Skip to main content
NCBI home page
Annals of Medicine logoLink to Annals of Medicine
. 2026 Oct 1;58(1):2735163. doi: 10.1080/07853890.2026.2735163

Identification of clinical phenotypes and development of a predictive model for rapid lung function decline in COPD patients

Zehua Yang a,b,c,d,e,f,g,h, Wei Li b,c,d,e,f,g,h, Yanan Cui b,c,d,e,f,g,h, Tingting Huang b,c,d,e,f,g,h, Xingyao Tang b,c,d,e,f,g,h, Yaodie Peng b,c,d,e,f,g,h, Rui Su b,c,d,e,f,g,h, Xu Chu b,c,d,e,f,g,h, Yong Li b,c,d,e,f,g,h, Chunyu Zhang b,c,d,e,f,g,h, Cunbo Jia i, Ke Huang b,c,d,e,f,g,h,✉, Ting Yang a,b,c,d,e,f,g,h,✉
PMCID: PMC13637773  PMID: 42817919

Abstract

Background

Chronic obstructive pulmonary disease (COPD) exhibits significant clinical heterogeneity. This study integrates phenotype identification with longitudinal lung function analysis to develop a prediction tool for rapid decline in a large Chinese cohort.

Methods

This study included 38,863 COPD patients from a national multicenter screening program. Latent class analysis identified clinical phenotypes. Among 1,215 patients with baseline and 24-month follow-up spirometry, rapid decline was defined as annual FEV1 decline ≥40 mL or FEV1% predicted decline >1.5%. LASSO regression and the Boruta algorithm selected predictive variables. Six machine learning models were constructed and evaluated for discrimination, calibration, and utility. SHAP analysis was employed for interpretation.

Results

Three distinct phenotypes were identified: Phenotype 1 (GOLD Stage 2, non-smoking females, 25.8%), Phenotype 2 (GOLD Stage 1, smoking males, 44.8%), and Phenotype 3 (GOLD Stage 3, severe smokers, 29.4%). Phenotype 2 exhibited the highest rapid decline rate (55.8%). The CatBoost model achieved optimal performance (AUC 0.712) using six variables: region, fuel type, income level, wheezing, hip circumference, and baseline FEV1. SHAP identified baseline FEV1 as the top predictor. At a threshold of 0.365, the model attained 95.5% sensitivity and 86.4% negative predictive value, with significant net benefit.

Conclusion

Three clinical phenotypes were identified among COPD patients. Phenotype 2 exhibits the highest risk of rapid progression, representing a key target group for intervention. The predictive model, based on six simple indicators, serves as a valuable tool for screening high-risk rapid decliners.

Keywords: Chronic obstructive pulmonary disease, clinical phenotypes, lung function decline, predictive models, machine learning

Introduction

Chronic obstructive pulmonary disease (COPD) is a highly heterogeneous respiratory condition characterized by progressive decline in lung function [1]. COPD affects over 400 million individuals and ranks as the third leading cause of death worldwide, with low- and middle-income countries contributing significantly to its growing burden [2]. The China Pulmonary Health Study (CPHS) revealed that COPD prevalence among individuals aged over 40 years in China reached 13.7%, a notable increase from 8.2% reported between 2002 and 2004 [3,4]. Traditional management strategies primarily rely on the Global Initiative for Chronic Obstructive Lung Disease (GOLD) classification, which uses forced expiratory volume in one second (FEV1) as the core indicator for assessing airflow obstruction severity [5]. In clinical practice, explaining substantial differences among patients regarding symptom presentation, acute exacerbation frequency, and treatment responses poses a significant challenge [6].

Precision medicine has been introduced in COPD, focusing on identifying subgroups with distinct disease mechanisms, clinical courses, and treatment needs [7]. The identification of clinical phenotypes based on observable characteristics—differentiating emphysema-dominant, chronic bronchitis-dominant, and frequent exacerbator phenotypes—has become a cornerstone for personalized management [8]. Notably, these classical phenotypes cannot be differentiated purely on the basis of spirometry or patient-reported outcomes; their definition additionally requires imaging and exacerbation data. Unsupervised machine learning methods have been extensively applied in respiratory disease phenotyping, providing powerful tools to uncover intrinsic disease heterogeneity [9]. Longitudinal evidence suggests that lung function decline among COPD patients is neither linear nor homogeneous, exhibiting significant trajectory heterogeneity [10]. Rapid decliners are of particular concern, as their FEV1 deteriorates sharply over brief periods, closely associated with increased hospitalization rates, mortality risk, and healthcare burden [11]. Early identification of such high-risk trajectories is crucial for modifying disease course and implementing targeted interventions. Identified predictors of rapid decline include clinical and behavioral factors such as current smoking, frequent exacerbations, and higher comorbidity burden [12]. Greater initial disease severity, imaging evidence of emphysema, and systemic inflammation have likewise been associated with accelerated FEV1 decline [12–14].

