An early screening system for chronic obstructive pulmonary disease based on blood routine test data

Abstract

An early screening system for chronic obstructive pulmonary disease based on blood routine test data. It belongs to the technical field of digital medical treatment, and specifically relates to the technical field of intelligent early screening of chronic obstructive pulmonary disease. It solves the technical problem that the generalization ability of a COPD prediction model is limited due to differences between different research data sets. The system comprises a data collection module that collects blood routine test data of a plurality of chronic obstructive pulmonary disease patients and blood routine test data of a plurality of normal personnel; a data preprocessing module that performs standardization processing on the blood routine test data to obtain a standardized data set; a screening model training module that trains N classification models using the standardized data set and calculates the prediction accuracy of each classification model; and a screening model application module that obtains the fusion weight of each model by using a normalization processing method based on the prediction accuracy of each classification model, fuses the risk scores predicted by the multiple models, and obtains a final screening result.

Description

Technical Field

[0001] This invention belongs to the field of digital medical technology, specifically relating to the field of intelligent early screening technology for COPD. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a chronic airway disease characterized by persistent airflow limitation, primarily manifesting as chronic cough, sputum production, and shortness of breath. Globally, COPD is a leading cause of death and disability. Its early symptoms are often subtle, leading to low diagnosis rates, and once lung function is impaired, the damage is irreversible. Traditional COPD diagnosis relies on pulmonary function testing, but this method is costly, complex, and difficult to implement in primary healthcare institutions.

[0003] Addressing the challenges of COPD presents numerous difficulties. The high cost of pulmonary function testing equipment deters many primary healthcare institutions from acquiring such equipment. Furthermore, operating this equipment requires professional training, which is often lacking in primary healthcare institutions, further limiting the widespread adoption of pulmonary function testing at the grassroots level. Moreover, because early COPD symptoms are often atypical and difficult for patients to detect, by the time noticeable symptoms appear, lung function may have already suffered significant damage, missing the optimal window for treatment.

[0004] Currently, early screening for COPD mainly relies on lung function tests and clinical symptom assessments, but these methods have the following limitations:

[0005] Lung function testing requires specialized equipment and operators, making large-scale deployment difficult. This is not only due to the high cost of the equipment but also because its maintenance and calibration require specialized knowledge and skills. In some remote or economically underdeveloped areas, even if the equipment is available, its proper operation and accurate testing cannot be guaranteed. Clinical symptom assessment relies heavily on physician experience, is highly subjective, and is prone to missed or misdiagnosis. Different doctors may have different criteria for judging symptoms, and some mild symptoms may be overlooked by patients or misinterpreted by doctors as symptoms of other common diseases, leading to delayed detection of COPD. Furthermore, data utilization is insufficient. Blood routine examination data, as a routine physical examination item, contains rich health information but has not been fully utilized for early COPD screening. Although blood routine data is readily available, its potential relationship with COPD is complex and difficult to interpret and utilize accurately.

[0006] In recent years, machine learning and big data analytics technologies have made significant progress in the field of disease prediction. For example:

[0007] Disease prediction based on blood data: Existing studies have used routine blood test data to build machine learning models for predicting the risk of chronic diseases and cancer. However, these studies face the challenge of selecting data features when applied to COPD prediction. Routine blood test data contains numerous indicators, and determining which indicators are closely related to early COPD lesions is a challenge. Furthermore, differences between different study datasets may limit the generalization ability of the models.

[0008] COPD Prediction Models: Some studies have constructed early screening models for COPD by combining clinical data and machine learning algorithms. However, these models often rely on complex, multi-source data, making them difficult to apply in primary healthcare institutions. When integrating multi-source data, data consistency and accuracy are difficult to guarantee, and different data sources may have different formats and standards, requiring extensive data preprocessing. Summary of the Invention

[0009] To address the technical problem that existing early COPD screening models often rely on complex, multi-source data, and the differences between different research datasets can limit the model's generalization ability, this invention provides an early screening system for COPD based on routine blood test data. The system includes:

[0010] S1, Data Collection Module: Collects blood routine test data from several COPD patients and blood routine test data from several healthy individuals;

[0011] S2, Data Preprocessing Module: Standardizes the blood routine test data to obtain a standardized dataset;

[0012] S3, Screening Model Training Module: Trains N classification models using a standardized dataset and calculates the prediction accuracy of each classification model;

[0013] S4. Screening Model Application Module: Based on the prediction accuracy of each classification model, the fusion weight of each model is obtained by normalization, and the risk scores of multiple model predictions are fused to obtain the final screening result.

