the order of the strength of the synergy between different disease combinations. For example, “hypertension + diabetes (0.75)”, “depression + hyperlipidemia (0.68)”, “hypertension + coronary heart disease (0.52)”, etc. are listed in descending order of synergy weight, which intuitively show which disease combinations have significant synergistic risks.

[0096] A pre-defined interpretability algorithm is used to calculate the comprehensive contribution of each key physical examination feature to the predicted values ​​of single-disease and comorbidity conditions. Based on this comprehensive contribution, a third visual representation of the contribution of each key physical examination feature to the prediction results is generated. In this embodiment, the pre-defined algorithm can be either the SHAP algorithm or the LIME algorithm. Taking the SHAP algorithm as an example, for a specific prediction scenario (e.g., female, diabetes, hypertension + diabetes comorbidity, 3-year time window), the SHAP contribution value of each enhanced feature (including the original key physical examination features and the first, second, and third related information features generated by the attention mechanism) is calculated. and A positive contribution indicates that the characteristic increases the risk of disease, while a negative contribution indicates that the risk of disease decreases.

[0097] Based on the contribution calculation results, a feature contribution waterfall plot is generated. This waterfall plot displays the marginal contribution of each feature to the prediction result from left to right. The length of the bar corresponding to each feature represents its absolute contribution value, and the direction of the bar (left or right) represents a positive or negative impact. For example, for single-disease prediction of diabetes, the SHAP value of fasting blood glucose is +0.35 (the largest positive contribution), while the SHAP value of regular exercise is -0.12 (a negative contribution). The first, second, and third visualizations mentioned above, along with the ranking table and waterfall plot, constitute various specific forms of the third visualization.

[0098] Integrating the first, second, and ranking tables, as well as the third visualization, the system outputs a single-disease independent risk association report and a comorbidity synergistic risk association report for each key physical examination feature. Specifically, the single-disease independent risk association report, based on the first association information and the SHAP contribution results, provides the independent risk association of each feature with respect to a specific single disease. (ratio ratio, This indicates an increased risk. (This indicates a reduction in risk). For example, "The independent risk association between fasting blood glucose and diabetes is 3.2, meaning that for every unit increase in fasting blood glucose, the risk of developing diabetes increases by 2.2 times."

[0099] The comorbidity risk correlation report, based on secondary and tertiary correlation information and SHAP contribution results, provides the comorbidity risk correlation of each feature with respect to a specific comorbidity combination. For example, "the synergistic risk association between fasting blood glucose and hypertension + diabetes comorbidity is 4.1, and the disease synergistic weight of this comorbidity combination is 0.75, suggesting that there is a significant synergistic enhancement effect between the two diseases."

[0100] In the above implementation process, a single-framework multi-objective collaborative integration model is constructed. Within this model, the single-disease prediction module and the comorbidity prediction module reuse the same feature representation layer and computational unit, allowing them to share the underlying feature space. Simultaneously, a global loss function is used to collaboratively optimize the two modules, enabling the adjustment of feature weights in single-disease prediction and the mining of association information in comorbidity prediction to mutually enhance each other, thereby improving the model's ability to model complex disease relationships. Simultaneous output of single-disease and comorbidity prediction values ​​in a single inference reduces the accumulation of errors from multiple inferences. Calculating the comorbidity risk gain value quantifies the additional risk contributed by disease synergy, allowing the structured assessment results to more comprehensively reflect the true disease risk, thus improving the accuracy of joint risk assessment for chronic diseases and comorbidities.

[0101] The following is an example illustrating the application of the chronic disease and comorbidity risk assessment method provided in this application in practice, specifically for the long-term risk prediction of four single diseases (hypertension, diabetes, depression, and hyperlipidemia) and three comorbidity combinations (hypertension + diabetes, depression + hyperlipidemia, and hypertension + coronary heart disease).

[0102] The data sources for the physical examination-related data include data from community hospitals and physical examination centers, encompassing 1183 initial characteristics such as CT scans, MRI scans, blood tests, and internal medicine examinations. This data covers demographic information (gender, age), examination indicators (heart rate, blood glucose, tumor markers, etc.), single-disease diagnosis results, comorbidity diagnosis results, diagnosis time, and examination time.

