Cerebral stroke screening data preprocessing method and system

Abstract

The invention discloses a cerebral apoplexy screening data preprocessing method and system, and aims to solve the problems of difficult fusion of multi-source heterogeneous data, low data quality, feature redundancy and the like in cerebral apoplexy screening. The method comprises the following steps: collecting the multi-source heterogeneous data, and encrypting and desensitizing the multi-source heterogeneous data to form an initial data set; carrying out standardized alignment on the multi-modal data, and constructing a data set in a unified format; missing values are filled through a mixing strategy, abnormal values are detected and corrected, and noise is smoothed to enhance data quality; extracting and screening key features in combination with a medical knowledge graph and an algorithm; and dividing the data set to train the evaluation prediction model. The system comprises a data acquisition and safety module, a data standardization module, a data quality enhancement module, a feature engineering module and a model preparation and evaluation module, and full-process automation is achieved. The method breaks information islands, improves data quality and feature effectiveness, reduces labor cost, can be deployed in a standardized manner, and provides efficient technical support for early screening of cerebral apoplexy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of stroke screening, in particular to a stroke screening data preprocessing method and system. BACKGROUND

[0002] Stroke (commonly known as "stroke") is an acute cerebrovascular disease, which has the characteristics of high incidence, high disability rate, high mortality and high recurrence rate. Early screening and risk assessment are crucial for preventing and reducing the occurrence of stroke. With the development of medical informatization, medical institutions have accumulated a large amount of patient data, including electronic medical records, medical images, laboratory test results, etc., which makes it possible to build stroke risk prediction models based on artificial intelligence.

[0003] However, before using these data to build models, there are serious data preprocessing challenges. First, the data sources are diverse (HIS, LIS, PACS, wearable devices, etc.), with heterogeneous formats, different standards, and serious "information island" phenomenon, which makes it difficult to be directly fused. Second, the quality of the original medical data is generally low, with a large number of missing values, outliers (such as negative blood pressure values) and noise (such as ECG signal interference), which will cause the model performance to drop sharply if used directly. Third, the feature dimension related to stroke is high and complex, which contains a large number of irrelevant or redundant features, not only increasing the computational overhead, but also causing "dimension disaster", leading to model overfitting.

[0004] Currently, the existing data preprocessing methods are mostly general techniques, lacking special optimization for the stroke screening scenario. For example, simple mean imputation of missing values may introduce bias; feature selection only relies on statistical methods and ignores important medical prior knowledge. Therefore, there is an urgent need in the field for a preprocessing method and system specifically for stroke screening that can systematically solve the above-mentioned problems of multi-source heterogeneous, low-quality, high-dimensional data. SUMMARY

[0005] (I) Technical problem

[0006] The present application provides a stroke screening data preprocessing method and system, aiming to solve the technical problems in the background art.

[0007] (II) Technical content

[0008] To solve the above technical problems, the technical solution of the present application is as follows: a stroke screening data preprocessing method, comprising the following steps:

[0009] S1: Collecting original data related to stroke risk from multiple heterogeneous data sources, and performing encrypted transmission and desensitization processing on the original data to form an initial data set;

[0010] S2: performing multimodal data standardization and alignment processing on the initial data set to construct a standardized data set in a unified format;

[0011] S3: performing data quality enhancement processing on the standardized data set, including missing value filling, outlier detection and correction, and noise data smoothing;

[0012] S4: based on a medical prior knowledge graph and a feature engineering algorithm, extracting and screening a key feature set for stroke risk prediction from the quality-enhanced data;

[0013] S5: dividing the key feature set into a training set and a test set, training a pre-set stroke risk prediction model using the training set, evaluating the trained model using the test set, and finally outputting preprocessed data and a model that can be used for clinical screening.

[0014] Further, the plurality of heterogeneous data sources in step S1 include: structured electronic medical record data in a hospital information system, unstructured image reports in a medical image archive and communication system, laboratory test data in a laboratory information management system, and continuous physiological time series data collected by wearable devices.

