Method for constructing chronic obstructive pulmonary disease screening model based on quantitative CT (Computed Tomography)

Abstract

The invention provides a method for constructing a chronic obstructive pulmonary disease screening model based on quantitative CT, and the method comprises the steps: carrying out the screening in a research population according to a preset standard, and obtaining a derivation queue and a first external verification queue; performing questionnaire, lung function examination and chest CT scanning analysis on the subjects in the derivative queue to obtain a training data set and an internal verification data set, performing clinical data collection, lung function examination and chest CT scanning on the subjects in the first external verification queue, and combining with a second external verification queue to obtain an external verification data set; performing single-mode modeling and multi-mode modeling according to the type of each data in the training data set, and verifying and evaluating the modeled model by using the internal verification data set to obtain an optimal model; and performing performance evaluation on the optimal model by using an external verification data set. According to the invention, the early screening, diagnosis and management level of the chronic obstructive pulmonary disease can be obviously improved, so that the life quality and prognosis of a patient are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of disease screening, and in particular to a method for constructing a chronic obstructive pulmonary disease screening model based on quantitative CT. Background Art

[0002] Underdiagnosis of chronic obstructive pulmonary disease (COPD) is a worldwide problem, especially in developing countries. According to a large epidemiological study in my country, about 60.2% of COPD patients did not report typical symptoms, and only 12% of COPD patients underwent pulmonary function tests (PFTs) before the survey, which fully demonstrated the hidden nature of early symptoms of COPD and the low public awareness of it. However, although PFT is the current "gold standard" for the diagnosis of COPD, it is not popularized in primary care settings. Common reasons include the lack of standardized PFT technology, the high cost and time consumption of PFT, and the limited ability to interpret the results. Early diagnosis of COPD is essential for the timely initiation of appropriate lifestyle and treatment interventions to reduce the prevalence, delay disease progression, and improve the quality of life of patients. Therefore, alternative strategies are urgently needed to provide rapid detection and accurate assessment of COPD to achieve optimal clinical decision-making.

[0003] It is well known that systematic active case finding in primary care settings by mailing screening questionnaires is an effective method. A widely used tool that is recognized by the Chinese clinical diagnosis and treatment guidelines is the COPD Screening Questionnaire (COPD-SQ). Other tools include the COPD Population Screening Questionnaire (COPD-PS) and the COPD Diagnostic Questionnaire (CDQ). They are simple and cost-effective methods for screening COPD based on symptoms and exposure factors, and are of great significance for identifying high-risk populations. However, the accuracy of questionnaires is difficult to improve significantly. Some even contain multiple subjective questions, which are prone to subject recall bias and time-consuming and labor-intensive to collect. Whether the extrapolation of different questionnaires is good needs further verification. In contrast, chest computed tomography (CT) is very popular, especially in large-scale lung cancer screening. At the same time, the Global Initiative for Chronic Obstructive Lung Disease (GOLD) in 2024 emphasizes the importance of CT in the assessment of patients with stable COPD and the use of lung cancer imaging for COPD screening, highlighting the role of imaging. This evidence-based recommendation is based on the limitations of PFT. CT has the advantages of being non-invasive and high-resolution, and can be used to evaluate changes in lung parenchyma, airways, and pulmonary vasculature that occur with declining lung function and aging. However, the traditional manual reading method has certain limitations in the evaluation of COPD using CT images. On the one hand, the interpretation of CT is somewhat subjective and will be affected by factors such as different judgment standards of radiologists and uneven reading levels. It requires long-term systematic training to correct and reduce evaluation bias. On the other hand, as a highly heterogeneous disease, COPD cannot be finely analyzed through single-dimensional images, and all-round, multi-dimensional visual observation is bound to increase the workload of medical staff. It may even cause problems on different scanning devices due to differences in detection parameters, reconstruction algorithms, image color differences, etc., resulting in different CT imaging effects. Therefore, with the help of rapidly developing artificial intelligence (AI) technology, it is crucial to find an objective, efficient, and accurate COPD intelligent evaluation method for the whole lung level of CT.

[0004] Radiomics is a relatively new method that can quickly obtain quantitative features from medical images (such as CT) and shows great potential in clinical decision-making. Early studies performed segmentation and quantification by manually marking regions of interest (generally lesions). However, manual segmentation of diffuse and heterogeneous lung lesions (such as emphysema, bronchiolitis, and interstitial lung disease) is very difficult and time-consuming due to unclear boundaries or low contrast on CT images, and the error rate may also increase depending on the operator's lack of experience. Therefore, automatic segmentation of the entire lung area will help to comprehensively quantify lung abnormalities in COPD patients and assist in clinical treatment decisions.