Current evidence regarding COPD phenotypes and lung function trajectories primarily derives from Western cohorts such as ECLIPSE, SPIROMICS, and COPDGene [15]. In China, never-smokers constitute 41% of COPD patients, rising to 81.3% among females [3]. Biomass fuel exposure, early respiratory infections, and tuberculosis history represent significant non-smoking risk factors [16–18]. Directly applying phenotype classifications from international cohorts may not adequately represent Chinese patient characteristics. Large cohort-based research from the Chinese population remains limited.

We conducted a study involving 38,863 COPD patients in China, incorporating 36 variables and utilizing unsupervised machine learning to identify clinical phenotypes. We identified rapid lung function decliners and assessed their association with baseline clinical phenotypes. Machine learning algorithms were employed to develop a straightforward predictive model for screening individuals at high risk of rapid lung function decline. This study aims to establish a precise COPD risk stratification system, facilitating a transition from passive universal management to proactive, individualized care.

Methods

Study design and population

This analysis utilized data from “The National COPD Screening Programme in China (2022–2025)” [19]. Community-based screening for individuals at high risk of COPD was conducted using standardized questionnaires and spirometry, with subsequent risk-stratified comprehensive interventions implemented based on screening findings. The project encompasses 160 sites across 31 provinces. In accordance with the Declaration of Helsinki, the study protocol received approval from the Ethics Committee of the National Respiratory Medicine Center at the Japan-China Friendship Hospital (approval number: 2021-145-K103), and all participants provided written informed consent.

The selection process is illustrated in Figure S1. The initial screening comprised 2,193,579 subjects who completed the COPD screening questionnaire. Among these, 45,965 individuals were diagnosed with COPD based on pulmonary function tests, defined as post-bronchodilator FEV1/FVC < 0.7. After excluding individuals with FEV1 reproducibility exceeding 0.2 L between acceptable maneuvers, 38,863 patients were included in the cross-sectional analysis [20]. Patients who completed both baseline and 24-month follow-up pulmonary function tests, with FEV1 reproducibility within 0.2 L at both assessments, formed a final cohort of 1,215 patients for longitudinal analyses.

Feature selection and data preprocessing

The study included seven categories of baseline characteristics: demographic characteristics, socioeconomic status, lifestyle, respiratory symptoms, vaccination history, physical indicators, and pulmonary function indicators. Outliers were addressed for all continuous variables, and variables with missing rates exceeding 20% were excluded from the analysis. Missing data were imputed using multiple imputation methods.

Exploration of clinical phenotypes

Clinical phenotypes were explored through latent class analysis. The “poLCA” package in R version 4.5.2 was utilized to fit models containing between 2 and 6 latent classes. Model selection was guided by a combination of statistical indicators and clinical interpretability, including the Bayesian Information Criterion (BIC), Akaike Information Criterion (AIC), entropy, average posterior probabilities, and the proportion of samples within each class.

This study employs effect sizes to rank the importance of variables and identify key distinguishing features of each phenotype. For continuous variables, η2 effect sizes are calculated using analysis of variance, while Cramer’s V effect sizes are computed for categorical variables using chi-square tests. Variables that exhibit p < 0.05 and are ranked in the top 15 for effect size—where pulmonary function indicators are limited to FEV1% predicted—are selected as core variables. Based on the distribution characteristics of each category in relation to these core variables, and in conjunction with the GOLD pulmonary function grading criteria and clinical knowledge, the phenotypic characteristics are thoroughly described and appropriately named.

Analysis of lung function trajectories

The rapid decline subgroup was identified based on baseline and 24-month follow-up values of FEV1 and FEV1% predicted. Rapid decline was operationally defined as an annual decline in FEV1 of ≥40 mL or an annual rate of decline in FEV1% predicted exceeding 1.5% [21,22]. Conversely, the remaining patients were categorized as non-rapidly declining, encompassing both stable and improving subtypes. To compare differences in the proportion of rapidly declining patients across baseline clinical phenotypes, the chi-square test was utilized, followed by post-hoc pairwise comparisons with Bonferroni correction for multiple testing.