[0103] The samples were then labeled. Patients diagnosed with hypertension, diabetes, depression, or hyperlipidemia were selected from the diagnostic records and marked as positive samples for the corresponding single disease. The time difference between the physical examination time and the diagnosis time was calculated and matched with the corresponding preset time window. Patients diagnosed with hypertension + diabetes, depression + hyperlipidemia, or hypertension + coronary heart disease were selected from the diagnostic records and marked as positive samples for the corresponding comorbidities. The time difference between the physical examination time and the first diagnosis time of the comorbidity was calculated and matched with the corresponding preset time window. Healthy individuals with no history of disease and no diagnostic records were selected and marked as negative samples. The time difference between the physical examination time and the virtual diagnosis time was set to 99 years (maximum value) to ensure that they were judged to be healthy in all time windows.

[0104] The samples were divided into male and female groups based on gender. The male group retained male-specific characteristics such as prostate examination, while the female group retained female-specific characteristics such as breast examination. Single-disease positive samples and comorbid-positive samples were segmented according to matched time windows (1 year, 3 years, 5 years, and 10 years), and negative samples were assigned to all time windows. Finally, the physical examination correlation data under different genders and different time windows were obtained.

[0105] The vertical, individual examination records (one line per person per examination) were transformed into a horizontal, comprehensive record of a single physical examination. The original 1.63 million records were consolidated into over 100,000 structured records, including single-disease tags and comorbidity combination tags. The handling of tiered missing values ​​included: 1. Remove feature columns with a missing rate > 50%, reducing the initial 1183 features to 140; 2. For data with a missing rate of ≤30% in the remaining features, priority should be given to filling the missing data with the corresponding values ​​from the adjacent physical examination records of the examinee (e.g., if the blood glucose data is missing in this examination, the blood glucose value from the previous physical examination should be used to fill the missing data). 3. For data that was not successfully filled, the KNN algorithm was used to fill in the missing data to obtain a clean dataset.

[0106] The steps of feature standardization include: 1. Categorical variable processing: Convert text results such as “weak positive”, “negative”, and “positive” to binary classification (positive = 1, negative = 0); 2. Numerical variable processing: Units are standardized and normalized for indicators such as heart rate and blood glucose to eliminate the influence of dimensions; 3. Feature selection: 129 key physical examination features were ultimately retained and integrated to obtain a standardized dataset.

[0107] The construction of a single-framework multi-objective collaborative ensemble model based on a standardized dataset consists of two stages. The first stage includes attention association mining and enhanced feature generation. A single-head self-attention mechanism with clinical prior constraints is adopted to calculate three types of association information: feature-single disease association information, such as the association information between fasting blood glucose and diabetes (0.85) and hypertension (0.72); feature-comorbidity combination association information, such as the association information between fasting blood glucose and hypertension + diabetes (0.88); and disease-disease synergistic weights, such as the synergistic weight between hypertension and diabetes (0.75) and depression and hyperlipidemia (0.68). The three types of association information are fused with 129 original features to generate an enhanced feature set containing 129+4+3+16=152 features (4 for the number of single diseases, 3 for the number of comorbidity combinations, and 16 for the number of disease-disease synergistic weights). The second stage includes training a multi-objective collaborative ensemble model. For the assessment scenario of "4 single diseases × 3 comorbidity combinations × 4 time windows," a dedicated mapping rule is constructed within the dual prediction dimensions, eliminating the need for separate independent training and achieving full scenario coverage. To address the different scenario adaptation needs of the dual prediction dimensions, a hybrid algorithm set (LightGBM, multilayer perceptron, etc.) is embedded into a unified architecture, and scenario adaptation is achieved through dynamic allocation of integrated weights. Sample sampling (handling class imbalance), hyperparameter tuning (Bayesian optimization), and 10-fold cross-validation are employed to ensure model stability. The input variable dimensions (±20% of features), feature weights (±20%), and data distribution (introducing 10% noise) are adjusted to select stable models with performance changes ≤5%. After collaborative training based on a global loss function, a single-framework multi-objective collaborative ensemble model is formed, providing both full and lightweight modes.

[0108] Afterwards, risk prediction and classification are performed. The physical examination data of the subjects to be tested are preprocessed according to the method in step 2 and then input into the single-frame multi-objective collaborative integration model. The model outputs 48 disease prediction values ​​(each prediction corresponds to a single disease probability and a comorbidity probability). The preset risk thresholds are set (low risk ≤30%, medium risk 30%-70%, high risk ≥70%). The predicted values ​​are compared with the thresholds, the comorbidity risk gain value is calculated, and the structured assessment results are generated (such as "hypertension - 3 years - medium risk (45%)" "hypertension + diabetes - 3 years - high risk (68%) - gain value 23%)".