[0005] It is worth noting that the results of two recent studies indicate that whole lung CT imaging-based indicators are suitable for intelligent screening of COPD. However, one of the studies used a smoking-related Western cohort to train the model, which may not be applicable to other regions. Although the other used the Chinese population to develop a model, most of the cohorts were confirmed COPD patients, who had significant differences in clinical phenotype, biochemical indicators, and CT imaging compared with the general population. In addition, the above indicators may also be affected by imaging processing algorithms and effects, making them difficult to apply in routine practice. Therefore, it is necessary to provide a more appropriate CT evaluation scheme for COPD screening in my country. Summary of the invention

[0006] In order to overcome the deficiencies of the prior art, the object of the present invention is to provide a method for constructing a COPD screening model based on quantitative CT.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] A method for constructing a chronic obstructive pulmonary disease screening model based on quantitative CT, comprising:

[0009] The study population was screened according to the pre-set criteria to obtain the derivation cohort and the first external validation cohort;

[0010] Conducting questionnaire surveys, pulmonary function tests, and chest CT scans on the subjects of the derivative cohort to obtain a training data set and an internal validation data set, and collecting clinical data, performing pulmonary function tests, and performing chest CT scans on the subjects of the first external validation cohort to obtain an external validation data set in combination with a preset second external validation cohort;

[0011] Performing single-modal modeling and multi-modal modeling according to the type of each data in the training data set, and using the internal validation data set to validate and evaluate the model after modeling to obtain the optimal model;

[0012] The external validation dataset is used to evaluate the performance of the optimal model.

[0013] Preferably, the optimal model is constructed based on the XGBoost algorithm.

[0014] Preferably, screening is performed in the study population according to preset criteria to obtain a derivation cohort and a first external validation cohort, including:

[0015] The subjects from the derivation cohort who met the pre-set inclusion and exclusion criteria were randomly divided into a training cohort and an internal validation cohort at a ratio of 8:2;

[0016] The preset hospital cohorts were used as the first independent external validation cohorts.

[0017] Preferably, the subjects of the derivative cohort are subjected to questionnaires, pulmonary function tests, and chest CT scans to obtain a training data set and an internal validation data set, and the subjects of the first external validation cohort are subjected to clinical data collection, pulmonary function tests, and chest CT scans, and combined with the preset second external validation cohort to obtain an external validation data set, including:

[0018] A questionnaire is administered to the subjects of the derivation cohort to obtain questionnaire data; the indicators of the questionnaire include: demographic characteristics, anthropometric information, smoking history, personal and family medical history, respiratory symptoms, and life and dietary patterns;

[0019] The gender, age, smoking history, lung function test data and peripheral blood eosinophil count of all subjects in the first external validation cohort were collected through the electronic medical record system to obtain clinical data;

[0020] The data of the pre-set second external validation cohort were obtained through public databases; the data of the second external validation cohort included demographic characteristics, smoking history, disease history, and chest CT images;

[0021] Performing pulmonary function tests and chest CT scans on the subjects of the derivation cohort and the subjects of the first external validation cohort, respectively, to obtain pulmonary function test data and chest CT images;

[0022] Performing lung parenchyma analysis and airway analysis on the chest CT image using an image registration algorithm to obtain quantitative CT data;

[0023] Performing structured processing on the diagnosis result according to the CT report corresponding to the chest CT image to obtain structured sign data; the structured sign data includes: whether there are bullae, chronic inflammation, fibrous foci, calcified foci, nodules, bronchiectasis, tuberculosis, bronchitis and pleural thickening;

[0024] Determining the questionnaire data, the pulmonary function test data, the quantitative CT data, and the structured sign data corresponding to the subjects of the derivation cohort as the training data set and the internal validation data set;

[0025] The clinical data, the pulmonary function test data, the quantitative CT data, the structured sign data corresponding to the subjects of the first external validation cohort and the data of the second external validation cohort are determined as the external validation data set.

[0026] Preferably, the single-modal modeling includes: modeling using questionnaire data in the training data set, modeling using quantitative CT data in the training data set, and modeling using structured sign data in the training data set; the multi-modal modeling includes: modeling using questionnaire data and quantitative CT data in the training data set, modeling using questionnaire data and structured sign data in the training data set, modeling using quantitative CT data and structured sign data in the training data set, and modeling using questionnaire data, quantitative CT data, and structured signs in the training data set.

[0027] Preferably, the indicators for validating and evaluating the model after modeling using the internal validation data set include: AUC, sensitivity, specificity, accuracy, positive predictive value, negative predictive value and F1 score.

[0028] Preferably, the quantitative CT data include: LAA%-950, LAA%-950 left upper lung, LAA%-950 left lower lung, LAA%-950 right upper lung, LAA%-950 right middle lung, LAA%-950 right lower lung, LAA%-910 left lower lung, LAA%-910 right lower lung, maximum diameter of the first-level airway lumen and average diameter of the fourth-level airway lumen.