[0015] Further, the missing value filling in step S3 adopts a hybrid filling strategy: for numerical features, a filling method based on a k-nearest neighbor algorithm is used; for categorical features, a filling method based on inter-feature association rules is used; for randomly missing data, multiple imputation method is used to generate multiple complete data sets.

[0016] Further, the step S4 specifically includes:

[0017] S41: based on a medical knowledge graph, a feature pool containing stroke risk factors and their associated weights is constructed;

[0018] S42: using a recursive feature elimination algorithm, a feature subset with high importance ranking is preliminarily screened from the feature pool;

[0019] S43: using a maximum information coefficient algorithm to calculate the non-linear correlation between each feature in the feature subset and the stroke label, and further removing redundant features;

[0020] S44: combining and deriving the screened features to generate high-order features containing interaction terms to form the final key feature set.

[0021] Another aspect of the present application provides a stroke screening data preprocessing system, comprising a data acquisition and security module, a data standardization module, a data quality enhancement module, a feature engineering module and a model preparation and evaluation module connected in sequence.

[0022] Furthermore, the feature engineering module has a built-in knowledge graph database that stores the medical relationships between risk factors, diseases, symptoms, and medications for stroke; the feature engineering module calls a feature filtering algorithm engine to perform feature extraction and filtering processes.

[0023] (III) Technical Effects

[0024] The advantages of this invention compared to the prior art are:

[0025] 1. Multimodal data standardization and alignment technology effectively breaks down the "information silos" between various systems within the hospital, integrating scattered and heterogeneous data into a unified and well-organized high-value dataset, laying a solid data foundation for subsequent analysis.

[0026] 2. Hybrid missing value imputation, intelligent anomaly detection and correction, and targeted denoising algorithms can systematically and meticulously improve data quality, thereby directly enhancing the accuracy and robustness of subsequent risk prediction models.

[0027] 3. Combining medical knowledge graphs with statistical feature selection algorithms. This method not only utilizes data-driven objective laws but also incorporates profound prior medical knowledge, resulting in a selection of key feature sets that not only have strong predictive capabilities but also possess clear clinical significance, thereby enhancing doctors' trust in the model.

[0028] 4. The methodology was systematized, and a fully automated system was built, encompassing data acquisition and model preparation. This not only significantly reduced the time and cost of manual intervention and minimized human error, but also made the technical solution easy to standardize and promote in medical institutions at all levels, demonstrating high practical value and commercial potential. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating a stroke screening data preprocessing method according to the present invention.

[0030] Figure 2 This is a schematic diagram of a stroke screening data preprocessing system according to the present invention. Detailed Implementation

[0031] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "provided with," "installed," "connected," "linked," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0032] The present invention will now be described in further detail with reference to the accompanying drawings.

[0033] Combined with appendix Figure 1 To be continued Figure 2 A method for preprocessing stroke screening data includes the following steps:

[0034] S1: Collect raw data related to stroke risk from multiple heterogeneous data sources, and encrypt and desensitize the raw data to form an initial dataset. The multiple heterogeneous data sources include: structured electronic medical record data in hospital information systems, unstructured image reports in medical image archiving and communication systems, test data in laboratory information management systems, and continuous physiological time-series data collected by wearable devices.

[0035] S2: Perform multimodal data standardization and alignment on the initial dataset to construct a standardized dataset with a unified format. The multimodal data standardization and alignment specifically includes:

[0036] S21: For structured data, establish a unified data dictionary to map similar data from different sources to standard terms;

[0037] S22: For unstructured text data, natural language processing technology is used to perform entity recognition and relation extraction, transforming it into structured data;

[0038] S23: Resample and align time-series data with timestamps to ensure all data is on a unified time base;

[0039] S24: Generate a globally unique patient identifier for each record to enable the association and fusion of the same patient data from different data sources.

[0040] S3: Perform data quality enhancement processing on the standardized dataset, including missing value imputation, outlier detection and correction, and noisy data smoothing. The missing value imputation adopts a hybrid imputation strategy: for numerical features, an imputation method based on the k-nearest neighbor algorithm is used; for categorical features, an imputation method based on the association rules between features is used; for randomly missing data, multiple imputation methods are used to generate multiple complete datasets.