[0109] After obtaining the structured risk assessment results, a feature-single disease / comorbidity association heatmap can be generated, showing the strong association between fasting blood glucose and diabetes, and hypertension + diabetes comorbidity. The SHAP algorithm is then used to calculate the contribution of enhancing features to each prediction result; for example, the SHAP value for "fasting blood glucose × diabetes association information" reaches 0.35 (ranked first). Based on the assessment image, a feature impact report is output, clearly stating that "fasting blood glucose is a core risk factor for diabetes as a single disease and hypertension + diabetes comorbidity; controlling fasting blood glucose can simultaneously reduce the risk of both single disease and comorbidity," helping doctors develop targeted intervention measures.

[0110] Based on the same concept, this application also provides a device for assessing the combined risk of chronic diseases and comorbidities, please refer to [reference needed]. Figure 2 , Figure 2 This is a schematic diagram of the architecture of the joint risk assessment device for chronic diseases and comorbidities provided in the embodiments of this application.

[0111] The chronic disease and comorbidity joint risk assessment device 20 may include: The first construction module 21 is used to perform hierarchical preprocessing on the physical examination association data of multiple physical examination subjects to generate a standardized dataset containing single disease labels and comorbidity combination labels; the physical examination association data includes chronic disease diagnosis data and comorbidity diagnosis data; The second construction module 22 is used to construct a single-framework multi-objective collaborative ensemble model based on the standardized dataset. The single-framework multi-objective collaborative ensemble model is an integrated ensemble framework, which includes a single-disease prediction module and a comorbidity prediction module. The single-disease prediction module and the comorbidity prediction module reuse the feature representation layer and computing unit of the integrated ensemble framework. The single-disease prediction module and the comorbidity prediction module are collaboratively optimized through a global loss function. Prediction module 23 inputs target data into the single-frame multi-objective collaborative integration model to obtain single-disease incidence prediction values ​​and comorbidity incidence prediction values; the single-disease incidence prediction values ​​and the comorbidity incidence prediction values ​​are obtained synchronously by the single-frame multi-objective collaborative integration model through a single inference. The result generation module 24 is used to calculate the comorbidity risk gain value based on the single disease prevalence prediction value and the comorbidity prevalence prediction value, and generate a structured assessment result based on the comorbidity risk gain value.

[0112] In some embodiments, the result generation module 24 can also be used for: Based on the first association information, the second association information, and the third association information, a first visual representation of the association strength between the key physical examination features and individual diseases, a second visual representation of the association strength between the key physical examination features and comorbidity combinations, and a ranking table of synergistic effects among diseases are generated; wherein, the first visual representation is used to display the association strength distribution of each key physical examination feature to each individual disease, the second visual representation is used to display the association strength distribution of each key physical examination feature to each comorbidity combination, and the ranking table is used to display the order of the strength of synergistic effects among different disease combinations; The comprehensive contribution of each of the key physical examination features to the predicted values ​​of the single disease and the comorbidity is calculated based on a preset algorithm, and a third visual representation of the contribution of each of the key physical examination features to the prediction results is generated based on the comprehensive contribution. The third visual representation is used to distinguish the positive or negative impact of each of the key physical examination features on the prediction results. By integrating the first visualization representation, the second visualization representation, the sorting table, and the third visualization representation, a single-disease independent risk correlation report and a comorbidity synergistic risk correlation report corresponding to each of the key physical examination features are output.

[0113] It should be understood that when the various modules of the system provided in the above embodiments are working, the division of each functional module in the above description is only used as an example. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0114] The functional modules in the above embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of the embodiments of this application.

[0115] Based on the same concept, embodiments of this application also provide a computer device, which may include a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the methods described above.