[0014] Furthermore, the standardization process for the blood routine test data specifically involves:

[0015] S21. An improved K-nearest neighbor filling algorithm is used to process missing feature values;

[0016] S22. Use the interquartile range method to detect and remove extreme feature values;

[0017] S23. Perform Z-score standardization on all feature values ​​so that the mean of each feature is 0 and the standard deviation is 1.

[0018] S24. The SMOTE method is introduced to oversample minority class features and generate new virtual feature values ​​through interpolation, thereby expanding the minority class feature space.

[0019] Furthermore, step S21 specifically includes:

[0020] Step 1: Dynamic feature space distance calculation: For feature values ​​with missing values, the distance between samples is measured by Euclidean distance. The missing items are reconstructed by weighting similar samples to obtain a preliminary incomplete sample set.

[0021] Step 2: Inverse distance weighted interpolation calculation: Based on the information of the complete samples in the preliminary sample set, the missing values ​​are estimated again in a more refined manner, and the inverse distance weighted interpolation method is used to complete the missing features to obtain the reconstructed sample set;

[0022] Step 3: Sample Identification: For the locations in the reconstructed sample set where missing values ​​once existed, mark them by setting binary mask bits to obtain a reconstructed sample set with mask marking.

[0023] Furthermore, the classification models include multilayer perceptron models, random forest models, and XGBoost models.

[0024] Furthermore, the specific steps for training the N classification models are as follows:

[0025] S31. Divide the standardized dataset into a training set, a validation set, and a test set;

[0026] S32. Train N classification models sequentially using the training set;

[0027] S33. Use the validation set to optimize and adjust the hyperparameters of the N classification models after training.

[0028] S34. Use the test set to perform blind testing on N classification models to verify their generalization ability and test the initial prediction accuracy of different models.

[0029] S35. Remove the feature columns that were masked as "previously missing" from the training set, validation set, and test set, and construct a comparison training set, a comparison validation set, and a comparison test set.

[0030] S36. Repeat the process of S32-S34 using a comparison training set, a comparison validation set, and a comparison test set to obtain the comparative prediction accuracy of different models.

[0031] Furthermore, the prediction accuracy is achieved through... Obtain, among which, Indicates the accuracy of the prediction. Represents the balance coefficient. Indicates the initial prediction accuracy. This represents the difference between the initial prediction accuracy and the comparison prediction accuracy.

[0032] Furthermore, based on the prediction accuracy of each classification model, a normalization process is used to obtain the fusion weights of each model, denoted as . And satisfy Based on the aforementioned fusion weights, a weighted average is calculated for the risk scores predicted by each classification model. The specific calculation method is as follows: ,in Indicates the first The risk score predicted by the classification model This represents the final risk score after fusion.

[0033] The beneficial effects of the system described in this invention are as follows: It proposes an early screening method for COPD based on routine blood test data, which has the following characteristics:

[0034] Low cost and easy to promote: It utilizes routine blood test data, requires no additional testing equipment, and is suitable for use in primary healthcare institutions.

[0035] High precision and high efficiency: Through machine learning algorithms, COPD screening and risk prediction are automated, improving the early diagnosis rate.

[0036] To overcome the limitation of model generalization ability caused by the differences between different blood routine datasets: Since there may be some differences in the blood routine test items of different testing institutions or hospitals in different regions, the generalization ability of the model may be limited. To overcome this deficiency, the system described in this invention adopts a dual difference elimination mechanism. The first mechanism is to determine the final prediction accuracy by fusing the initial prediction accuracy and the comparative prediction accuracy. The existence of comparative prediction accuracy is to balance the impact of missing data sources on the model's prediction accuracy, thereby more objectively evaluating the model's prediction accuracy. The second mechanism is to perform a weighted average of the output results of each classification model, fusing the prediction results of multiple models to obtain the final screening result. Since each model has different prediction characteristics for different data, a weighted fusion method is finally adopted to maximize the fusion of the prediction characteristics of different models to adapt to different blood routine data. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the workflow of an early screening system for COPD based on routine blood test data, as described in this invention. Detailed Implementation

