Brain disease risk prediction method and system based on big data analysis

Abstract

The invention discloses a brain disease risk prediction method and system based on big data analysis, and belongs to the technical field of brain disease risk prediction. The method comprises the following steps: carrying out standardized preprocessing and tagged classification on brain disease related big data to generate a feature data set; mining specific disease characteristics and risk factors in the set, and carding an association rule; training a risk prediction sub-model for each disease type based on the data, and building a multi-sub-model hierarchical prediction system; and collecting to-be-predicted object data, matching a disease type, and calling the corresponding sub-model to complete risk assessment. The system comprises multiple modules for collaborative operation, and a full-process closed loop of data storage, feature processing, model management and result output is realized. According to the scheme, the pertinence, the accuracy and the efficiency of risk prediction are improved, the traceability of the whole process and the dynamic optimization of the model are realized, and reliable technical support is provided for early screening and risk early warning of brain diseases.

Description

Technical Field

[0001] This invention relates to the field of brain disease risk prediction technology, and in particular to a brain disease risk prediction method and system based on big data analysis. Background Technology

[0002] With the advancement of medical informatization and the deep penetration of big data and artificial intelligence technologies into the medical field, the field of brain disease risk prediction is gradually developing towards data-driven intelligence. Currently, the data collection channels related to brain diseases are becoming increasingly abundant, covering various types such as clinical diagnosis and treatment data, daily health monitoring data, and medical history records, providing a massive data foundation for risk prediction. Simultaneously, machine learning, deep learning, and other algorithmic models are being used more and more widely in disease risk assessment, and various predictive models are being tested to identify the potential risk of developing brain diseases, assisting medical personnel in conducting early screening and intervention. Different types of brain diseases exhibit significant differences in pathogenesis and influencing factors, and related research is increasingly focusing on the impact of disease type specificity on prediction results. The industry's demand for precise and targeted risk prediction solutions continues to grow, with the hope of improving the scientific rigor and practicality of risk prediction through technological means, providing stronger support for brain health management.

[0003] However, existing brain disease risk prediction technologies still face many unresolved issues. First, due to the diversity of data sources, the collected big data related to brain diseases often exhibits heterogeneous formats and inconsistent data quality. A unified, standardized preprocessing process is lacking, and systematic labeling and classification by disease type is not implemented, making it difficult to directly use the data for efficient feature mining and model training. Second, existing technologies often employ generalized processing methods in feature extraction, failing to accurately uncover the specific pathogenesis characteristics and risk factors of different types of brain diseases, nor clarifying the correlation between them, resulting in insufficiently targeted input data for model training. Third, most current prediction models are single-model architectures attempting to adapt to the prediction needs of multiple types of brain diseases, making it difficult to consider the specificity of different diseases, leading to limited prediction accuracy. Furthermore, the lack of a unified model management and retrieval system prevents the rapid on-demand retrieval of models. Finally, existing solutions lack standardized processes in matching the disease type of the target object and conducting risk assessment, making it difficult to accurately connect the data to be predicted with the corresponding disease prediction model, affecting the efficiency and accuracy of risk assessment. The existence of these problems has resulted in the overall effectiveness of brain disease risk prediction failing to meet expectations, and making it difficult to satisfy the needs of clinical applications and health management for precise and efficient risk prediction. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for predicting the risk of brain diseases based on big data analysis.

[0005] The objective of this invention is achieved through the following technical solution: A method for predicting the risk of brain diseases based on big data analysis is provided, which includes the following steps: S1. Standardize the preprocessing of the collected big data related to brain diseases, remove invalid data, and then classify the data by label according to the specific type of brain disease to generate a feature data set corresponding to each type of brain disease. S2. Perform feature mining operations on the feature datasets corresponding to various types of brain diseases, extract the specific pathogenesis features and core risk factors of each type of brain disease, and sort out the correlation patterns between specific pathogenesis features and core risk factors through the correlation analysis module and generate a correlation pattern dataset. S3. Based on the feature data set corresponding to each type of brain disease, combined with the extracted specific pathogenesis features, core risk factors and correlation pattern datasets, corresponding risk prediction sub-models are trained for each type of brain disease for risk feature matching and evaluation calculation of each type of brain disease. A multi-sub-model hierarchical brain disease risk prediction system is built to integrate the risk prediction sub-models corresponding to each type of brain disease and realize the on-demand retrieval of risk prediction sub-models. Each risk prediction sub-model forms a unique mapping connection relationship with the corresponding brain disease type. S4. Collect basic health characteristics and vital signs data of the object to be predicted, match the corresponding brain disease type, and then call the risk prediction sub-model corresponding to the type from the multi-sub-model level brain disease risk prediction system. Input the basic health characteristics and vital signs data into the risk prediction sub-model to complete the risk assessment calculation for the corresponding brain disease type.

[0006] Furthermore, step S1 includes the following sub-steps: S1.1. Perform unified format processing on brain disease-related big data, converting heterogeneous storage formats of brain disease-related big data into a preset unified data format; S1.2. Perform missing value supplementation processing on the brain disease-related big data after the format is unified, and use interpolation method to improve the brain disease-related big data with missing data; S1.3. Perform outlier identification and removal on the brain disease-related big data after missing value completion to generate preprocessed brain disease-related big data; S1.4. Based on the clinical classification standards for brain diseases, add disease type labels to the preprocessed big data related to brain diseases, complete the labeling classification, and generate feature data sets corresponding to each type of brain disease.

[0007] Furthermore, step S2 includes the following sub-steps: S2.1. Perform initial feature screening on the feature datasets corresponding to various types of brain diseases, remove redundant feature data that are not related to the onset of brain diseases, and retain valid feature data; S2.2. Perform feature quantization processing on the retained effective feature data to convert non-numerical effective feature data into numerical effective feature data; S2.3. Rank the numerical effective feature data by feature importance, and extract the top-ranked feature data as the specific pathogenesis characteristics and core risk factors of various types of brain diseases; S2.4. Input the extracted specific pathogenesis features and core risk factors into the association analysis module. The module will then mine the association between the two, organize them into a corresponding association pattern dataset according to the type of brain disease, and output it synchronously.

[0008] Furthermore, step S3 includes the following sub-steps: S3.1. Divide the feature datasets corresponding to each type of brain disease into training datasets and validation datasets. Both datasets contain complete data dimensions of specific pathogenesis features, core risk factors, and correlation patterns. S3.2. Build the network structure of the initial prediction model. The initial prediction model is used for subsequent training to obtain risk prediction sub-models corresponding to various types of brain diseases. Set the loss function and optimizer parameters for model training. Iteratively train the initial prediction model using the training dataset and dynamically adjust the model's weight parameters and bias parameters. S3.3. Input the validation dataset into the trained initial prediction model, verify the model's fit and generalization ability through the validation module, and generate sub-models for risk prediction of various types of brain diseases for pre-calculation of risk assessment of various types of brain diseases. S3.4. Retain the validated risk prediction sub-models to be validated, classify the sub-models according to the types of brain diseases, and build a multi-level brain disease risk prediction system composed of multi-level sub-models to integrate the risk prediction sub-models corresponding to various types of brain diseases and realize the on-demand retrieval of risk prediction sub-models, and confirm the unique mapping connection between each level of sub-model and the corresponding disease type.

[0009] Furthermore, step S4 includes the following sub-steps: S4.1. Collect basic health characteristics and vital signs data of the subject to be predicted, and perform standardized processing on the collected data according to the preprocessing standard of S1 to generate standardized data to be predicted; S4.2. Perform feature matching operation between the standardized data to be predicted and the specific pathogenesis characteristics of each type of brain disease, and determine the type of brain disease corresponding to the object to be predicted based on the operation results; S4.3. Retrieve the risk prediction sub-models corresponding to the matching results from the multi-sub-model hierarchical brain disease risk prediction system, and establish a computational connection between the standardized data to be predicted and the risk prediction sub-models. S4.4. Input the standardized data to be predicted into the established operational connection of the risk prediction sub-model, and generate the risk assessment results of the brain disease type corresponding to the object to be predicted through the forward propagation operation of the model.

[0010] Furthermore, in step S1, after completing the labeling and classification, a feature data storage module is built. The feature data sets corresponding to each type of brain disease are stored in the independent data partitions of the feature data storage module. A unique disease type identifier is set for each data partition to achieve partitioned and isolated storage of each feature data set. The feature data storage module establishes a real-time data transmission connection with the feature mining operation after the preprocessing in S1.

[0011] Furthermore, in step S2, when sorting out the correlation between specific pathogenesis characteristics and core risk factors, the weight calculation module configures the correlation weight value for each group of correlations, and the correlation weight value and the corresponding correlation pattern are entered into the correlation pattern dataset. The correlation pattern dataset establishes a one-to-one correlation connection with the feature data set of each type of brain disease, and the calculation data of the weight calculation module is synchronously stored in the feature data storage module.