[0041] S4: Based on medical prior knowledge graphs and feature engineering algorithms, a key feature set for stroke risk prediction is extracted and screened from the enhanced data, specifically including:

[0042] S41: Based on medical knowledge graph, construct a feature pool containing stroke risk factors and their associated weights;

[0043] S42: Using a recursive feature elimination algorithm, a subset of features with high importance ranking is initially selected from the feature pool;

[0044] S43: Calculate the nonlinear correlation between each feature in the feature subset and the stroke label using the maximum information coefficient algorithm, and further eliminate redundant features;

[0045] S44: Combine and derive the selected features to generate higher-order features containing interaction terms, forming the final key feature set.

[0046] S5: Divide the key feature set into a training set and a test set, use the training set to train the preset stroke risk prediction model, use the test set to evaluate the trained model, and finally output preprocessed data and model that can be used for clinical screening.

[0047] Another aspect of the present invention provides a stroke screening data preprocessing system, comprising a data acquisition and security module, a data standardization module, a data quality enhancement module, a feature engineering module, and a model preparation and evaluation module connected in sequence.

[0048] The output of the data acquisition and security module is connected to the input of the data standardization module to provide an encrypted and de-identified initial dataset; the output of the data standardization module is connected to the input of the data quality enhancement module to provide a standardized dataset in a unified format; the output of the data quality enhancement module is connected to the input of the feature engineering module to provide a cleaned and enhanced high-quality dataset; the output of the feature engineering module is connected to the input of the model preparation and evaluation module to provide an optimized set of key features; the model preparation and evaluation module receives the set of key features, performs data partitioning, model training and evaluation, and finally outputs preprocessed data and a usable risk assessment model.

[0049] The data acquisition and security module includes: a multi-source adaptation submodule, configured with API interfaces and data parsers for interfacing with different data source systems; a secure transmission submodule, which uses national cryptographic algorithms to encrypt the transmission channel; and a data desensitization submodule, used to perform irreversible hashing on direct identifiers such as patient names and ID numbers, and to perform generalization processing on quasi-identifiers such as dates.

[0050] The data quality enhancement module integrates: a missing value processing unit for executing the hybrid imputation strategy according to claim 4; an anomaly detection unit for detecting numerical anomalies using the isolated forest algorithm and logical anomalies based on a medically reasonable range; and a noise filtering unit for smoothing and denoising time-series physiological signals using a wavelet transform algorithm.

[0051] The feature engineering module has a built-in knowledge graph database that stores the medical relationships between risk factors, diseases, symptoms, and medications for stroke. The feature engineering module calls the feature filtering algorithm engine to perform feature extraction and filtering processes.

[0052] The working principle of this invention is as follows: Addressing the pain points of data processing in stroke screening scenarios, this invention constructs a comprehensive and professional preprocessing technology solution. Regarding data source control, this solution achieves secure collection and integration of multi-source heterogeneous data. By adapting to dedicated API interfaces and data parsers of Hospital Information Systems (HIS), Picture Archiving and Communication Systems (PACS), Laboratory Information Systems (LIS), and wearable devices, it breaks down the "information silos" between systems. Simultaneously, it employs national cryptographic algorithms to encrypt the transmission channel and performs irreversible hashing and alignment identifier generalization processing on patient direct identifiers. This ensures data transmission security and patient privacy while achieving multi-source data collection.

[0053] In the data standardization phase, the solution innovatively adopts a unified multimodal data organization strategy. For structured data, a unified data dictionary is established to map standard terms for similar data. For unstructured text data such as image reports, natural language processing technology is used to complete entity recognition and relation extraction, transforming it into a structured format. For time-series physiological data generated by wearable devices, resampling and timestamp alignment are used to ensure a unified time reference. Combined with a globally unique patient identifier, the solution achieves the association and fusion of multi-source data for the same patient, forming a standardized dataset with a unified format.