[0038] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0039] Example 1

[0040] This embodiment provides an early screening system for COPD based on routine blood test data, such as... Figure 1 The diagram shown is a flowchart of the system's workflow. The system includes:

[0041] S1, Data Collection Module: Collects blood routine test data from several COPD patients and blood routine test data from several healthy individuals;

[0042] S2, Data Preprocessing Module: Standardizes the blood routine test data to obtain a standardized dataset;

[0043] S3, Screening Model Training Module: Trains N classification models using a standardized dataset and calculates the prediction accuracy of each classification model;

[0044] S4. Screening Model Application Module: Based on the prediction accuracy of each classification model, the fusion weight of each model is obtained by normalization, and the risk scores of multiple model predictions are fused to obtain the final screening result.

[0045] Example 2

[0046] This embodiment further defines Embodiment 1 and describes the working method of the system. It acquires the patient's complete blood count data, including white blood cell count (WBC), red blood cell count (RBC), and platelet count (PLT). During data collection, challenges arise from the diversity of data sources and inconsistent data quality. Different medical institutions may use different testing equipment and methods, making it difficult to guarantee data consistency. Furthermore, some data may contain missing or outlier values, requiring effective processing.

[0047] To preserve the structural integrity of the data to the greatest extent, we improve upon the existing K-nearest neighbor filling algorithm. The specific method of the improvement is as follows:

[0048] Step 1: Dynamic feature space distance calculation: For feature values ​​with missing values, the distance between samples is measured by Euclidean distance. The missing items are reconstructed by weighting similar samples to obtain a preliminary incomplete sample set.

[0049] Step 2: Inverse Distance Weighted Interpolation Calculation: Based on the information of the complete samples in the preliminary sample set, the missing values ​​are estimated again in a more refined manner. The inverse distance weighted interpolation method (IDW) is used to complete the missing features and obtain the reconstructed sample set.

[0050] Step 3: Sample Identification: For the locations in the reconstructed sample set where missing values ​​once existed, mark them by setting binary mask bits to obtain a reconstructed sample set with mask marking.

[0051] The method employed calculates the feature distance between samples and estimates the values ​​of neighboring samples using an inverse distance-weighted average. Simultaneously, specific representation bits are used to record interpolated samples, limiting their impact on the model's training and validation processes. Applying the missing value preprocessing method designed in this invention to the current modeling process for high-dimensional features and medical data with significant differences in feature quality ultimately improved model accuracy by 5-8%. Secondly, outliers can interfere with model training, especially common in medical data. The interquartile range (IQR) method is used to detect and remove extreme values.

[0052] Furthermore, the different value ranges and scales of different features may lead to unbalanced learning in the model. Z-score standardization was applied to all features, making the mean of each feature 0 and the standard deviation 1, thereby accelerating model convergence and improving training performance.

[0053] Finally, to address the issue of imbalanced sample labels, direct training can lead to a bias in the model towards the majority class. To mitigate this problem, the SMOTE (Synthetic Minority Over-sampling Technique) method is introduced to oversample minority class samples. This method generates new dummy samples through interpolation, expanding the minority class sample space and effectively balancing the class distribution of the training data.

[0054] In terms of model construction, considering the complex distribution of medical data features and the varying capabilities of different models in capturing feature patterns, a multi-model joint modeling strategy was adopted. Three models—Multilayer Perceptron (MLP), Random Forest (FR), and XGBoost—were trained on the data, and their performance was evaluated on independent test sets. Unlike the traditional "single optimal model" selection approach, this method designed and implemented a performance-weighted fusion of model prediction results based on key metrics such as accuracy and AUC on the test set, achieving more robust discriminative inference. This method fully leverages the complementary advantages of each model, improving the system's robustness, generalization ability, and clinical adaptability in scenarios with imbalanced samples, feature noise, or missing features. It demonstrates significant innovation and practicality. The specific technical methods are as follows.

[0055] Training N classification models specifically involves:

[0056] S31. Divide the standardized dataset into a training set, a validation set, and a test set;

[0057] S32. Train N classification models sequentially using the training set;

[0058] S33. Use the validation set to optimize and adjust the hyperparameters of the N classification models after training.