[0012] Furthermore, in step S3, a multi-sub-model management module is built. This module forms a bidirectional data connection with the multi-sub-model hierarchical brain disease risk prediction system. The multi-sub-model management module performs independent parameter retrieval, model updates, and performance testing for each risk prediction sub-model. The training logs, weight parameters, and bias parameters of each risk prediction sub-model are synchronously stored in the multi-sub-model management module. The multi-sub-model management module establishes a data interaction connection with the feature data storage module.

[0013] Furthermore, in step S4, after the risk assessment calculation is completed, the risk assessment result is input into the result output module for structured processing. During the structured processing, the feature matching data and parameter calculation data of the model calculation are retained. The result output module establishes a real-time data transmission connection with the risk prediction sub-model, synchronously receiving the intermediate calculation data and the final risk assessment result of the model. The result output module also synchronously interacts with the multi-sub-model management module to exchange calculation data.

[0014] In some embodiments, a brain disease risk prediction system based on big data analysis is provided. The system includes a data preprocessing module, a feature mining module, a model training module, a risk assessment module, a feature data storage module, a weight calculation module, a multi-sub-model management module, and a result output module. The data preprocessing module is unidirectionally connected to the feature data storage module. The data preprocessing module outputs the generated feature data sets corresponding to various types of brain diseases to the feature data storage module for partitioned and isolated storage. The feature mining module is bidirectionally connected to both the feature data storage module and the weight calculation module. The feature mining module retrieves the feature data sets from the feature data storage module, and the association weight values ​​configured by the weight calculation module for the association relationships are synchronized to the feature mining module and entered into the association pattern dataset. The model training module is bidirectionally connected to the feature data storage module, the weight calculation module, and the multi-sub-model management module. The model training module trains corresponding risk prediction sub-models for each type of brain disease based on the feature data sets and the association pattern dataset. The system performs risk feature matching and assessment calculations for various types of brain diseases. The multi-sub-model management module manages each risk prediction sub-model throughout the entire process and synchronously stores model-related data. The risk assessment module is bidirectionally connected to both the multi-sub-model management module and the result output module. The risk assessment module retrieves the corresponding risk prediction sub-model from the multi-sub-model management module to complete the risk assessment calculation and outputs the calculation results to the result output module. The result output module maintains real-time data transmission with the risk assessment module, performs structured processing on the risk assessment results, and retains all calculation data. The model training module builds a multi-sub-model hierarchical brain disease risk prediction system to integrate risk prediction sub-models corresponding to various types of brain diseases and realizes on-demand retrieval of risk prediction sub-models.

[0015] The beneficial effects of this invention are: (1) By standardizing the complete process of data preprocessing, targeted feature mining, classification model training and accurate risk assessment, we can standardize each prediction link and improve the pertinence and accuracy of brain disease risk prediction. (2) The combination of multi-module collaborative operation and multi-sub-model hierarchical system integrates various disease prediction resources, breaks through the limitations of single model adaptation, and enhances the applicability and efficiency of prediction methods; (3) Full-process data traceability, dynamic model management and closed-loop operation design provide support for method optimization and function expansion, and improve the operability and reliability of risk prediction as a whole. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the steps of a brain disease risk prediction method based on big data analysis; Figure 2 The following is a flowchart illustrating the specific steps of a brain disease risk prediction method based on big data analysis, provided as an example. Detailed Implementation

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

[0018] Example 1 See Figure 1 This embodiment provides a method for predicting the risk of brain diseases based on big data analysis. The method includes the following steps: S1. Standardize the preprocessing of the collected big data related to brain diseases, remove invalid data, and then classify the data by label according to the specific type of brain disease to generate a feature data set corresponding to each type of brain disease. S2. Perform feature mining operations on the feature datasets corresponding to various types of brain diseases, extract the specific pathogenesis features and core risk factors of each type of brain disease, and sort out the correlation patterns between specific pathogenesis features and core risk factors through the correlation analysis module and generate a correlation pattern dataset. S3. Based on the feature data set corresponding to each type of brain disease, combined with the extracted specific pathogenesis features, core risk factors and correlation pattern datasets, corresponding risk prediction sub-models are trained for each type of brain disease for risk feature matching and evaluation calculation of each type of brain disease. A multi-sub-model hierarchical brain disease risk prediction system is built to integrate the risk prediction sub-models corresponding to each type of brain disease and realize the on-demand retrieval of risk prediction sub-models. Each risk prediction sub-model forms a unique mapping connection relationship with the corresponding brain disease type. S4. Collect basic health characteristics and vital signs data of the object to be predicted, match the corresponding brain disease type, and then call the risk prediction sub-model corresponding to the type from the multi-sub-model level brain disease risk prediction system. Input the basic health characteristics and vital signs data into the risk prediction sub-model to complete the risk assessment calculation for the corresponding brain disease type.

[0019] In some embodiments, step S1 includes the following sub-steps: S1.1. Perform unified format processing on brain disease-related big data, converting heterogeneous storage formats of brain disease-related big data into a preset unified data format; S1.2. Perform missing value supplementation processing on the brain disease-related big data after the format is unified, and use interpolation method to improve the brain disease-related big data with missing data; S1.3. Perform outlier identification and removal on the brain disease-related big data after missing value completion to generate preprocessed brain disease-related big data; S1.4. Based on the clinical classification standards for brain diseases, add disease type labels to the preprocessed big data related to brain diseases, complete the labeling classification, and generate feature data sets corresponding to each type of brain disease.

[0020] In some embodiments, step S2 includes the following sub-steps: S2.1. Perform initial feature screening on the feature datasets corresponding to various types of brain diseases, remove redundant feature data that are not related to the onset of brain diseases, and retain valid feature data; S2.2. Perform feature quantization processing on the retained effective feature data to convert non-numerical effective feature data into numerical effective feature data; S2.3. Rank the numerical effective feature data by feature importance, and extract the top-ranked feature data as the specific pathogenesis characteristics and core risk factors of various types of brain diseases; S2.4. Input the extracted specific pathogenesis features and core risk factors into the association analysis module. The module will then mine the association between the two, organize them into a corresponding association pattern dataset according to the type of brain disease, and output it synchronously.

[0021] In some embodiments, step S3 includes the following sub-steps: S3.1. Divide the feature datasets corresponding to each type of brain disease into training datasets and validation datasets. Both datasets contain complete data dimensions of specific pathogenesis features, core risk factors, and correlation patterns. S3.2. Build the network structure of the initial prediction model. The initial prediction model is used for subsequent training to obtain risk prediction sub-models corresponding to various types of brain diseases. Set the loss function and optimizer parameters for model training. Iteratively train the initial prediction model using the training dataset and dynamically adjust the model's weight parameters and bias parameters. S3.3. Input the validation dataset into the trained initial prediction model, verify the model's fit and generalization ability through the validation module, and generate sub-models for risk prediction of various types of brain diseases for pre-calculation of risk assessment of various types of brain diseases. S3.4. Retain the validated risk prediction sub-models to be validated, classify the sub-models according to the types of brain diseases, and build a multi-level brain disease risk prediction system composed of multi-level sub-models to integrate the risk prediction sub-models corresponding to various types of brain diseases and realize the on-demand retrieval of risk prediction sub-models, and confirm the unique mapping connection between each level of sub-model and the corresponding disease type.

[0022] In some embodiments, step S4 includes the following sub-steps: S4.1. Collect basic health characteristics and vital signs data of the subject to be predicted, and perform standardized processing on the collected data according to the preprocessing standard of S1 to generate standardized data to be predicted; S4.2. Perform feature matching operation between the standardized data to be predicted and the specific pathogenesis characteristics of each type of brain disease, and determine the type of brain disease corresponding to the object to be predicted based on the operation results; S4.3. Retrieve the risk prediction sub-models corresponding to the matching results from the multi-sub-model hierarchical brain disease risk prediction system, and establish a computational connection between the standardized data to be predicted and the risk prediction sub-models. S4.4. Input the standardized data to be predicted into the established operational connection of the risk prediction sub-model, and generate the risk assessment results of the brain disease type corresponding to the object to be predicted through the forward propagation operation of the model.

[0023] In some embodiments, after completing the labeling and classification in step S1, a feature data storage module is built, and the feature data sets corresponding to each type of brain disease are stored in the independent data partitions of the feature data storage module. A unique disease type identifier is set for each data partition to achieve partitioned and isolated storage of each feature data set. The feature data storage module establishes a real-time data transmission connection with the feature mining operation after the preprocessing in S1.