[0029] Preferably, the questionnaire data, the pulmonary function test data, the quantitative CT data and the structured sign data corresponding to the subjects of the derivation cohort are determined as the training data set and the internal validation data set, further comprising:

[0030] Using the createDataPartition function of the caret R package, randomly dividing the data set consisting of the questionnaire data, quantitative CT data, and structured sign data corresponding to the subjects of the derivation cohort into a training cohort and an internal validation cohort in a ratio of 80% and 20%;

[0031] Dividing the data of the training cohort and the internal validation cohort into measurement data and counting data, respectively, and further dividing the counting data into binary classification data and ordered multi-classification data;

[0032] Determining the measurement data as a first sequence;

[0033] The values ​​of each binary count data are recoded as 1 and 2 to obtain the second series;

[0034] Re-encoding the ordered multi-classification data by adopting a fusion weight encoding method to obtain a third sequence;

[0035] Use the missforest R package to interpolate missing values ​​of the first sequence, the second sequence, and the third sequence in the training cohort and the internal validation cohort, respectively, to obtain interpolated data sequences;

[0036] Using the nearZeroVar function of the caret R package to identify and remove the features of zero variance and near zero variance in the interpolated data sequence corresponding to the training cohort, to obtain a data sequence after removal;

[0037] The createFolds function is used to randomly split the eliminated data sequence corresponding to the training queue into multiple groups, and finally a preprocessed training data set and a preprocessed internal verification data set are obtained.

[0038] Preferably, the formula for the fusion weight encoding is: Where N is the maximum integer value after recoding, k is the total number of categories, and w i is the contribution weight of category i to the target variable Y, Among them, C i is an event of category i, P(c i ,y) is the probability of category i and target variable Y occurring simultaneously, P(c i ) and P(y) are C i and the marginal probability of Y, c i is the value of category i, and y is the specific value of the target variable Y.

[0039] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0040] The present invention provides a method for constructing a screening model for chronic obstructive pulmonary disease based on quantitative CT, comprising: screening according to preset standards in the research population to obtain a derivative cohort and a first external validation cohort; conducting questionnaires, pulmonary function tests and chest CT scan analysis on the subjects of the derivative cohort to obtain a training data set and an internal validation data set, and collecting clinical data, pulmonary function tests and chest CT scans on the subjects of the first external validation cohort, and combining with a preset second external validation cohort to obtain an external validation data set; performing single-modality modeling and multi-modality modeling according to the type of each data in the training data set, and using the internal validation data set to validate and evaluate the model after modeling to obtain an optimal model; and using the external validation data set to evaluate the performance of the optimal model. The present invention can significantly improve the level of early screening, diagnosis and management of chronic obstructive pulmonary disease, thereby improving the quality of life and prognosis of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0042] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0043] Figure 2 A schematic diagram of a research overview provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a method for constructing a chronic obstructive pulmonary disease screening model based on quantitative CT, comprising:

[0047] Step 1: Screen the study population according to the preset criteria to obtain the derivation cohort and the first external validation cohort;

[0048] Step 2: Conducting questionnaire surveys, pulmonary function tests, and chest CT scans on the subjects of the derivative cohort to obtain a training data set and an internal validation data set, and collecting clinical data, pulmonary function tests, and chest CT scans on the subjects of the first external validation cohort, and combining with the preset second external validation cohort to obtain an external validation data set;

[0049] Step 3: Performing single-modal modeling and multi-modal modeling according to the types of each data in the training data set, and using the internal validation data set to validate and evaluate the model after modeling to obtain the optimal model;

[0050] Step 4: Use the external validation dataset to evaluate the performance of the optimal model.

[0051] Preferably, the optimal model is constructed based on the XGBoost algorithm.

[0052] Preferably, screening is performed in the study population according to preset criteria to obtain a derivation cohort and an external validation cohort, including:

[0053] The subjects from the derivation cohort who met the pre-set inclusion and exclusion criteria were randomly divided into a training cohort and an internal validation cohort at a ratio of 8:2;

[0054] The preset hospital cohorts were used as the first independent external validation cohorts.

[0055] Specifically, all participants from China in this example were retrospectively recruited from four centers between April 2017 and May 2024, including four blocks in Guangzhou, the First Affiliated Hospital of Guangzhou Medical University, Xiangyang Central Hospital, and the Second Affiliated Hospital of Xi'an Jiaotong University. The inclusion criteria were as follows: 1) age 35-80 years; 2) at least one PFT was completed; 3) inspiratory chest CT scan in the supine position; 4) chest CT scan and PFT examination were performed in the same center; 5) participants from the community were surveyed by respiratory epidemiology questionnaire. The exclusion criteria were as follows: 1) incomplete CT images; 2) a large number of image artifacts; 3) a history of lung resection; 4) the questionnaire information missing rate was >50%. Finally, a total of 3653 participants were included. Among them, 1950 participants from the community (650 COPD patients and 1300 controls, fully frequency matched by sex and age) were assigned to the derivation cohort and randomly divided into the training cohort (n=1560) and the internal validation cohort (n=390) in an 8:2 ratio. 1186 patients from the First Affiliated Hospital of Guangzhou Medical University, 225 patients from Xiangyang Central Hospital, and 292 patients from the Second Affiliated Hospital of Xi'an Jiaotong University were assigned to external validation cohorts 1, 2, and 3, respectively. A random subset of the US NLST cohort (n = 453) was further used as an external validation cohort 4 to verify the multiethnic extrapolation and robustness of low-dose CT scanning. This study was approved by all institutional review boards (approval numbers: ES–2024–K060–01; NLST–1167), and written informed consent was not required due to the retrospective nature of the study.