[0059] S34. Use the test set to perform blind testing on N classification models to verify their generalization ability and test the initial prediction accuracy of different models.

[0060] S35. Remove the feature columns that were masked as "previously missing" from the training set, validation set, and test set, and construct a comparison training set, a comparison validation set, and a comparison test set.

[0061] S36. Repeat the process of S32-S34 using a comparison training set, a comparison validation set, and a comparison test set to obtain the comparative prediction accuracy of different models.

[0062] The prediction accuracy is achieved through Obtain, among which, Indicates the accuracy of the prediction. Represents the balance coefficient. Indicates the initial prediction accuracy. This represents the difference between the initial prediction accuracy and the comparison prediction accuracy.

[0063] Based on the prediction accuracy of each classification model, the fusion weights of each model are obtained using a normalization process, denoted as: And satisfy Based on the aforementioned fusion weights, a weighted average is calculated for the risk scores predicted by each classification model. The specific calculation method is as follows: ,in Indicates the first The risk score predicted by the classification model This represents the final risk score after fusion.

[0064] Risk Prediction: Based on the final risk score, patients are categorized into low-risk, medium-risk, and high-risk groups, and personalized health management recommendations are provided. Clinical experience and professional knowledge should be combined to interpret and apply the risk score appropriately. Furthermore, the model should be regularly updated and optimized to adapt to the needs of different periods and regions.

[0065] System Implementation: The system supports access via web and mobile devices, providing data upload, result query, and report generation functions for user convenience. During system implementation, user experience and interface design should be prioritized to ensure ease of operation and use. Simultaneously, robust system security measures should be implemented to prevent data leaks and malicious attacks.

[0066] Example 3

[0067] This embodiment further defines Embodiment 2 and explains the data details involved in Embodiment 2.

[0068] Regularization parameter (alpha) in LASSO regression:

[0069] Optimal range: 0.01-0.1

[0070] Results: When the regularization parameter alpha is between 0.01 and 0.1, LASSO regression can effectively screen out key features related to COPD while avoiding excessive sparsity.

[0071] Specific example: When alpha=0.05, the number of features selected by LASSO regression is 15, the model's AUC value is 0.84, the sensitivity is 0.81, and the specificity is 0.77.

[0072] The K value in feature selection (KNN interpolation):

[0073] Optimal range: 5-10

[0074] Results: When the K value is between 5 and 10, KNN interpolation can effectively handle missing values ​​in the data while avoiding the introduction of too much noise.

[0075] Specific implementation: When K=7, the dataset after KNN interpolation performs best in model training, with an AUC of 0.85, a sensitivity of 0.82, and a specificity of 0.78.

[0076] Model ensemble optimization:

[0077] Number of trees in the Random Forest algorithm (n_estimators):

[0078] Optimal range: 100-200

[0079] Results: The model's predictive performance is optimal when the number of trees is between 100 and 200. Too few trees may lead to underfitting, while too many trees may lead to overfitting and increased computational complexity.

[0080] Specific implementation example: When n_estimators=150, the model's AUC value is 0.85, sensitivity is 0.82, and specificity is 0.78.

[0081] Maximum depth (max_depth) in the Random Forest algorithm:

[0082] Optimal range: 10-20

[0083] Results: When the maximum depth is between 10 and 20, the model can maintain high prediction accuracy while avoiding overfitting.

[0084] Specific example: When max_depth=15, the model's AUC value is 0.84, sensitivity is 0.81, and specificity is 0.77.

[0085] Learning rate in XGBoost:

[0086] Optimal range: 0.01-0.1

[0087] Results: When the learning rate is between 0.01 and 0.1, the XGBoost model can gradually converge during training, avoiding overfitting.

[0088] Specific example: When learning_rate=0.05, the AUC value of the XGBoost model is 0.86, the sensitivity is 0.83, and the specificity is 0.79.

[0089] The number of hidden layer nodes (hidden_layer_sizes) in a multilayer perceptron (MLP):

[0090] Optimal range: 50-100

[0091] Results: When the number of hidden layer nodes is between 50 and 100, the MLP model can maintain high prediction accuracy while avoiding overfitting.

[0092] Specific example: When hidden_layer_sizes=75, the AUC value of the MLP model is 0.84, the sensitivity is 0.81, and the specificity is 0.77.