[0024] In some embodiments, in step S2, when sorting out the correlation between specific pathogenesis characteristics and core risk factors, the weight calculation module configures a correlation weight value for each group of correlations, and the correlation weight value and the corresponding correlation pattern are entered into the correlation pattern dataset. The correlation pattern dataset establishes a one-to-one correlation connection with the feature data set of each type of brain disease, and the calculation data of the weight calculation module is synchronously stored in the feature data storage module.

[0025] In some embodiments, in step S3, a multi-sub-model management module is built. This module forms a bidirectional data connection with the multi-sub-model hierarchical brain disease risk prediction system. The multi-sub-model management module performs independent parameter retrieval, model updates, and performance testing for each risk prediction sub-model. The training logs, weight parameters, and bias parameters of each risk prediction sub-model are synchronously stored in the multi-sub-model management module. The multi-sub-model management module establishes a data interaction connection with the feature data storage module.

[0026] In some embodiments, in step S4, after the risk assessment calculation is completed, the risk assessment result is input into the result output module for structured processing. During the structured processing, the feature matching data and parameter calculation data of the model calculation are retained. The result output module establishes a real-time data transmission connection with the risk prediction sub-model, synchronously receives the intermediate calculation data and the final risk assessment result of the model, and synchronously interacts with the multi-sub-model management module to exchange calculation data.

[0027] In some embodiments, a brain disease risk prediction system based on big data analysis is provided. The system includes a data preprocessing module, a feature mining module, a model training module, a risk assessment module, a feature data storage module, a weight calculation module, a multi-sub-model management module, and a result output module. The data preprocessing module is unidirectionally connected to the feature data storage module. The data preprocessing module outputs the generated feature data sets corresponding to various types of brain diseases to the feature data storage module for partitioned and isolated storage. The feature mining module is bidirectionally connected to both the feature data storage module and the weight calculation module. The feature mining module retrieves the feature data sets from the feature data storage module, and the association weight values ​​configured by the weight calculation module for the association relationships are synchronized to the feature mining module and entered into the association pattern dataset. The model training module is bidirectionally connected to the feature data storage module, the weight calculation module, and the multi-sub-model management module. The model training module trains corresponding risk prediction sub-models for each type of brain disease based on the feature data sets and the association pattern dataset. The system performs risk feature matching and assessment calculations for various types of brain diseases. The multi-sub-model management module manages each risk prediction sub-model throughout the entire process and synchronously stores model-related data. The risk assessment module is bidirectionally connected to both the multi-sub-model management module and the result output module. The risk assessment module retrieves the corresponding risk prediction sub-model from the multi-sub-model management module to complete the risk assessment calculation and outputs the calculation results to the result output module. The result output module maintains real-time data transmission with the risk assessment module, performs structured processing on the risk assessment results, and retains all calculation data. The model training module builds a multi-sub-model hierarchical brain disease risk prediction system to integrate risk prediction sub-models corresponding to various types of brain diseases and realizes on-demand retrieval of risk prediction sub-models.

[0028] Example 2 This embodiment provides a specific implementation process for a brain disease risk prediction method based on big data analysis, such as... Figure 2 As shown, the specific implementation process is as follows: S1. Preprocess big data related to brain diseases, classify and label it to generate a feature data set: The big data related to brain diseases comes from various data collection channels, covering a wide range of basic information and clinical data related to brain health. In its raw state, this data suffers from issues such as inconsistent formats, missing data, and the presence of abnormal or interfering data, making it unsuitable for direct use in subsequent feature mining and model training. Therefore, standardized preprocessing is required before labeling and classification. "Heterogeneous storage formats" refer to the different storage standards used by data from different collection channels, such as JSON, XML, Excel, and TXT formats. These formats have different data structures and field definitions, requiring unified conversion before they can be recognized by subsequent modules.

[0029] S1.1. Perform format unification processing on brain disease-related big data, converting heterogeneous storage formats of brain disease-related big data into a preset unified data format: Big data related to brain diseases acquired from different collection channels uses different storage specifications and formats. These heterogeneous storage formats cannot be uniformly identified and parsed by subsequent processing modules, so format unification processing is required. In the processing, the storage format of all raw data is first identified one by one using a data format identification tool to clarify the structural characteristics of various types of data (such as whether it is a key-value pair structure, whether it contains table header information), field meanings (such as the specific definitions of fields such as "blood pressure value" and "onset time"), and storage rules (such as the number of decimal places to retain for numerical data). Then, according to the preset unified data format requirements (CSV format is preferred because of its strong compatibility and high parsing efficiency, with fields separated by commas and each line representing a data record), the data in different formats is converted. For example, the "health indicator" field nested in the JSON format is split into an independent column, and the tag nodes in the XML format are mapped to the field names in the CSV format. This ensures that the converted data is consistent in terms of structure, field type (such as converting string-type date data to standard date format strings), and encoding method (UTF-8 encoding is used uniformly). This allows subsequent operations such as missing value supplementation and outlier removal to be carried out smoothly, ensuring the uniformity and efficiency of data processing.

[0030] In some specific implementations, a total of 100,000 big data entries related to brain diseases were collected, including 32,000 entries in JSON format, 28,000 entries in XML format, 25,000 entries in Excel format, and 15,000 entries in TXT format. After format conversion, all data were converted into CSV files. Each file contained 20 fields, corresponding to core dimensions such as basic information, clinical indicators, and medical history records. The total number of rows in the converted data remained at 100,000, with no data loss or redundancy.

[0031] S1.2. Perform missing value imputation processing on the standardized brain disease-related big data, and use interpolation to improve the brain disease-related big data with missing data: Even after standardizing the data format, some fields may be missing (e.g., the "blood lipid index" and "sleep duration" fields are empty). Directly retaining these missing data can lead to incomplete feature extraction and biased model training data. Therefore, missing value supplementation is necessary. Interpolation is a common method for estimating missing values ​​based on the overall distribution of the data. In this embodiment, linear interpolation is preferred, while spline interpolation is used for time-series data (e.g., multiple consecutive blood pressure monitoring data). The specific implementation of linear interpolation is as follows: for a field with missing values, find two adjacent valid data points in the dataset. Based on the values ​​and locations of these two data points, calculate the replacement value for the missing value using a linear equation. Spline interpolation, on the other hand, constructs a smooth spline curve to fit the changing trend of adjacent valid data, thereby obtaining the estimated result for the missing value. During the supplementation process, it is necessary to first determine the type of missing values ​​(whether they are random or systematic; random missing values ​​are those unrelated to the data itself, while systematic missing values ​​are those caused by limitations in the data collection conditions, resulting in a general lack of certain types of data). Only randomly missing data should be interpolated and supplemented, while systematically missing data should be marked separately before processing to avoid affecting subsequent analysis results due to the supplemented values ​​deviating from the actual data characteristics. This operation improves the completeness of the data and provides comprehensive data support for subsequent processing.

[0032] In some specific implementations, among the 100,000 data entries after standardization, 8,600 entries contained missing values. The missing fields mainly fell into four categories: "blood lipid indicators," "sleep duration," "exercise frequency," and "blood glucose levels." Specifically, 2,300 entries were missing for "blood lipid indicators," 2,100 for "sleep duration," 2,600 for "exercise frequency," and 1,600 for "blood glucose levels." Linear interpolation was used to supplement the missing data for "blood lipid indicators" and "blood glucose levels," spline interpolation was used to supplement the missing data for "sleep duration," and the missing data for "exercise frequency" was supplemented using the mean of a similar population (the similar population was divided by age and gender). After supplementation, the missing data rate was reduced to below 0.5%, meeting the requirements for subsequent processing.