[0056] The derivation cohort used forced expiratory volume in one second (FEV1) / forced vital capacity (FVC) < 0.7 before inhalation of bronchodilators as the diagnostic criterion for COPD, and the external validation cohort (except the NLST cohort) used FEV1 / FVC < 0.7 after inhalation of bronchodilators as the diagnostic criterion for COPD. Further classification was made with reference to the GOLD guidelines: GOLD grade 1, FEV1% pred ≥ 80%; GOLD grade 2: 50% ≤ FEV1% pred < 80%; GOLD grade 3: 30% ≤ FEV1% pred < 50%; GOLD grade 4: FEV1% pred < 30%.

[0057] Furthermore, this embodiment performs sample size calculation, which is as follows:

[0058] This study adopted a double-gate design diagnostic test, the outcome index was a qualitative index, and the sample size of the case group and the control group was equal. The main parameters were as follows: sensitivity = 80%, the tolerance of sensitivity = 5%; specificity = 70%, the tolerance of specificity = 5%; the confidence level 1-α = 0.95 was selected, and the prevalence (prior probability) was 0.33 in this study (COPD group: control group = 1:2). The invalid sensitivity and specificity were both 0.5.

[0059] According to the calculation of PASS software, a total of 800 samples are needed, of which 264 patients with COPD diagnosis are clearly excluded by the gold standard, in order to reach the preset 95% confidence interval range of the test. Considering the 20% loss rate, 1000 samples are ultimately required. A total of 1950 subjects were included in the development of this experimental model, with 650 and 1300 samples in the COPD group and control group, respectively, meeting the minimum sample size requirements for model training.

[0060] Preferably, the subjects of the derivation cohort are subjected to questionnaires, pulmonary function tests and chest CT scans to obtain a training data set, and the subjects of the external validation cohort are subjected to clinical data collection, pulmonary function tests and chest CT scans to obtain an external validation data set, including:

[0061] A questionnaire is administered to the subjects of the derivation cohort to obtain questionnaire data; the indicators of the questionnaire include: demographic characteristics, anthropometric information, smoking history, personal and family medical history, respiratory symptoms, and life and dietary patterns;

[0062] The gender, age, smoking history, lung function test data and peripheral blood eosinophil count of all subjects in the first external validation cohort were collected through the electronic medical record system to obtain clinical data;

[0063] The data of the pre-set second external validation cohort were obtained through public databases; the data of the second external validation cohort included demographic characteristics, smoking history, disease history, and chest CT images;

[0064] The subjects of the derivative cohort and the subjects of the first external validation cohort are respectively subjected to pulmonary function tests and chest CT scans to obtain pulmonary function test data and chest CT images; the pulmonary function test data are used to assign disease labels to the subjects and are not included in the modeling process.

[0065] Performing lung parenchyma analysis and airway analysis on the chest CT image using an image registration algorithm to obtain quantitative CT data;

[0066] Performing structured processing on the diagnosis result according to the CT report corresponding to the chest CT image to obtain structured sign data; the structured sign data includes: whether there are bullae, chronic inflammation, fibrous foci, calcified foci, nodules, bronchiectasis, tuberculosis, bronchitis and pleural thickening;

[0067] Determining the questionnaire data, the pulmonary function test data, the quantitative CT data, and the structured sign data corresponding to the subjects of the derivation cohort as the training data set and the internal validation data set;

[0068] The clinical data, the pulmonary function test data, the quantitative CT data, the structured sign data corresponding to the subjects of the first external validation cohort and the data of the second external validation cohort are determined as the external validation data set.

[0069] Specifically, the steps of data collection and processing in this embodiment are as follows:

[0070] 1) Questionnaire and clinical data collection

[0071] All subjects in the derivative cohort filled out the questionnaire offline under the guidance of staff with proficient epidemiological survey skills. The questionnaire contained 40 indicators, mainly including demographic characteristics, anthropometric information, smoking history, personal and family medical history, respiratory symptoms, and lifestyle and dietary patterns. For the external validation cohort, the gender, age, smoking history, lung function test data, and peripheral blood eosinophil count of all subjects from China were collected through the electronic medical record system, and the US NLST external validation cohort data, including demographic characteristics, smoking history, medical history, chest CT images, etc., were provided through the US NIH official website.

[0072] 2) Pulmonary function test

[0073] All subjects underwent pulmonary function tests using a domestically produced new differential pressure spirometer, the U-Pneumo PF680 (Yiliankang, Zhejiang, China) or the MasterScreen Pneumo spirometer (Yeger, Bavaria, Germany). Subjects from the external validation cohort in China underwent bronchial dilation tests. All pulmonary function tests were performed by professional technicians. The FEV1 / FVC and FEV1%pred indicators were mainly included in this study.