[0093] System performance optimization:

[0094] Database caching mechanism:

[0095] Optimal range: Cache size 100MB-500MB

[0096] Effect: When the cache size is between 100MB and 500MB, the system response speed is significantly improved, while avoiding excessive memory usage.

[0097] Specific example: When the cache size is 300MB, the system response time is reduced from 5 seconds to 2 seconds, significantly improving the user experience.

[0098] Number of replicas in Kubernetes automated deployment:

[0099] Optimal range: 3-5

[0100] Effect: When the number of replicas is between 3 and 5, the system can maintain stable operation under high concurrency while avoiding resource waste.

[0101] Specific implementation: When replicas=4, the system response time under high concurrency is kept within 2 seconds, and the system stability is significantly improved.

[0102] The above specific examples clearly demonstrate the effects of algorithm parameters, feature selection, model integration, and system performance optimization under different parameter values, providing a reference for practical applications.

Claims

1. A system for early screening of COPD based on routine blood test data, characterized in that, The system screens patients for COPD based on routine blood test data. The system includes: S1, Data Collection Module: Collects blood routine test data from several COPD patients and blood routine test data from several healthy individuals; S2, Data Preprocessing Module: Standardizes blood routine test data, completes missing feature values, and marks them by setting binary mask bits to obtain a standardized dataset; S3, Screening Model Training Module: Trains N classification models using a standardized dataset and calculates the prediction accuracy of each classification model; Training N classification models specifically involves: S31. Divide the standardized dataset into a training set, a validation set, and a test set; S32. Train N classification models sequentially using the training set; S33. Use the validation set to optimize and adjust the hyperparameters of the N classification models after training. S34. Use the test set to perform blind testing on N classification models to verify their generalization ability and test the initial prediction accuracy of different models. S35. Remove the feature columns that were masked as "previously missing" from the training set, validation set, and test set, and construct a comparison training set, a comparison validation set, and a comparison test set. S36. Repeat the process of S32-S34 using the comparison training set, comparison validation set, and comparison test set to obtain the comparison prediction accuracy of different models; The prediction accuracy is achieved through Obtain, among which, Indicates the accuracy of the prediction. Represents the balance coefficient. Indicates the initial prediction accuracy. This represents the difference between the initial prediction accuracy and the comparative prediction accuracy. S4. Screening Model Application Module: Based on the prediction accuracy of each classification model, the fusion weight of each model is obtained by normalization, and the risk scores of multiple model predictions are fused to obtain the final screening result.

2. The early screening system for COPD based on routine blood test data according to claim 1, characterized in that, The standardization process for routine blood test data specifically involves: S21. An improved K-nearest neighbor filling algorithm is used to process missing feature values; S22. Use the interquartile range method to detect and remove extreme feature values; S23. Perform Z-score standardization on all feature values ​​so that the mean of each feature is 0 and the standard deviation is 1. S24. The SMOTE method is introduced to oversample minority class features and generate new virtual feature values ​​through interpolation, thereby expanding the minority class feature space.

3. The early screening system for COPD based on routine blood test data according to claim 2, characterized in that, Step S21 specifically involves: Step 1: Dynamic feature space distance calculation: For feature values ​​with missing values, the distance between samples is measured by Euclidean distance. The missing items are reconstructed by weighting similar samples to obtain a preliminary incomplete sample set. Step 2: Inverse distance weighted interpolation calculation: Based on the information of the complete samples in the preliminary sample set, the missing values ​​are estimated again in a more refined manner, and the inverse distance weighted interpolation method is used to complete the missing features to obtain the reconstructed sample set; Step 3: Sample Identification: For the locations in the reconstructed sample set where missing values ​​once existed, mark them by setting binary mask bits to obtain a reconstructed sample set with mask marking.

4. The early screening system for COPD based on routine blood test data according to claim 3, characterized in that, The classification models include multilayer perceptron model, random forest model and XGBoost model.

5. The early screening system for COPD based on routine blood test data according to claim 4, characterized in that, Based on the prediction accuracy of each classification model, the fusion weights of each model are obtained using a normalization process, denoted as . And satisfy Based on the aforementioned fusion weights, a weighted average is calculated for the risk scores predicted by each classification model. The specific calculation method is as follows: ,in Indicates the first The risk score predicted by the classification model This represents the final risk score after fusion.

Images