[0033] S1.3. Perform outlier identification and removal on the brain disease-related big data after missing value imputation to generate preprocessed brain disease-related big data: Even after formatting and missing value supplementation, data may contain outliers due to data acquisition errors (e.g., blood pressure entered as 2000 mmHg), transmission failures (e.g., truncated values), or special anomalies (e.g., extreme physiological indicators of a single patient). These outliers deviate significantly from the overall data distribution and can interfere with subsequent feature mining and model training results. Therefore, outlier identification and removal are necessary. This embodiment uses a combination of the 3σ principle and box plot method to identify outliers: the 3σ principle states that for normally distributed data, data exceeding the mean ± 3 standard deviations are considered outliers; the box plot method calculates the quartiles (Q1, Q2, Q3) to determine the outlier threshold (below Q1 - 1.5 × IQR or above Q3 + 1.5 × IQR, where IQR = Q3 - Q1), and is suitable for non-normally distributed data. During the identification process, a distribution test is first performed on each numerical field to determine whether it conforms to a normal distribution. Then, the corresponding outlier identification method is selected to filter out outliers that clearly do not conform to the overall characteristics of the data. Afterwards, through manual review (for fields with an outlier rate of less than 0.1%) or comparison with the original data collection records, it is confirmed that the removed data is truly invalid data, rather than valid data with special significance (such as extreme indicators of a patient due to a sudden illness that need to be retained and marked separately). Finally, complete and valid preprocessed data is generated. The specific outlier identification and removal for various fields are shown in Table 1 below: Table 1. Outlier Identification and Removal Results Field Name Distribution type Outlier Judgment Criteria Number of outliers Number of items to be removed Retain and mark the quantity Blood pressure (mmHg) normal distribution Mean ±3σ (80-180) 320 305 15 Blood glucose level (mmol / L) normal distribution Mean ± 3σ (3.9-11.1) 280 272 8 Blood lipid levels (mmol / L) Nonnormal distribution Q1 - 1.5 × IQR to Q3 + 1.5 × IQR (0.5 - 6.5) 450 438 12 Sleep duration (h) Nonnormal distribution Q1 - 1.5 × IQR to Q3 + 1.5 × IQR (3-10) 180 175 5 S1.4. Based on the clinical classification standards for brain diseases, add disease type labels to the preprocessed big data related to brain diseases, complete the labeling classification, and generate feature data sets corresponding to each type of brain disease: The preprocessed big data encompasses information on various types of brain diseases. To achieve targeted feature mining and model training, it needs to be categorized by disease type. The "Clinical Classification Standard for Brain Diseases" references the "Classification and Codes of Neurological Diseases," dividing brain diseases into five major categories: cerebrovascular diseases, neurodegenerative diseases, infectious brain diseases, brain tumors, and other brain diseases. Each major category is further subdivided into specific subtypes (e.g., cerebrovascular diseases include subtypes such as cerebral hemorrhage and cerebral infarction). Based on this standard, the disease type corresponding to each piece of preprocessed data is analyzed one by one. By matching fields such as "disease diagnosis result" and "clinical symptoms" in the data, a unique disease type label is added (using numerical coding: 1 represents cerebrovascular disease, 2 represents neurodegenerative disease, 3 represents infectious brain disease, 4 represents brain tumor, and 5 represents other brain diseases). The label clearly identifies the brain disease category to which the data belongs. After adding labels, the data with the same disease type labels are categorized and integrated to form feature data sets corresponding to various types of brain diseases. The data in each set is associated with a specific disease type. For example, the dataset with label 1 corresponds to data related to cerebrovascular diseases, and the dataset with label 2 corresponds to data related to neurodegenerative diseases. This lays the foundation for subsequent feature mining and model training for different diseases.

[0034] In some specific implementations, a total of 98,000 valid data entries were obtained after preprocessing. Among them, there were 35,000 entries of cerebrovascular disease with label 1, 22,000 entries of neurodegenerative disease with label 2, 18,000 entries of brain infectious disease with label 3, 13,000 entries of brain tumor with label 4, and 10,000 entries of other brain diseases with label 5. Each feature dataset was named "disease type label_feature dataset.csv" and stored in a designated folder.

[0035] After generating feature datasets corresponding to various types of brain diseases, a feature data storage module is built. This module adopts a distributed storage architecture (based on the HDFS distributed file system), storing the feature datasets corresponding to each type of brain disease in independent data partitions within the feature data storage module. Each data partition is allocated independent storage resources (100GB disk space, 100Mbps read / write bandwidth), and a unique disease type identifier (consistent with the data label, 1-5) is assigned to each data partition. This achieves partitioned and isolated storage of each feature dataset, avoiding interference between different types of data. The feature data storage module establishes a real-time data transmission connection with the subsequent feature mining operations in S2, enabling data retrieval through a RESTful API interface. The interface response time is no more than 500ms, ensuring that the required feature datasets can be retrieved quickly and accurately during feature mining. The isolated design of the data partitions makes data management clearer. If updates or maintenance are needed for data of a certain type of disease, the corresponding partition can be directly accessed without affecting the normal use of other types of data.

[0036] S2. Mine the feature dataset, extract relevant features and risk factors, and identify correlation patterns: The feature datasets corresponding to various types of brain diseases contain a large amount of disease-related feature information (such as various indicators corresponding to 20 fields). However, this feature information is complex, and some features have limited role in predicting disease risk (such as fields like "patient number" which are only used for identification). Therefore, it is necessary to use feature mining operations to filter out key features and sort out their correlations to provide core input data for model training. Among them, the "correlation pattern dataset" refers to a structured dataset containing information such as the correlation between specific pathogenesis features and core risk factors, correlation weight values, and correlation strength, which is used to clarify the intrinsic relationship between different features.

[0037] S2.1. Perform initial feature screening on the feature datasets corresponding to various types of brain diseases, removing redundant feature data that is not related to the onset of brain diseases, and retaining valid feature data: The initial feature screening is based on the pathogenesis of various types of brain diseases (such as the correlation between cerebrovascular diseases and indicators such as blood pressure and blood lipids) and influencing factors. It employs a combination of Pearson correlation coefficient and mutual information methods to analyze the association between each feature and the disease status (whether or not the disease is present). The Pearson correlation coefficient is used to calculate the linear correlation between numerical features and disease status (converted to a 0-1 variable, where 0 represents no disease and 1 represents disease), with a value range of [-1, 1]. The closer the absolute value is to 1, the stronger the association. The mutual information method is used to calculate the informational association between non-numerical features (such as "lifestyle habits" and "past medical history") and disease status; a larger value indicates a stronger association. A correlation threshold is set (Pearson correlation coefficient absolute value ≥ 0.1, mutual information value ≥ 0.05). Features that are not directly or indirectly related to disease onset and cannot provide effective information for risk prediction (such as "patient number" and "data acquisition device number", whose Pearson correlation coefficient absolute value < 0.01 and mutual information value < 0.005) are identified as redundant features and removed. Features that can reflect the correlation with disease onset and support risk prediction (such as "blood pressure value", "blood lipid index", "smoking history", etc., whose correlation indicators all meet the threshold requirements) are retained as valid feature data. This operation can reduce the workload of subsequent data processing, improve feature processing efficiency, and avoid redundant features from interfering with subsequent analysis results, ensuring that subsequent processing focuses on effective key features.

[0038] In some specific implementations, 20 original features from a dataset of 35,000 data points on cerebrovascular diseases were initially screened. Using the Pearson correlation coefficient method, the correlation coefficients for "blood pressure" (0.32), "blood lipid levels" (0.28), "blood glucose" (0.21), and "smoking history" (converted to a 0-1 variable) (0.18) all met the threshold requirements. "Patient ID" (0.003) and "data acquisition device ID" (0.002) were identified as redundant features. Using the mutual information method, the mutual information values ​​for "lifestyle habits" (0.12) and "past medical history" (0.15) met the threshold requirements. Ultimately, 15 valid features were retained, and 5 redundant features were removed.

[0039] S2.2. Perform feature quantization on the retained effective feature data to convert non-numerical effective feature data into numerical effective feature data: Among the retained valid feature data, some are non-numerical features (such as "lifestyle habits" including classification values ​​such as "regular exercise", "sedentary", and "occasional exercise", and "past medical history" being descriptive text). These non-numerical features cannot be directly identified and calculated by the model, so they need to be quantified. During the quantification process, appropriate transformation methods are selected based on the attributes and characteristics of non-numerical features: For categorical features (such as "lifestyle habits," "smoking history," and "drinking history"), if they are binary features (such as "smoking history" divided into "yes" and "no"), label encoding is used (converted to 0 and 1); if they are multi-category features without a logical order (such as "lifestyle habits" divided into "regular exercise," "occasional exercise," and "sedentary lifestyle"), one-hot encoding is used (converted to multiple 0-1 variables, such as "lifestyle habits_regular exercise," "lifestyle habits_occasional exercise," and "lifestyle habits_sedentary lifestyle"). For textual features (such as "hypertension history for 5 years" in "past medical history"), keyword extraction (extracting keywords such as "hypertension" and "5 years") is used followed by numerical mapping ("hypertension" is mapped to 1, no related medical history is mapped to 0, and the duration of the medical history is directly retained as a numerical value). During the transformation process, it is necessary to ensure that the numerical values ​​can accurately reflect the meaning and differences of the original non-numerical features without changing the information value of the features themselves. For example, after one-hot encoding, the variable corresponding to each category only takes the value of 0 or 1 to avoid ambiguity. Quantization transforms all effective feature data into numerical data, meeting the input data format requirements for subsequent model training. This standardization of feature data ensures unified model computation. Table 2 shows the quantization encoding methods for some common non-numerical features. Table 2 Non-numerical feature quantization coding table Original non-numerical features Feature type Quantification methods Quantified results smoking history Two categories Tag Encoding Yes = 1, No = 0 Lifestyle Multiple categories (no order) One-hot encoding Regular exercise = 100, occasional exercise = 010, sedentary lifestyle = 001 drinking history Two categories Tag Encoding Yes = 1, No = 0 Past medical history (hypertension) Text type Keyword mapping Present = 1, Absent = 0 Past medical history (diabetes) Text type Keyword mapping Present = 1, Absent = 0 S2.3. Rank the numerical effective feature data by feature importance, and extract the top-ranked features as specific pathogenesis characteristics and core risk factors for various types of brain diseases: Feature importance ranking is based on the degree of influence of features on the prediction results of brain disease risk, and is implemented using the random forest algorithm. This algorithm constructs multiple decision trees and calculates the importance score of each feature based on its contribution to splitting in each decision tree (i.e., the information gain brought by the feature when splitting a decision tree node). The higher the score, the greater the influence of the feature on the prediction result. In the specific implementation, quantified numerical effective feature data is used as input, and disease incidence status (0-1 variable) is used as output. A random forest model containing 100 decision trees is trained. After the model is trained, the importance score of each feature is output, and the features are then ranked from high to low. After ranking, the top-ranked feature data is extracted (the top 8 features are extracted to ensure coverage of major influencing factors and avoid excessive features that would complicate the model). These features can most intuitively and accurately reflect the unique pathogenesis characteristics of various types of brain diseases and are the key basis for distinguishing different disease types and judging the risk of disease occurrence. They are identified as the specific pathogenesis characteristics and core risk factors of each type of brain disease. For example, in the ranking results of cerebrovascular diseases, features such as "blood pressure value", "blood lipid index", "smoking history" and "previous history of hypertension" scored highly and were identified as specific pathogenesis features and core risk factors of this type of disease. This extraction process ensures that the input data for model training is targeted and effective, which can improve the model's prediction accuracy of disease risk.