[0116] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for assessing the combined risk of chronic diseases and comorbidities, characterized in that, include: Hierarchical preprocessing was performed on the physical examination data of multiple subjects to generate a standardized dataset containing single disease labels and comorbidity combination labels; The physical examination-related data includes chronic disease diagnosis data and comorbidity diagnosis data; A single-framework multi-objective collaborative ensemble model is constructed based on the standardized dataset. The single-framework multi-objective collaborative ensemble model is an integrated ensemble framework, which includes a single-disease prediction module and a comorbidity prediction module. The single-disease prediction module and the comorbidity prediction module reuse the feature representation layer and computation unit of the integrated ensemble framework. The single-disease prediction module and the comorbidity prediction module are collaboratively optimized through a global loss function. The target data is input into the single-frame multi-objective collaborative integration model to obtain the single-disease incidence prediction value and the comorbidity incidence prediction value; the single-disease incidence prediction value and the comorbidity incidence prediction value are obtained by the single-frame multi-objective collaborative integration model through a single inference synchronization. The comorbidity risk gain value is calculated based on the single disease prevalence prediction value and the comorbidity prevalence prediction value, and a structured assessment result is generated based on the comorbidity risk gain value.

2. The method for assessing the combined risk of chronic diseases and comorbidities according to claim 1, characterized in that, The construction of a single-framework multi-objective collaborative ensemble model based on the standardized dataset includes: The standardized dataset is processed using a single-head self-attention mechanism with prior constraints to calculate first association information, second association information, and third association information. The first association information is used to quantify the association strength of each key physical examination feature with a single chronic disease, the second association information is used to quantify the association strength of each key physical examination feature with common clinical comorbidity combinations, and the third association information is used to quantify the risk synergistic effect between two or more chronic diseases. The first association information, the second association information, the third association information, and the key features of the physical examination are fused together to generate an enhanced feature set; The single-framework multi-objective collaborative integration model is constructed based on the enhanced feature set.

3. The method for assessing the combined risk of chronic diseases and comorbidities according to claim 2, characterized in that, The single-disease prediction module is used to adapt the risk prediction of each chronic disease under each preset time window, and dynamically adjust the feature weights according to the differences in biomarkers of different chronic diseases. The comorbidity prediction module is used to adapt the risk prediction of each comorbidity combination under each preset time window, and call the third association information to calculate the comorbidity risk.

4. The method for assessing the combined risk of chronic diseases and comorbidities according to claim 2, characterized in that, The output modes of the single-frame multi-objective collaborative integration model include full mode and lightweight mode; In the full-scale mode, the single-frame multi-objective collaborative integration model performs a single inference based on all features of the enhanced feature set, and outputs single-disease morbidity prediction values ​​and comorbidity prediction values. In the lightweight mode, the single-frame multi-objective collaborative integration model performs a single inference based on a core feature subset, outputting single-disease prediction values ​​and comorbidity prediction values; the core feature subset is determined by performing sensitivity analysis on the single-frame multi-objective collaborative integration model and optimizing the number of variables.

5. The method for assessing the combined risk of chronic diseases and comorbidities according to claim 1, characterized in that, The step involves hierarchical preprocessing of the health examination association data of multiple examinees to generate a standardized dataset containing single-disease labels and comorbidity combination labels, including: Based on the disease diagnosis results in the physical examination-related data, determine whether the physical examination subject is a patient with a chronic disease or a person with a comorbid disease; If the examinee is a person suffering from a single chronic disease, the examinee's examination-related data is marked as a single-disease positive sample. The time difference between the examinee's examination time and the time of disease diagnosis is calculated, and a corresponding preset time window is matched for the single-disease positive sample based on the time difference. If the examinee is a comorbidity group, the examinee's examination-related data is marked as a comorbidity positive sample. The time difference between the examination time and the first diagnosis time of the comorbidity is calculated, and a corresponding preset time window is matched for the comorbidity positive sample based on the time difference. If the examinee is a healthy person without a diagnosis record, the examinee's examination-related data will be marked as a negative sample, and the time difference between the examination time and the virtual diagnosis time will be set to a preset maximum value. The single-disease positive samples and the comorbid-disease positive samples are segmented according to their respective matched preset time windows, and the negative samples are allocated according to all preset time windows to obtain physical examination correlation data under different time windows; All physical examination-related data are divided into male and female groups according to gender. Male-specific physical examination characteristics are retained in the male group, and female-specific physical examination characteristics are retained in the female group.