[0074] 3) Chest CT Scan Protocol

[0075] All cohorts used CT detection instruments from seven equipment manufacturers: SIEMENS, GE, NMS, TOSHIBA, Philips, UIH, and Anke. The subjects were in a supine position with their arms raised above their heads for a chest CT scan at the end of maximum inspiration. The scanning area extended from the thoracic inlet to the level of the bilateral adrenal glands. All images were exported in the Digital Imaging and Communications in Medicine (DICOM) format for quantitative data extraction.

[0076] 4) Image analysis

[0077] CT images were analyzed using NeuLungCare–QA software (version 1.0, Neusoft Software Co., Ltd., Shenyang), which was commercially certified and validated in a previous study (13) for quantification of CT emphysema and airway lesions. A total of 47 QCT features were extracted in this study. Briefly, CT emphysema was quantified for the whole lung and each lobe using low attenuation areas below −950 (LAA-950) and −910 (LAA-910) Hounsfield units on full inspiratory images, respectively.

[0078] CT airway measurements were generated from bronchi from generation 0 to 4. Percent airway wall area (%WA) was measured by dividing the total bronchial cross-sectional area by the wall area. For airway size measurements, airway length was defined as the distance from the start to the end of the branch point by placing a smooth center line through the lumen. Wall thickness (WT) and lumen diameter (LD) were calculated for each airway every 2 mm. Maximum, minimum, and mean values ​​of WT and LD were generated from five measurements.

[0079] All original chest CT images in DICOM format were imported into the NeuLungCARE-QA system (version 1.0, Neusoft Software Co., Ltd., Shenyang) for quantification of CT emphysema and airway lesions. The analysis process is briefly described as follows: the system processes the read-in CT images using the optimized Demons registration algorithm and performs quantitative measurements according to the classification standards of previous literature:

[0080] (1) Lung parenchyma analysis: Each lung lobe was automatically segmented along the interlobar fissures, and the thresholds were set at -950 Hounsfield units (HU) and -910 HU, respectively. Lung attenuation voxels less than -950 HU or -910 HU were considered low attenuation areas, expressed as LAA%-950 and LAA%-910, and the LAA% of the whole lung, left upper lobe, left lower lobe, right upper lobe, right middle lobe, and right lower lobe were calculated, respectively.

[0081] (2) Airway analysis: The bronchial tree was reconstructed in three dimensions, and then the 0-4th level airways were automatically selected as target airways. The percentage of airway wall area (%WA) was obtained by dividing the total bronchial cross-sectional area by the wall area. For airway size measurement, airway length was defined as the distance from the start point to the end point of the branch point by placing a smooth center line in the lumen. The wall thickness (WT) and lumen diameter (LD) were calculated for each airway every 2 mm. The maximum, minimum, and average values ​​of WT and LD were generated from five measurement results.

[0082] 5) CT report text processing

[0083] This embodiment simultaneously collects CT reports interpreted by professional radiologists, structures the diagnostic results, and finally records the nine signs of bullae (including emphysema), chronic inflammation (including acute and chronic infection), fibrous foci, calcification foci, nodules (including granulomas, neoplasms and suspected tumors), bronchiectasis, tuberculosis, bronchitis (including bronchiolitis and chronic inflammation of small airways) and pleural thickening.

[0084] Preferably, before performing single-modal modeling and multi-modal modeling, the method further includes:

[0085] Using the createDataPartition function of the caret R package, randomly dividing the data set consisting of the questionnaire data, quantitative CT data, and structured sign data corresponding to the subjects of the derivation cohort into a training cohort and an internal validation cohort in a ratio of 80% and 20%;

[0086] Dividing the data of the training cohort and the internal validation cohort into measurement data and counting data, respectively, and further dividing the counting data into binary classification data and ordered multi-classification data;

[0087] Determining the measurement data as a first sequence;

[0088] The values ​​of each binary count data are recoded as 1 and 2 to obtain the second series;

[0089] Re-encoding the ordered multi-classification data by adopting a fusion weight encoding method to obtain a third sequence;

[0090] Use the missforest R package to interpolate missing values ​​of the first sequence, the second sequence, and the third sequence in the training cohort and the internal validation cohort, respectively, to obtain interpolated data sequences;

[0091] Using the nearZeroVar function of the caret R package to identify and remove the features of zero variance and near zero variance in the interpolated data sequence corresponding to the training cohort, to obtain a data sequence after removal;

[0092] The createFolds function is used to randomly split the eliminated data sequence corresponding to the training queue into multiple groups, and finally a preprocessed training data set and a preprocessed internal verification data set are obtained.

[0093] Preferably, the formula for the fusion weight encoding is: Where N is the maximum integer value after recoding, k is the total number of categories, and w i is the contribution weight of category i to the target variable Y, Among them, C i is an event of category i, P(c i ,y) is the probability of category i and target variable Y occurring simultaneously, P(c i ) and P(y) are C i and the marginal probability of Y, c i is the value of category i, and y is the specific value of the target variable Y.

[0094] Furthermore, in this embodiment, for binary count data, in addition to simply encoding as 1 and 2, a robust encoding method for noise and outliers can be introduced. For example, outliers can be detected and redistributed by calculating quantiles to ensure more robust encoding.