[0040] In some specific implementations, the importance of 15 quantified effective features of neurodegenerative diseases was ranked. The feature importance scores (out of 10) output by the random forest model were as follows: cognitive function score 8.9, age-related indicators 8.2, genetic factors-related indicators 7.8, sleep quality 7.5, exercise frequency 7.1, dietary structure 6.8, history of neurological diseases 6.5, mental stress level 6.2, and other indicators 5.8 and below. Finally, the top 8 features were extracted as the specific pathogenesis features and core risk factors of this type of disease, and their cumulative importance score accounted for 75%, which can cover the main influencing factors.

[0041] S2.4. Input the extracted specific pathogenesis features and core risk factors into the association analysis module. The module will then mine the association between the two, organize them according to the type of brain disease to form a corresponding association pattern dataset, and output it synchronously. The association analysis module incorporates the Apriori association mining algorithm. This algorithm discovers the relationship between specific disease characteristics and core risk factors by mining frequent itemsets (i.e., feature combinations with a high probability of co-occurrence). The core idea is that "all subsets of a frequent itemset are frequent itemsets, and all supersets of a non-frequent itemset are non-frequent itemsets." After inputting the extracted specific disease characteristics and core risk factors into the module, minimum support (0.1) and minimum confidence (0.5) are first set. Minimum support refers to the frequency threshold of a feature combination appearing in the dataset, and minimum confidence refers to the conditional probability threshold of the occurrence of one feature after the occurrence of another feature in a feature combination. The module scans the dataset to filter out frequent 1-itemsets (single features) that meet the minimum support, then generates candidate 2-itemsets (combinations of two features) based on the frequent 1-itemsets, and scans the dataset again to filter out frequent 2-itemsets, and so on, until no new frequent itemsets can be generated. Next, the association strength (support, confidence, and lift) between features is calculated based on frequent itemsets. Support reflects the frequency of feature combinations, confidence reflects the degree of causal association between features, and lift reflects the effectiveness of the association (lift > 1 indicates a positive correlation, < 1 indicates a negative correlation). For example, the support of "history of hypertension" and "onset of cerebrovascular disease" is 0.15, the confidence is 0.65, and the lift is 2.3, indicating a strong positive correlation between the two. Subsequently, these association information are categorized and organized according to brain disease types. The associations (including feature combinations, support, confidence, and lift) between specific pathogenesis features and core risk factors under the same disease type are integrated to form a structured association pattern dataset (stored in JSON format, with each record corresponding to a set of associations). Each dataset corresponds to a type of brain disease, clearly presenting the association patterns between the specific pathogenesis features and core risk factors of that type of disease. This dataset is simultaneously output to provide association basis for subsequent model training.

[0042] When analyzing the correlation between specific disease characteristics and core risk factors, a weight calculation module was built. This module calculates the correlation weight value based on the lift obtained from the correlation analysis. The specific calculation formula is: Correlation weight value = (lift - 1) / (max lift - 1), where the max lift is the maximum lift of all correlations under this disease type. The weight value ranges from [0, 1]. The closer the correlation and the greater its contribution to the prediction result, the higher the weight value is assigned (e.g., if the lift is 2.3 and the max lift is 3.0, then the correlation weight value = (2.3 - 1) / (3.0 - 1) = 0.65). The correlation weight value and the corresponding correlation patterns (feature combination, support, confidence, lift) are entered into the correlation pattern dataset. This ensures that the correlation pattern dataset not only contains the correlation itself but also its importance information, providing a basis for the subsequent model training process to prioritize different correlations. The association pattern dataset is linked one-to-one with the feature data sets of various types of brain diseases (associated through disease type labels) to ensure that the association data of the corresponding disease type can be accurately called during the subsequent model training process. The operation data of the weight calculation module (including the calculation process of association weight values ​​and intermediate parameters) is synchronously stored in the feature data storage module to realize data backup and unified management, which facilitates subsequent data traceability and verification.

[0043] S3. Train disease-specific prediction sub-models, build a multi-sub-model prediction system, and establish a mapping: The model training is based on specific data for various types of brain diseases. Through a targeted training process, sub-models capable of accurately predicting the risk of corresponding diseases are obtained. Then, by building a multi-sub-model hierarchical prediction system, the effective integration and efficient use of all sub-models are achieved. Among them, the "multi-sub-model hierarchical brain disease risk prediction system" refers to a centralized management system that manages diseases hierarchically, includes a sub-model index library and calling interfaces, and can quickly locate and retrieve the corresponding sub-model based on the disease type.

[0044] S3.1. Divide the feature datasets corresponding to various types of brain diseases into training datasets and validation datasets. Both datasets contain complete data dimensions of specific pathogenesis characteristics, core risk factors, and correlation patterns. The dataset was split in an 8:2 ratio (80% training dataset and 20% validation dataset). Stratified sampling was used during the split to ensure that both the training and validation datasets comprehensively covered all dimensions of the specific pathogenesis characteristics, core risk factors, and correlation patterns of the corresponding disease type (such as the distribution of feature values ​​and the frequency of correlations), leaving no key information unexplored. The specific implementation of stratified sampling was as follows: stratification was performed by disease subtype (e.g., cerebrovascular diseases are divided into cerebral hemorrhage, cerebral infarction, etc.) and disease status (affected / not affected). Data was extracted from each stratum in an 8:2 ratio to form the training and validation datasets, avoiding data distribution bias. The training dataset was used for iterative model training, providing data support for model parameter adjustment; the validation dataset was used for performance testing after model training, verifying the model's predictive effectiveness. During the split, it was ensured that the feature distributions (e.g., the mean and standard deviation of "blood pressure values") and correlation patterns (e.g., the strength of the correlation between "history of hypertension" and "onset of disease") of the two datasets remained consistent with the original feature dataset, enabling the trained model to have good generalization ability and ensuring that the validation results accurately reflected the model's actual performance.

[0045] In some specific implementations, the brain tumor feature dataset totaled 13,000 records, divided into a training dataset of 10,400 records and a validation dataset of 2,600 records in an 8:2 ratio. Statistical analysis verified that the mean and standard deviation of the "tumor marker index" in the training dataset were 5.2 and 1.8, respectively, while the mean and standard deviation of the same index in the validation dataset were 5.1 and 1.9, respectively, with a difference of less than 5%. The support for the association between "smoking history + tumor incidence" was 0.12 in the training dataset and 0.11 in the validation dataset, with a difference of less than 10%, meeting the data distribution consistency requirement.