[0054] In terms of data quality improvement, the solution constructs a refined hybrid processing mechanism. Missing value imputation adopts a classification strategy: k-nearest neighbor algorithm for numerical features, feature association rules for categorical features, and multiple imputation method for randomly missing data; outlier detection combines the isolated forest algorithm (for numerical anomalies) with medical reasonable range judgment (for logical anomalies, such as negative blood pressure values); and wavelet transform algorithm is used to smooth and denoise time-series physiological signal noise, comprehensively improving data quality and laying the foundation for subsequent analysis.

[0055] In the feature engineering phase, the solution pioneers a feature optimization model that combines medical knowledge with algorithms. It internally stores a medical knowledge graph containing relationships between stroke risk factors, diseases, symptoms, and medications, and constructs a feature pool with risk factors and associated weights. First, a recursive feature elimination algorithm is used to select the most important feature subset. Then, a maximum information coefficient algorithm is used to calculate the non-linear correlation between features and stroke labels to eliminate redundant features. Finally, higher-order features with interaction terms are generated through feature combination, forming a key feature set that combines predictive power and clinical significance.

[0056] In terms of the overall process, the solution achieves end-to-end automation from data acquisition to model output. Through modular design (data acquisition and security, data standardization, data quality enhancement, feature engineering, model preparation and evaluation), the key feature set is automatically divided into training and testing sets to complete model training and performance evaluation. The final output is preprocessed data and risk assessment models that can be directly used for clinical screening, significantly reducing manual intervention and human error. At the same time, it has the practicality for standardized deployment in medical institutions at all levels, providing efficient and professional technical support for early stroke screening.

[0057] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for preprocessing stroke screening data, characterized in that, Includes the following steps: S1: Collect raw data related to stroke risk from multiple heterogeneous data sources, and encrypt and de-identify the raw data to form an initial dataset; S2: Perform multimodal data standardization and alignment on the initial dataset to construct a standardized dataset with a unified format; S3: Perform data quality enhancement processing on the standardized dataset, including missing value imputation, outlier detection and correction, and noisy data smoothing; S4: Based on medical prior knowledge graphs and feature engineering algorithms, extract and filter key feature sets for stroke risk prediction from the enhanced data; S5: Divide the key feature set into a training set and a test set, use the training set to train the preset stroke risk prediction model, use the test set to evaluate the trained model, and finally output preprocessed data and model that can be used for clinical screening.

2. The method for preprocessing stroke screening data according to claim 1, characterized in that, The multiple heterogeneous data sources in step S1 include: structured electronic medical record data in the hospital information system, unstructured image reports in the medical image archiving and communication system, test data in the laboratory information management system, and continuous physiological time-series data collected by wearable devices.

3. The method for preprocessing stroke screening data according to claim 1, characterized in that, The missing value imputation in step S3 adopts a hybrid imputation strategy: for numerical features, an imputation method based on the k-nearest neighbor algorithm is used; For categorical features, an imputation method based on the association rules between features is used; for randomly missing data, a multiple imputation method is used to generate multiple complete datasets.

4. The method for preprocessing stroke screening data according to claim 1, characterized in that, Step S4 specifically includes: S41: Based on medical knowledge graph, construct a feature pool containing stroke risk factors and their associated weights; S42: Using a recursive feature elimination algorithm, a subset of features with high importance ranking is initially selected from the feature pool; S43: Calculate the nonlinear correlation between each feature in the feature subset and the stroke label using the maximum information coefficient algorithm, and further eliminate redundant features; S44: Combine and derive the selected features to generate higher-order features containing interaction terms, forming the final key feature set.

5. A stroke screening data preprocessing system, used to implement the method as described in any one of claims 1-4, characterized in that, It includes a data acquisition and security module, a data standardization module, a data quality enhancement module, a feature engineering module, and a model preparation and evaluation module, which are connected in sequence.

6. The stroke screening data preprocessing system according to claim 5, characterized in that, The feature engineering module has a built-in knowledge graph database that stores the medical relationships between risk factors, diseases, symptoms, and medications for stroke. The feature engineering module calls the feature filtering algorithm engine to perform feature extraction and filtering processes.

Images