[0095] For ordered multi-class data, fusion weighted encoding is introduced. Based on the contribution of each class in the prediction of the target variable (which can be determined by the preliminary feature importance analysis), the class values ​​are weighted and encoded into continuous positive integers.

[0096] In addition, this embodiment can also adopt an optimized data segmentation strategy, that is, using Adaptive Stratified Partitioning, which not only relies on a simple 80 / 20 split, but also achieves a more balanced split in the distribution of target variables and key features. The consistency of data distribution on the training and validation sets is ensured by using indicators such as weighted Kappa.

[0097] Furthermore, this embodiment also introduces Neighborhood Dynamic Regression Imputation (NDRI), which is an enhanced version of the classic missForest method. This method combines local linear regression with the K nearest neighbor algorithm to achieve dynamic prediction of missing values.

[0098] In addition, this embodiment also uses an Augmented Generative Adversarial Network (AGAN) for data augmentation to generate diverse samples in the feature space, thereby improving the generalization ability of the model. Multi-level cross-validation is further implemented. In addition to the classic ten-fold, multi-level feature and model validation is performed by introducing multi-scale partitioning and subsampling strategies to reduce the risk of overfitting.

[0099] Preferably, the single-modality modeling includes: modeling using questionnaire data in the training data set, modeling using quantitative CT data in the training data set, and modeling using CT reports in the training data set; the multi-modality modeling includes: modeling using questionnaire data and quantitative CT data in the training data set, modeling using questionnaire data and structured sign data in the training data set, modeling using quantitative CT data and structured sign data in the training data set, and modeling using questionnaire data, quantitative CT data, and structured signs in the training data set.

[0100] Preferably, the model after modeling is verified and evaluated to obtain the optimal model, including:

[0101] The internal validation cohort is used to validate and evaluate the model after modeling to obtain the optimal model; the evaluation indicators include: AUC, sensitivity, specificity, accuracy, positive predictive value, negative predictive value and F1 score.

[0102] Furthermore, this embodiment also includes:

[0103] 7) Model construction and performance evaluation

[0104] The XGBoost algorithm was used as the core of the screening model. The specific construction method is as follows: On the preprocessed training cohort, based on the ten-fold cross-validation framework, the model hyperparameters were automatically optimized using the BayesianOptimization function of the rBayesianOptimization R package, including eta, max_depth, subsample, colsample_byTree, gamma, min_child_weight, and nrounds. The parameters init_points and n_iter were set to 5 and 1, respectively, and the remaining parameters were set to the default values. AUC was used as the cross-validation loss function value for evaluation. After determining the optimal hyperparameters, SHAP analysis was used to screen out the top ten features, and the final screening model was rebuilt in the same way.

[0105] To identify the optimal screening program, the above process was used to model the single-modality (questionnaire, quantitative CT, and CT report) and multimodality (questionnaire + quantitative CT, questionnaire + CT report, quantitative CT + CT report, and questionnaire + quantitative CT + CT report) programs. Finally, they were uniformly evaluated on the internal validation cohort, and the optimal model was named AutoCOPD. AUC was used as the overall discriminant index for the classification effect of the model. In addition, sensitivity, specificity, accuracy, positive predictive value, negative predictive value, and F1 score were used as additional evaluation indicators. Finally, the area under the receiver operating characteristic curve (AUC) of different models was compared using the Delong test, and P < 0.05 indicated that the difference between AUCs was statistically significant. The above indicators were calculated using the pROC R package. Then, the performance of AutoCOPD was further evaluated on all external validation cohorts, and the stats R package was used to draw lowess curves to reflect the calibration of the model. The ResourceSelection R package was used to perform the Hosmer-Lemeshow test, and P > 0.05 indicated that the model calibration was good. The Brier score was then calculated to comprehensively evaluate the calibration. The rmdaR package was used to draw clinical decision curves (DCA) of the model applied to all cohorts. Subgroup analysis was performed in the internal and external validation cohorts to clarify the applicability of the model. The COPD identification performance of the model was compared with that of the classic screening tool COPD-SQ in the internal validation cohort.

[0106] 8) Core Results

[0107] Finally, this embodiment determined that the XGBoost model constructed based on 10 quantitative CT features was the best screening model (i.e., AutoCOPD), and the features included LAA%-950, LAA%-950 left upper lung, LAA%-950 left lower lung, LAA%-950 right upper lung, LAA%-950 right middle lung, LAA%-950 right lower lung, LAA%-910 left lower lung, LAA%-910 right lower lung, maximum diameter of the first-level airway lumen, and average diameter of the fourth-level airway lumen. The model can be well verified in different scenarios, different regions, and ethnic populations, and the model calibration is good. The decision curve analysis shows that the overall clinical net benefit ranges from 0.12 to 0.66, and the overall recognition ability is also much better than COPD-SQ. In summary, AutoCOPD can be used as a new COPD screening tool.