Modeling and evaluation

A two-step feature selection strategy was employed. Initially, multicollinearity was addressed by eliminating variables with a variance inflation factor (VIF) greater than 5 or a Pearson correlation coefficient exceeding 0.7. Subsequently, the LASSO and Boruta algorithms were utilized to identify key predictors. The stability of the features was assessed using 500 bootstrap resamples, and features selected in over 80% of the iterations were incorporated into the final model development.

We compared six machine learning algorithms to construct predictive models: Naive Bayes, Logistic Regression (Elastic Net), Support Vector Machines, Random Forests, XGBoost, and CatBoost. Utilizing stratified sampling, we randomly divided the dataset into training and test sets in a 7:3 ratio, ensuring that the distribution of outcomes remained consistent. The 7:3 split is commonly adopted in prediction model development, balancing adequate training sample size against a sufficiently large independent test set (365 participants, 164 events) for unbiased evaluation. We implemented 10-fold cross-validation for internal validation, using the area under the curve (AUC) as the primary optimization metric, and adopted a staged grid search strategy for hyperparameter tuning.

Model discriminatory power was evaluated using AUC, accuracy, precision, sensitivity, specificity, and F1 score on an independent test set. Additionally, the model’s calibration was assessed through calibration curves and its clinical utility via decision curve analysis. The SHAP algorithm was utilized to interpret feature importance and evaluate the contribution of individual variables. To further enhance clinical utility, logistic regression with restricted cubic splines was implemented to model the relationship between the optimal model’s predictions and the original features, thereby generating visualized contour plots.

Statistical analysis

Continuous variables that follow a normal distribution are presented as mean ± standard deviation and are compared using Student’s t-test. Variables that do not follow a normal distribution are reported as median (interquartile range) and are compared using the Wilcoxon signed-rank test. Categorical variables are expressed as proportions and are compared using either the chi-square test or Fisher’s exact test. All statistical analyses were conducted using R version 4.5.2, with a two-sided p < 0.05 considered statistically significant.

Results

Patient characteristics

The initial phase of the study encompassed a total of 38,863 eligible patients diagnosed with COPD. As illustrated in Table 1, the median age of the study population was 66 years (60,70); 81.9% of the participants were male, and 93.6% were of Han ethnicity. The overall median FEV1% of the predicted value was 69.00% (51.80,84.70), and the smoking index averaged 28 pack-years (0,47). Notably, only 25.8% of the patients reported having never smoked. Among these, 1,215 patients underwent lung function tests both at baseline and at the 24-month follow-up, forming the cohort for the longitudinal analyses presented in Parts II and III. Comprehensive baseline characteristics are detailed in Table S1.

Table 1.

Baseline characteristics of the study population (top 15 variables by importance).

Variables Overall (n = 38,863)
Baseline characteristics  
Sex, n (%)  
Female 7,041 (18.1%)
Male 31,817 (81.9%)
Area, n (%)  
Urban 18,370 (47.3%)
Rural 20,493 (52.7%)
Nation, n (%)  
Han 36,316 (93.6%)
Minority 2,494 (6.4%)
Education level, n (%)  
Primary school or below 20,727 (53.4%)
Junior high school 12,503 (32.2%)
High school/Technical secondary school 4,517 (11.6%)
College or above 1,063 (2.7%)
Region, n (%)  
North China 4,935 (12.7%)
Northeast China 1,327 (3.4%)
Central China 5,749 (14.8%)
East China 13,326 (34.3%)
Northwest China 4,161 (10.7%)
South China 2,964 (7.6%)
Southwest China 6,401 (16.5%)
Employment status, n (%)  
Employed 4,702 (12.1%)
Retired 8,671 (22.3%)
Housework 24,149 (62.2%)
Unemployed 1,288 (3.3%)
Income(CNY), n (%)  
<10,000 5,545 (14.3%)
10,000–20,000 8,359 (21.5%)
20,000–50,000 12,162 (31.3%)
50,000–100,000 9,153 (23.6%)
100,000–200,000 2,936 (7.6%)
200,000–300,000 542 (1.4%)
300,000–500,000 99 (0.3%)
500,000–1,000,000 14 (0.0%)
Smoke, n (%)  
Never 10,045 (25.8%)
Former (quit ≥6 months) 9,544 (24.6%)
Current or quit <6 months 19,274 (49.6%)
Respiratory_diseases_before14, n (%)  
No 35,007 (90.1%)
Yes 3,854 (9.9%)
Cold_cough, n (%)  
No 22,214 (57.2%)
Yes 16,647 (42.8%)
Breath_shortness, n (%)  
No 10,036 (25.8%)
Yes 28,825 (74.2%)
Asthma_attack, n (%)  
No 29,292 (75.4%)
Yes 9,569 (24.6%)
Oxygen_inhalation, n (%)  
No 37,566 (96.7%)
Yes 1,295 (3.3%)
Smoke_index (pack-years)  
  36,683; 28.00 [0.00, 47.00]
Missing 2,180 (5.6%)
FEV1_percent.y(%) 38,817; 69.00 [51.80, 84.70]
Open in a new tab