[0046] S3.2. Construct the network structure of the initial prediction model. This initial prediction model is used for subsequent training to obtain risk prediction sub-models corresponding to various types of brain diseases. Set the loss function and optimizer parameters for model training, and iteratively train the initial prediction model using the training dataset, dynamically adjusting the model's weight and bias parameters. The initial prediction model uses a neural network model from deep learning. The network structure includes an input layer, hidden layers, and an output layer. The structural parameters of each layer are shown in Table 3. Table 3 Initial Prediction Model Network Structure Parameters Network layer Number of neurons Activation function Core Functions Input layer Consistent with the dimensions of the input data (8, corresponding to 8 core features). - Receive training data and convert it into a tensor format that the model can process. Hidden layer 1 64 ReLU function Perform preliminary nonlinear transformation on the input features Hidden layer 2 32 ReLU function Further extract higher-order correlation information of features Hidden layer 3 16 ReLU function Compress feature dimensions while retaining key information Output layer 1 Sigmoid function Output the probability of disease risk (between 0 and 1). Before model training, the loss function and optimizer parameters are set: the loss function is the binary cross-entropy loss function, which is suitable for binary risk prediction tasks. The calculation formula is: Loss=-y×log(p)-(1-y)×log(1-p), where y is the true label (0 or 1) and p is the risk probability predicted by the model; the optimizer is the Adam optimizer, which has fast convergence speed and good stability. The learning rate is set to 0.001 (initial value), and a dynamic adjustment strategy is adopted (the learning rate decays to 0.9 every 100 iterations). The maximum number of iterations is 1000, the batch size is 64 (64 data points are input for training each time), and an early stopping strategy is adopted (training is stopped when the value of the validation set loss function does not decrease for 20 consecutive iterations to avoid overfitting). During iterative training, the training dataset data is input into the initial prediction model in batches. The prediction results are obtained through forward propagation (the input data is passed from the input layer through the hidden layer to the output layer, and each layer is operated on by the weight matrix and activation function). The deviation between the prediction results and the actual results is calculated using the loss function. Then, the backpropagation algorithm (based on gradient descent, the gradient of the loss function with respect to the weight parameters and bias parameters of each layer is calculated, and the parameter values ​​are adjusted according to the gradient direction to reduce the loss function value) is used. This process is repeated until the model converges (the loss function value stabilizes below 0.1) or the preset upper limit of the number of iterations is reached.

[0047] In some specific implementations, during the initial prediction model training process for infectious brain diseases, the initial learning rate was 0.001, the batch size was 64, and the model converged after 500 iterations. The loss function value of the training dataset dropped to 0.08. At this time, the mean weight parameter of hidden layer 1 was 0.25 and the standard deviation was 0.12, the mean weight parameter of hidden layer 2 was 0.31 and the standard deviation was 0.15, the mean weight parameter of hidden layer 3 was 0.28 and the standard deviation was 0.13, and the bias parameters were all between [-0.1, 0.1].

[0048] S3.3. Input the validation dataset into the trained initial prediction model, and verify the model's fit and generalization ability through the validation module. Generate sub-models for risk prediction of various types of brain diseases for pre-calculation of risk assessment for each type of brain disease: The validation module incorporates multiple evaluation metrics to comprehensively validate the model's fit and generalization ability: fit is evaluated using accuracy, precision, recall, and F1 score on the training set, while generalization ability is evaluated using the same metrics on the validation set, along with the AUC (area under the ROC curve). The formulas for each metric are as follows: Accuracy = (True positive + True negative) / (True positive + True negative + False positive + False negative), reflecting the overall prediction accuracy of the model; Precision = True positive / (True positive + False positive), reflecting the proportion of true positives among the positive results predicted by the model; Recall rate = true positives / (true positives + false negatives), which reflects the model's ability to identify true positive samples; F1 score = 2 × precision × recall / (precision + recall), which comprehensively reflects both precision and recall; AUC value: The ROC curve is plotted with the false positive rate on the horizontal axis and the true positive rate on the vertical axis. The closer the AUC value is to 1, the stronger the model's generalization ability.

[0049] The validation module receives the prediction results (risk probabilities) output by the trained initial prediction model, sets a risk threshold (0.5, where a prediction probability ≥ 0.5 is considered high risk and < 0.5 is considered low risk), converts the prediction results into 0-1 labels, and then compares and analyzes them with the actual labels (disease status) in the validation dataset, calculating various evaluation indicators. Validation qualification criteria are set as follows: accuracy ≥ 0.85, precision ≥ 0.8, recall ≥ 0.8, F1 score ≥ 0.8, and AUC value ≥ 0.85. Models that meet all qualification criteria are selected and identified as the risk prediction sub-models to be validated for each type of brain disease. These sub-models possess preliminary risk assessment pre-computation capabilities, enabling them to perform risk prediction calculations on input data and output results.

[0050] S3.4. Retain the validated risk prediction sub-models to be validated, classify the sub-models according to the type of brain disease, and build a multi-level brain disease risk prediction system composed of multiple sub-models to integrate the risk prediction sub-models corresponding to various types of brain diseases and realize the on-demand retrieval of risk prediction sub-models. Confirm the unique mapping connection between each level of sub-model and the corresponding disease type: Validated sub-models for risk prediction are retained and categorized according to their corresponding brain disease types (classes 1-5). Sub-models of the same disease type are grouped into one level (5 levels in total). Each level stores information such as network structure parameters (weight matrix, bias parameters), evaluation results, and training logs. The multi-sub-model level brain disease risk prediction system internally sets up a sub-model index and a calling interface: the index uses a key-value pair storage structure, where the key is the disease type label (1-5), and the value contains information such as the sub-model's storage path (distributed storage address), model version number, and core parameter summary (such as input feature dimension, output type); the calling interface adopts a RESTful API design, supporting the input of disease type labels and data to be predicted via HTTP requests, with an interface response time of no more than 1 second. During the setup process, it is essential to ensure that the system can achieve centralized management and efficient integration of all sub-models. When a risk prediction request is received, the interface first parses the disease type label in the request, queries the index to obtain the storage path and parameter information of the corresponding sub-model, and then loads the sub-model into memory through the model loading module, enabling on-demand retrieval of the sub-model. Simultaneously, a unique mapping relationship between each level of sub-model and its corresponding disease type must be clearly defined. That is, one disease type corresponds to only one dedicated risk prediction sub-model, and one sub-model is used only for risk prediction of one disease type (e.g., cerebrovascular disease labeled 1 corresponds only to sub-model 1, and sub-model 1 is not used for prediction of other disease types). This ensures the accuracy of model calls and avoids situations where the model and disease type do not match.

[0051] A multi-sub-model management module is built, which forms a bidirectional data connection with the multi-sub-model hierarchical brain disease risk prediction system. This module possesses several core functions: Parameter retrieval allows the module to quickly obtain the current weight parameters, bias parameters, and other configuration information of each risk prediction sub-model through an index library, facilitating model status monitoring; Model update supports the module in retraining and adjusting the parameters of each risk prediction sub-model based on newly collected feature data sets and updated association pattern datasets (the update process is consistent with the initial training, and the updated model version number increments), achieving iterative optimization of the model and improving its adaptability to new data and patterns; Performance monitoring can periodically (every 7 days) monitor the running status, prediction accuracy, response speed, and other indicators of each risk prediction sub-model, generating performance monitoring reports. When the accuracy of a model drops below 0.8, an automatic update process is triggered, promptly identifying issues such as parameter drift and performance degradation during model operation and triggering corresponding maintenance or update operations. The training logs (iteration count, loss function change curve, evaluation index results), weight parameters, and bias parameters of each risk prediction sub-model are synchronously stored in the database of the multi-sub-model management module (using a MySQL database) to achieve centralized management and backup of model-related data. The multi-sub-model management module establishes a data interaction connection with the feature data storage module, which can retrieve the latest data required for model updates from the feature data storage module to ensure the smooth progress of the model update process.

[0052] S4. Collect the data to be predicted and match it with the disease type, then call the sub-model to complete the risk assessment: This step focuses on practical risk prediction applications. By collecting relevant data on the object to be predicted, processing it, and matching it with the corresponding disease type and sub-model, the risk assessment calculation is finally completed and the results are output. Among them, "feature matching calculation" refers to the process of determining the most likely corresponding disease type by calculating the degree of fit between the data to be predicted and the specific features of each disease.

[0053] S4.1. Collect basic health characteristics and vital signs data of the subjects to be predicted, and perform standardization processing on the collected data according to the preprocessing standards in S1 to generate standardized data to be predicted: The collected basic health characteristics and vital signs data must cover key information related to brain diseases. The data type and dimensions must be consistent with the brain disease-related big data processed in S1 (including fields corresponding to 8 core features) to ensure effective matching with the specific pathogenesis characteristics of various types of brain diseases. Data collection is achieved through standardized collection forms, which include field names, data types, and collection requirements (e.g., "blood pressure value" must be collected as systolic blood pressure, in mmHg, retaining integer digits), to avoid deviations in the format or dimensions of the collected data. The standardization process strictly follows the preprocessing standards in S1, including format unification, missing value supplementation, and outlier identification and removal: First, the collected data is converted into a CSV format consistent with the feature data set (single data is stored on a single row, and the field order is consistent with the model input dimensions). Then, missing fields are supplemented using the same method as in S1 (e.g., missing "blood lipid indicators" are estimated using linear interpolation based on the average level of the same population). Finally, outliers that do not conform to the data pattern are identified and removed using the 3σ principle (e.g., "blood pressure" values ​​exceeding 180 mmHg or below 80 mmHg are considered abnormal and need to be re-collected or re-labeled). These operations generate standardized data to be predicted, ensuring that the data format and quality are consistent with the data used for model training, providing a data foundation for accurate prediction.