[0108] 8) System Development

[0109] This embodiment also uses R Shiny to make a user-friendly free online App (https: / / lwj-lab.shinyapps.io / autocopd / ) based on the AutoCOPD model. This application includes routine, custom, and individual analysis. In routine analysis, users can build their own data sets according to the format of the sample file and upload them to the App for batch prediction. When the prediction score ≥50% indicates a COPD patient, <50% indicates non-COPD. Custom analysis is designed to perform external validation on self-built data sets, provide the best decision threshold based on the Youden index, and display standard evaluation indicators, ROC curves, calibration curves, and decision curves. All missing values ​​are automatically interpolated before analysis, and users can download interpolated data and prediction results for local research. In individual prediction, when the actual values ​​of the 10 features required by the model (at least 1 feature value is entered) are entered and the "Calculate" button is clicked, this application will automatically predict the individual's risk of COPD. In addition, a SHAP force diagram will be displayed to indicate feature contributions: purple features push the prediction toward "non-COPD", while yellow features push the prediction into "COPD". Overall, the App is easy to operate, responds quickly, and has focused functions, making it highly consistent with clinical practice.

[0110] The advantages of the quantitative CT-based chronic obstructive pulmonary disease (COPD) screening model of the present invention are mainly reflected in the following aspects:

[0111] (1) Early screening and diagnosis: Through a machine learning model that only includes 10 whole-lung inspiratory phase quantitative CT indicators, COPD can be detected early, thereby increasing the possibility of early intervention and treatment and slowing down disease progression.

[0112] (2) Multimodal comparison, integrating different data sources (such as questionnaires, quantitative CT, and structured signs) to make a more comprehensive comparison and derive the optimal screening program suitable for my country.

[0113] (3) Quantitative analysis: Commercialized quantitative CT analysis software certified by the China Food and Drug Administration can accurately measure changes in lung structure, such as emphysema, airway structure and volume changes, provide objective data support, and improve the credibility of screening.

[0114] (4) Standardized procedures: preset standard patient screening and data collection procedures ensure the homogeneity of the study population and the comparability of the data, thus improving the reliability and validity of the results.

[0115] (5) External validation: Through the evaluation of the external validation cohort, the generalization ability of the model can be effectively tested to ensure the effectiveness of the model in different populations and environmental conditions.

[0116] (6) Potential for personalized medicine: This model has strong clinical interpretability, which is conducive to combining other individual characteristics and clinical data during clinical diagnosis and treatment to tailor personalized management plans for subjects.

[0117] (7) Promote clinical decision-making. The construction and evaluation of the model will provide a scientific basis for clinicians, helping them make more accurate diagnosis and treatment decisions, thereby improving treatment outcomes.

[0118] (8) Resource optimization: accurate screening models can help clinical institutions optimize resource allocation, improve screening efficiency for high-risk populations, and reduce unnecessary examinations and medical expenses.

[0119] (9) Promote research progress. The construction of the model of the present invention provides a foundation for the study of COPD, helps to understand the disease progression mechanism and risk factors, and provides data support and theoretical basis for future related research.

[0120] Through the above advantages, the model constructed by the present invention can significantly improve the early screening, diagnosis and management of chronic obstructive pulmonary disease, thereby improving the quality of life and prognosis of patients.

[0121] The beneficial effects of the present invention are as follows:

[0122] (1) The research involved in this article is the largest multicenter quantitative CT-COPD screening model study in China and abroad;

[0123] (2) Using 10 highly interpretable whole-lung inspiratory phase chest quantitative CT features, we successfully constructed an intelligent COPD screening model (AutoCOPD) and online app based on the XGBoost algorithm. The screening performance in the training cohort, internal validation cohort, and external validation cohort was good, and the recognition performance in the highly heterogeneous community population was better than that of the classic screening tool COPD-SQ.

[0124] (3) The model of the present invention has good COPD identification ability before and after receiving bronchodilators and in different regions and ethnic groups, laying a new theoretical foundation for large-scale COPD imaging screening.

[0125] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0126] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for constructing a screening model for chronic obstructive pulmonary disease based on quantitative CT, characterized in that: include: The study population was screened according to the pre-set criteria to obtain the derivation cohort and the first external validation cohort; Conducting questionnaire surveys, pulmonary function tests, and chest CT scans on the subjects of the derivative cohort to obtain a training data set and an internal validation data set, and collecting clinical data, performing pulmonary function tests, and performing chest CT scans on the subjects of the first external validation cohort to obtain an external validation data set in combination with a preset second external validation cohort; Performing single-modal modeling and multi-modal modeling according to the type of each data in the training data set, and using the internal validation data set to validate and evaluate the model after modeling to obtain the optimal model; The external validation dataset is used to evaluate the performance of the optimal model.

2. The method for constructing a COPD screening model based on quantitative CT according to claim 1, characterized in that: The optimal model is constructed based on the XGBoost algorithm.

3. The method for constructing a COPD screening model based on quantitative CT according to claim 1, characterized in that: The study population was screened according to the pre-set criteria to obtain the derivation cohort and the first external validation cohort, including: The subjects from the derivation cohort who met the pre-set inclusion and exclusion criteria were randomly divided into a training cohort and an internal validation cohort at a ratio of 8:2; The preset hospital cohorts were used as the first independent external validation cohorts.