6. The method for joint risk assessment of chronic diseases and comorbidities according to claim 1, characterized in that, The step involves hierarchical preprocessing of the health examination association data of multiple examinees to generate a standardized dataset containing single-disease labels and comorbidity combination labels, including: The physical examination-related data is structured to obtain an initial dataset; the initial dataset includes single disease labels and comorbidity combination labels, wherein each chronic disease corresponds to a single disease label and each combination of chronic diseases corresponds to a comorbidity combination label; The initial dataset is subjected to hierarchical cleaning to obtain a clean dataset; Feature reconstruction processing is performed on the categorical variables in the clean dataset; Key physical examination features are extracted from the processed dataset. These key physical examination features include demographic features, routine examination indicators, specific examination indicators, and medical history-related features. The extracted data are then integrated into the standardized dataset.

7. The method for assessing the combined risk of chronic diseases and comorbidities according to claim 1, characterized in that, The step of calculating the comorbidity risk gain value based on the single-disease prevalence prediction value and the comorbidity prevalence prediction value, and generating a structured assessment result based on the comorbidity risk gain value, includes: Obtain the first risk threshold set corresponding to each single disease under each preset time window, and the second risk threshold set corresponding to each combination of comorbidities; The single disease incidence prediction value is compared with the first risk threshold set to determine the single disease risk level to which the single disease incidence prediction value belongs. The predicted comorbidity level is determined by comparing the predicted comorbidity level with the second set of risk thresholds. Based on the predicted comorbidity values ​​and the predicted individual disease values ​​constituting the comorbidity combination, a comorbidity risk gain value is calculated; wherein, if the comorbidity combination consists of two individual diseases, the comorbidity risk gain value is equal to the predicted comorbidity value minus the union probability of the predicted individual disease values; if the comorbidity combination consists of three or more individual diseases, the comorbidity risk gain value is calculated based on the difference between the predicted comorbidity value and the union probability of the predicted individual disease values. The output is a structured assessment result. Each record in the structured assessment result contains at least a single disease type identifier, a preset time window identifier, the single disease risk level, the single disease incidence prediction value, and a comorbidity combination type identifier, the comorbidity risk level, the comorbidity incidence prediction value, and the comorbidity risk gain value.

8. The method for assessing the combined risk of chronic diseases and comorbidities according to any one of claims 1-7, characterized in that, The method further includes: Based on the first association information, the second association information, and the third association information, a first visual representation of the association strength between the key physical examination features and individual diseases, a second visual representation of the association strength between the key physical examination features and comorbidity combinations, and a ranking table of synergistic effects among diseases are generated; wherein, the first visual representation is used to display the association strength distribution of each key physical examination feature to each individual disease, the second visual representation is used to display the association strength distribution of each key physical examination feature to each comorbidity combination, and the ranking table is used to display the order of the strength of synergistic effects among different disease combinations; The comprehensive contribution of each of the key physical examination features to the predicted values ​​of the single disease and the comorbidity is calculated based on a preset algorithm, and a third visual representation of the contribution of each of the key physical examination features to the prediction results is generated based on the comprehensive contribution. The third visual representation is used to distinguish the positive or negative impact of each of the key physical examination features on the prediction results. By integrating the first visualization representation, the second visualization representation, the sorting table, and the third visualization representation, a single-disease independent risk correlation report and a comorbidity synergistic risk correlation report corresponding to each of the key physical examination features are output.

9. A device for assessing the combined risk of chronic diseases and comorbidities, characterized in that, include: The first construction module is used to perform hierarchical preprocessing on the physical examination-related data of multiple physical examination subjects to generate a standardized dataset containing single disease labels and comorbidity combination labels. The physical examination-related data includes chronic disease diagnosis data and comorbidity diagnosis data; The second construction module is used to construct a single-framework multi-objective collaborative ensemble model based on the standardized dataset. The single-framework multi-objective collaborative ensemble model is an integrated ensemble framework, which includes a single-disease prediction module and a comorbidity prediction module. The single-disease prediction module and the comorbidity prediction module reuse the feature representation layer and computation unit of the integrated ensemble framework. The single-disease prediction module and the comorbidity prediction module are collaboratively optimized through a global loss function. The prediction module inputs target data into the single-frame multi-objective collaborative integration model to obtain single-disease and comorbidity prediction values; the single-disease and comorbidity prediction values ​​are obtained synchronously by the single-frame multi-objective collaborative integration model through a single inference. The results generation module is used to calculate the comorbidity risk gain value based on the single disease prevalence prediction value and the comorbidity prevalence prediction value, and generate a structured assessment result based on the comorbidity risk gain value.

10. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as claimed in any one of claims 1 to 8.

Images