[0054] In some specific implementations, the following eight core feature data of a subject to be predicted were collected: blood pressure 135 mmHg, blood lipid level 4.2 mmol / L, smoking history 1 (yes), history of hypertension 1 (yes), age-related index 65 (corresponding to 65 years old), exercise frequency 2 (occasional exercise), sleep quality 6.5 (out of 10), and diet structure 3 (balanced diet). Among them, the "exercise frequency" field was complete, and the "diet structure" was quantified to 3. After format standardization and outlier detection, there was no abnormal data, and the standardized data to be predicted was generated as [135, 4.2, 1, 1, 65, 2, 6.5, 3], which perfectly matched the input dimensions of the model.

[0055] S4.2. Perform feature matching operations between the standardized data to be predicted and the specific pathogenesis characteristics of each type of brain disease, and determine the type of brain disease corresponding to the object to be predicted based on the operation results: Feature matching operations employ a combination of cosine similarity and Euclidean distance calculations to comprehensively assess the degree of fit between standardized data to be predicted and the specific pathogenesis features of various types of brain diseases. Cosine similarity measures the directional consistency between two vectors (the data vector to be predicted and the disease-specific feature vector), with a value range of [-1, 1]. A value closer to 1 indicates a higher degree of fit. The formula is: cosθ = (A·B) / (||A||×||B||), where A is the data vector to be predicted, and B is the mean vector of the disease-specific features (calculated based on the mean values ​​of each feature in the disease feature dataset). Euclidean distance measures the spatial distance between two vectors; a smaller value indicates a higher degree of fit. The formula is: d = √Σ(Ai-Bi)², where Ai is the i-th element of the data vector to be predicted, and Bi is the i-th element of the mean vector of the disease-specific features. During the calculation, the cosine similarity and Euclidean distance between the standardized data to be predicted and the mean vectors of the specific features of five types of brain diseases are first calculated. Then, both indicators are normalized (converted to the [0,1] interval). The overall matching degree is calculated with a weight of 0.6 for cosine similarity and 0.4 for Euclidean distance (overall matching degree = normalized cosine similarity value × 0.6 + (1 - normalized Euclidean distance value) × 0.4). The 1 - Euclidean distance normalization value is used because a smaller Euclidean distance indicates a higher degree of fit and needs to be converted into a positive indicator. The disease type with the highest overall matching degree is determined as the brain disease type corresponding to the object to be predicted. If the highest overall matching degree does not reach the preset threshold (0.6), it is necessary to indicate insufficient data or that the disease type cannot be clearly identified, ensuring the accuracy of disease type judgment and providing a reliable basis for subsequent sub-model calls.

[0056] S4.3. Retrieve the risk prediction sub-models corresponding to the matching results from the multi-sub-model hierarchical brain disease risk prediction system, and establish a computational connection between the standardized data to be predicted and the risk prediction sub-models: Based on the determined brain disease type (e.g., cerebrovascular disease with the highest comprehensive matching degree as label 1), a model call request is sent to the multi-sub-model level brain disease risk prediction system. The request includes the corresponding disease type label (1) and the identifier of the data to be predicted. After receiving the request, the prediction system parses the request parameters through the call interface, queries the internal sub-model index library, and quickly locates the corresponding risk prediction sub-model (sub-model 1) based on the disease type label 1. It then obtains the storage path (distributed storage address: hdfs: / / xxx / model / type1_v2) and call parameters (input data dimension 8, data type is floating point, output is risk probability). Subsequently, the network structure and parameters of sub-model 1 are loaded into memory through the model loading module (loading time does not exceed 500ms). A computational connection is established between the standardized data to be predicted and the risk prediction sub-model. Data is transmitted using TensorFlow data flow pipeline to ensure that the standardized data to be predicted can be smoothly input into the risk prediction sub-model through the preset transmission channel. At the same time, a data verification mechanism is used to ensure the integrity and security of data transmission (such as verifying data dimensions and data types to avoid data loss or interference), preparing for subsequent risk assessment calculations.

[0057] S4.4. Input the standardized data to be predicted into the established risk prediction sub-model, and generate the risk assessment results for the corresponding brain disease type of the object to be predicted through the forward propagation operation of the model: After the standardized data to be predicted is fed into the risk prediction sub-model through computation, the model initiates forward propagation: the data first enters the input layer of the model, is converted into tensor format (shape [1,8], corresponding to 1 data point and 8 features), and is passed to hidden layer 1 through the input layer nodes. The 64 neurons in hidden layer 1 perform linear transformation on the input data through a weight matrix (shape [8,64]), and then perform nonlinear transformation through the ReLU activation function to output a 64-dimensional feature vector; this feature vector is passed to hidden layer 2, and after operation with a weight matrix (shape [64,32]) and the ReLU activation function, a 32-dimensional feature vector is output; then it is passed to hidden layer 3, and after operation with a weight matrix (shape [32,16]) and the ReLU activation function, a 16-dimensional feature vector is output; finally, it is passed to the output layer, where a linear transformation is performed through a weight matrix (shape [16,1]), and then the result is mapped to the [0,1] interval through the Sigmoid activation function to obtain the risk probability of the subject suffering from this type of brain disease. Based on the preset risk level classification rules (risk probability < 0.3 is low risk, 0.3-0.7 is medium risk, and > 0.7 is high risk), a risk assessment result is generated. The result includes information such as risk level, risk probability, key influencing features (the three features with the highest weight values ​​in the model), and risk-related factors (corresponding correlation patterns). This information can provide data support for subsequent health guidance and intervention.

[0058] After completing the risk assessment calculation for the brain disease type corresponding to the object to be predicted, the risk assessment results are input into the result output module for structured processing. During structured processing, feature matching data (comprehensive matching degree, cosine similarity, Euclidean distance values) and parameter calculation data (output feature vectors of each hidden layer, and the original calculation results of the output layer) from the model calculations are retained. This data provides detailed basis for subsequent result tracking and model optimization. The result output module supports multiple structured output formats, including JSON and tabular (Excel) formats. The default output format is JSON, which includes fields such as "Identifier of Data to be Predicted," "Disease Type," "Risk Level," "Risk Probability," "Key Influencing Features," "Feature Matching Data," and "Parameter Calculation Data Summary." The result output module establishes a real-time data transmission connection with the risk prediction sub-model, synchronously receiving intermediate calculation data and final risk assessment results from the model, ensuring the timeliness and integrity of data transmission and avoiding impact on result processing due to data delays or loss. The results output module synchronously interacts with the multi-sub-model management module to exchange computational data (risk assessment results, model call logs, computation time). The multi-sub-model management module can monitor the real-time running status of the risk prediction sub-model (such as call frequency, average response time, and prediction accuracy) based on this computational data, analyze the prediction accuracy and stability of the model, and provide a reference for subsequent model maintenance, parameter adjustment, and upgrades.

[0059] This embodiment provides a brain disease risk prediction method based on big data analysis. Through a standardized big data preprocessing workflow, it unifies the format of heterogeneous and incomplete raw data, fills in missing values, and removes outliers. It clarifies specific processing methods, thresholds, and operational steps, ensuring the integrity, validity, and consistency of the input data, laying a solid data foundation for subsequent feature mining and model training. Through targeted feature mining operations, it employs specific feature screening algorithms, quantification methods, importance ranking tools, and association analysis algorithms to screen effective features from massive amounts of data, quantify non-numerical features, extract key pathogenesis features and risk factors, and analyze their correlation patterns. This ensures that model training has clear targeting and scientific rigor, avoiding interference from invalid features in the prediction results. By constructing risk prediction sub-models hierarchically, it clarifies the model's network structure parameters, loss function, optimizer parameters, and training process, training a dedicated prediction model for each type of brain disease. Combined with correlation weights and training on a complete dataset, it improves the model's accuracy for specific diseases. To improve prediction accuracy, a multi-sub-model hierarchical prediction system was built, clarifying the system's structure, index design, and calling interfaces. This enabled centralized management and on-demand retrieval of all sub-models, overcoming the limitation of a single model adapting to multiple disease predictions. Through the collaborative work of multiple functional modules, including feature data storage, weight calculation, multi-sub-model management, and result output modules, the implementation methods, data interaction methods, and core functions of each module were clarified. A closed-loop management system was established for the entire process of data storage, feature processing, model training, risk assessment, and result output, ensuring the efficiency and orderliness of the entire prediction process and improving its standardization and manageability. Simultaneously, the bidirectional data connection and synchronous data storage design between modules enabled full-process traceability and management of data, models, and results, facilitating subsequent method optimization, model updates, and functional expansion. Overall, this improved the accuracy, relevance, and operability of brain disease risk prediction, providing reliable technical support for early screening and risk warning of brain diseases.