4. The method for constructing a COPD screening model based on quantitative CT according to claim 1, characterized in that: The subjects of the derivative cohort are subjected to questionnaires, pulmonary function tests, and chest CT scans to obtain a training data set and an internal validation data set, and the subjects of the first external validation cohort are subjected to clinical data collection, pulmonary function tests, and chest CT scans, and the external validation data set is obtained in combination with the preset second external validation cohort, including: A questionnaire is administered to the subjects of the derivation cohort to obtain questionnaire data; the indicators of the questionnaire include: demographic characteristics, anthropometric information, smoking history, personal and family medical history, respiratory symptoms, and life and dietary patterns; The gender, age, smoking history, lung function test data and peripheral blood eosinophil count of all subjects in the first external validation cohort were collected through the electronic medical record system to obtain clinical data; The data of the pre-set second external validation cohort were obtained through public databases; the data of the second external validation cohort included demographic characteristics, smoking history, disease history, and chest CT images; Performing pulmonary function tests and chest CT scans on the subjects of the derivation cohort and the subjects of the first external validation cohort, respectively, to obtain pulmonary function test data and chest CT images; Performing lung parenchyma analysis and airway analysis on the chest CT image using an image registration algorithm to obtain quantitative CT data; Performing structured processing on the diagnosis result according to the CT report corresponding to the chest CT image to obtain structured sign data; the structured sign data includes: whether there are bullae, chronic inflammation, fibrous foci, calcified foci, nodules, bronchiectasis, tuberculosis, bronchitis and pleural thickening; Determining the questionnaire data, the pulmonary function test data, the quantitative CT data, and the structured sign data corresponding to the subjects of the derivation cohort as the training data set and the internal validation data set; The clinical data, the pulmonary function test data, the quantitative CT data, the structured sign data corresponding to the subjects of the first external validation cohort and the data of the second external validation cohort are determined as the external validation data set.

5. The method for constructing a COPD screening model based on quantitative CT according to claim 4, characterized in that: The single-modal modeling includes: modeling using the questionnaire data in the training data set, modeling using the quantitative CT data in the training data set, and modeling using the structured sign data in the training data set; the multi-modal modeling includes: modeling using the questionnaire data and quantitative CT data in the training data set, modeling using the questionnaire data and structured sign data in the training data set, modeling using the quantitative CT data and structured sign data in the training data set, and modeling using the questionnaire data, quantitative CT data, and structured signs in the training data set.

6. The method for constructing a COPD screening model based on quantitative CT according to claim 3, characterized in that: The indicators used to verify and evaluate the model after modeling using the internal validation data set include: AUC, sensitivity, specificity, accuracy, positive predictive value, negative predictive value and F1 score.

7. The method for constructing a COPD screening model based on quantitative CT according to claim 4, characterized in that: The quantitative CT data include: LAA%-950, LAA%-950 left upper lung, LAA%-950 left lower lung, LAA%-950 right upper lung, LAA%-950 right middle lung, LAA%-950 right lower lung, LAA%-910 left lower lung, LAA%-910 right lower lung, maximum diameter of the first-level airway lumen and average diameter of the fourth-level airway lumen.

8. The method for constructing a COPD screening model based on quantitative CT according to claim 4, characterized in that: Determining the questionnaire data, the pulmonary function test data, the quantitative CT data, and the structured sign data corresponding to the subjects of the derived cohort as the training data set and the internal validation data set, further comprising: Using the createDataPartition function of the caret R package, randomly dividing the data set consisting of the questionnaire data, quantitative CT data, and structured sign data corresponding to the subjects of the derivation cohort into a training cohort and an internal validation cohort in a ratio of 80% and 20%; Dividing the data of the training cohort and the internal validation cohort into measurement data and counting data, respectively, and further dividing the counting data into binary data and ordered multi-classification data; Determining the measurement data as a first sequence; The values ​​of each binary count data are recoded as 1 and 2 to obtain the second series; Re-encoding the ordered multi-classification data by adopting a fusion weight encoding method to obtain a third sequence; Use the missforest R package to interpolate missing values ​​of the first sequence, the second sequence, and the third sequence in the training cohort and the internal validation cohort, respectively, to obtain interpolated data sequences; Using the nearZeroVar function of the caret R package, identifying and removing the features of zero variance and near zero variance in the interpolated data sequence corresponding to the training cohort, to obtain a data sequence after removal; The createFolds function is used to randomly split the eliminated data sequence corresponding to the training queue into multiple groups, and finally a preprocessed training data set and a preprocessed internal verification data set are obtained.

9. The method for constructing a COPD screening model based on quantitative CT according to claim 8, characterized in that: The formula for the fusion weight encoding is: Where N is the maximum integer value after recoding, k is the total number of categories, and w i is the contribution weight of category i to the target variable Y, Among them, C i is an event of category i, P(c i ,y) is the probability of category i and target variable Y occurring simultaneously, P(c i ) and P(y) are C i and the marginal probability of Y, c i is the value of category i, and y is the specific value of the target variable Y.

Images