[0060] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting the risk of brain diseases based on big data analysis, characterized in that, Includes the following steps: S1. Standardize the preprocessing of the collected big data related to brain diseases, remove invalid data, and then classify the data by label according to the specific type of brain disease to generate a feature data set corresponding to each type of brain disease. S2. Perform feature mining operations on the feature datasets corresponding to various types of brain diseases, extract the specific pathogenesis features and core risk factors of each type of brain disease, and sort out the correlation patterns between specific pathogenesis features and core risk factors through the correlation analysis module and generate a correlation pattern dataset. S3. Based on the feature data set corresponding to each type of brain disease, combined with the extracted specific pathogenesis features, core risk factors and correlation pattern datasets, corresponding risk prediction sub-models are trained for each type of brain disease for risk feature matching and evaluation calculation of each type of brain disease. A multi-sub-model hierarchical brain disease risk prediction system is built to integrate the risk prediction sub-models corresponding to each type of brain disease and realize the on-demand retrieval of risk prediction sub-models. Each risk prediction sub-model forms a unique mapping connection relationship with the corresponding brain disease type. S4. Collect basic health characteristics and vital signs data of the object to be predicted, match the corresponding brain disease type, and then call the risk prediction sub-model corresponding to the type from the multi-sub-model level brain disease risk prediction system. Input the basic health characteristics and vital signs data into the risk prediction sub-model to complete the risk assessment calculation for the corresponding brain disease type.

2. The method according to claim 1, characterized in that, Step S1 includes the following sub-steps: S1.

1. Perform unified format processing on brain disease-related big data, converting heterogeneous storage formats of brain disease-related big data into a preset unified data format; S1.

2. Perform missing value supplementation processing on the brain disease-related big data after the format is unified, and use interpolation method to improve the brain disease-related big data with missing data; S1.

3. Perform outlier identification and removal on the brain disease-related big data after missing value completion to generate preprocessed brain disease-related big data; S1.

4. Based on the clinical classification standards for brain diseases, add disease type labels to the preprocessed big data related to brain diseases, complete the labeling classification, and generate feature data sets corresponding to each type of brain disease.

3. The method according to claim 1, characterized in that, Step S2 includes the following sub-steps: S2.

1. Perform initial feature screening on the feature datasets corresponding to various types of brain diseases, remove redundant feature data that are not related to the onset of brain diseases, and retain valid feature data. S2.

2. Perform feature quantization processing on the retained effective feature data to convert non-numerical effective feature data into numerical effective feature data; S2.

3. Rank the numerical effective feature data by feature importance, and extract the top-ranked feature data as the specific pathogenesis characteristics and core risk factors of various types of brain diseases; S2.

4. Input the extracted specific pathogenesis features and core risk factors into the association analysis module. The module will then mine the association between the two, organize them into a corresponding association pattern dataset according to the type of brain disease, and output it synchronously.

4. The method according to claim 1, characterized in that, Step S3 includes the following sub-steps: S3.

1. Divide the feature datasets corresponding to each type of brain disease into training datasets and validation datasets. Both datasets contain complete data dimensions of specific pathogenesis features, core risk factors, and correlation patterns. S3.

2. Build the network structure of the initial prediction model. The initial prediction model is used for subsequent training to obtain risk prediction sub-models corresponding to various types of brain diseases. Set the loss function and optimizer parameters for model training. Iteratively train the initial prediction model using the training dataset and dynamically adjust the model's weight parameters and bias parameters. S3.

3. Input the validation dataset into the trained initial prediction model, verify the model's fit and generalization ability through the validation module, and generate sub-models for risk prediction of various types of brain diseases for pre-calculation of risk assessment of various types of brain diseases. S3.

4. Retain the validated risk prediction sub-models to be validated, classify the sub-models according to the types of brain diseases, and build a multi-level brain disease risk prediction system composed of multi-level sub-models to integrate the risk prediction sub-models corresponding to various types of brain diseases and realize the on-demand retrieval of risk prediction sub-models, and confirm the unique mapping connection between each level of sub-model and the corresponding disease type.

5. The method according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4.

1. Collect basic health characteristics and vital signs data of the subject to be predicted, and perform standardized processing on the collected data according to the preprocessing standard of S1 to generate standardized data to be predicted; S4.

2. Perform feature matching operation between the standardized data to be predicted and the specific pathogenesis characteristics of each type of brain disease, and determine the type of brain disease corresponding to the object to be predicted based on the operation results; S4.

3. Retrieve the risk prediction sub-models corresponding to the matching results from the multi-sub-model hierarchical brain disease risk prediction system, and establish a computational connection between the standardized data to be predicted and the risk prediction sub-models; S4.

4. Input the standardized data to be predicted into the established operational connection of the risk prediction sub-model, and generate the risk assessment results of the brain disease type corresponding to the object to be predicted through the forward propagation operation of the model.

6. The method according to claim 1, characterized in that, In step S1, after completing the labeling and classification, a feature data storage module is built. The feature data sets corresponding to each type of brain disease are stored in the independent data partitions of the feature data storage module. A unique disease type identifier is set for each data partition to achieve partitioned and isolated storage of each feature data set. The feature data storage module establishes a real-time data transmission connection with the feature mining operation after the preprocessing in S1.

7. The method according to claim 6, characterized in that, In step S2, when sorting out the correlation between specific pathogenesis characteristics and core risk factors, the weight calculation module configures the correlation weight value for each group of correlations, and the correlation weight value and the corresponding correlation pattern are entered into the correlation pattern dataset. The correlation pattern dataset establishes a one-to-one correlation connection with the feature data set of each type of brain disease, and the calculation data of the weight calculation module is synchronously stored in the feature data storage module.

8. The method according to claim 7, characterized in that, In step S3, a multi-sub-model management module is built. This module forms a bidirectional data connection with the multi-sub-model hierarchical brain disease risk prediction system. The multi-sub-model management module performs independent parameter retrieval, model updates, and performance testing for each risk prediction sub-model. The training logs, weight parameters, and bias parameters of each risk prediction sub-model are synchronously stored in the multi-sub-model management module. The multi-sub-model management module establishes a data interaction connection with the feature data storage module.

9. The method according to claim 8, characterized in that, In step S4, after the risk assessment calculation is completed, the risk assessment result is input into the result output module for structured processing. During the structured processing, the feature matching data and parameter calculation data of the model calculation are retained. The result output module establishes a real-time data transmission connection with the risk prediction sub-model, synchronously receiving the intermediate calculation data and the final risk assessment result of the model. The result output module also synchronously interacts with the multi-sub-model management module to exchange calculation data.

10. A brain disease risk prediction system based on big data analysis, used to execute the brain disease risk prediction method based on big data analysis as described in any one of claims 1-9, characterized in that, It includes a data preprocessing module, a feature mining module, a model training module, a risk assessment module, a feature data storage module, a weight calculation module, a multi-sub-model management module, and a result output module; The data preprocessing module is unidirectionally connected to the feature data storage module. The data preprocessing module outputs the generated feature data sets corresponding to various types of brain diseases to the feature data storage module for partitioned and isolated storage. The feature mining module is bidirectionally connected to both the feature data storage module and the weight calculation module. The feature mining module retrieves the feature data sets from the feature data storage module, and the association weight values ​​configured by the weight calculation module for the association relationships are synchronized to the feature mining module and entered into the association pattern dataset. The model training module is bidirectionally connected to the feature data storage module, the weight calculation module, and the multi-sub-model management module. The model training module trains corresponding risk prediction sub-models for each type of brain disease based on the feature data sets and the association pattern dataset. The system performs risk feature matching and assessment calculations for various types of brain diseases. The multi-sub-model management module manages each risk prediction sub-model throughout the entire process and synchronously stores model-related data. The risk assessment module is bidirectionally connected to both the multi-sub-model management module and the result output module. The risk assessment module retrieves the corresponding risk prediction sub-model from the multi-sub-model management module to complete the risk assessment calculation and outputs the calculation results to the result output module. The result output module maintains real-time data transmission with the risk assessment module, performs structured processing on the risk assessment results, and retains all calculation data. The model training module builds a multi-sub-model hierarchical brain disease risk prediction system to integrate risk prediction sub-models corresponding to various types of brain diseases and realizes on-demand retrieval of risk prediction sub-models.

Images