Continuous variables that follow a normal distribution are reported as mean ± standard deviation. Continuous variables that do not follow a normal distribution are reported as median (interquartile range). Categorical variables are reported as percentages.

Unless otherwise noted, missing values for all variables are <2%.

Clinical phenotypic characteristics

The results of the latent class analysis (Table S2 and Figure S2) indicate that as the number of classes increases, the values of the AIC and the BIC consistently decrease, although the rate of decline gradually diminishes. Notably, the AIC and BIC values for the three-class model reach an inflection point. Considering both statistical indicators and clinical interpretability, the three-class model was ultimately selected for clinical phenotyping. Smoking status (Cramer’s V = 0.72), gender (Cramer’s V = 0.67), and predicted FEV1% (η2 = 0.43) emerged as the core variables for distinguishing the three phenotypes (Table S3). This was followed by the smoking index, shortness of breath, and wheezing episodes. The distribution characteristics of the top 15 variables across the clinical phenotypes are presented in Figure 1 and Table S4.

Figure 1.

Heatmap of Z-scores showing the distribution of the top 15 variables across three classes, with variable types indicated by a color bar. The heatmap displays Z-scores for 15 health-related variables across three classes (Class 1, Class 2, and Class 3), with colors ranging from blue (low Z-scores) to red (high Z-scores); the legend shows a Z-score range from −2 to 2. The variables shown are respiratory_diseases_before14, nation, sex, smoke_index, asthma_attack, region, employment_status, oxygen_inhalation, breath_shortness, FEV1_percent.y, cold_cough, smoke, education_level, income, and area. A vertical annotation bar labeled “Type” indicates the variable types: orange for categorical variables and green for continuous variables. Hierarchical clustering highlights relationships among the variables, illustrating varying patterns across the three classes.

Open in a new tab

Heatmap of the distribution of the top 15 variables across phenotypes.

Phenotype 1 (GOLD Stage 2 – predominantly non-smoking females, n = 10,045, 25.8%): This phenotype is characterized by a predominance of non-smoking females (100% non-smokers, 62.1% females), exhibiting moderate lung function impairment (FEV1% of predicted 69.8% [53.2, 86.1]), more severe symptoms, and the lowest educational attainment. Phenotype 2 (GOLD Stage 1 – predominantly smoking males, n = 17,419, 44.8%): This is the largest subgroup, predominantly composed of male smokers (97.3% male, 100% smokers), demonstrating relatively well-preserved lung function (FEV1% of predicted 80.5% [71.0, 91.3]), relatively mild symptoms, and a comparatively high level of education. Phenotype 3 (GOLD Stage 3 – severe smokers, n = 11,399, 29.4%): This group is predominantly male (97%) and consists of heavy smokers (smoking index 41.3 pack-years [25.5, 54.0]), with severe lung function impairment (FEV1% predicted 46.8% [36.4, 55.7]). It exhibits the most severe symptoms, the highest proportion of individuals requiring home oxygen therapy (6.3%), and a lower level of education.

Rapid decliners and clinical phenotype

In a cohort of 1,215 patients who completed the 2-year follow-up, 289 (23.8%) were classified as phenotype 1, 373 (30.7%) as phenotype 2, and 553 (45.5%) as phenotype 3. The overall annual decline in FEV1 was 10.0 mL/year (−80.0, 95.0), while the annual rate of decline in FEV1% predicted was −0.15%/year (−4.00, 3.25). A total of 546 patients (44.9%) were identified as experiencing a rapid decline. The distributions of annual FEV1 decline and the annual rate of decline in FEV1% predicted, along with the threshold for rapid decline, are illustrated in Figure 2A and B. Baseline FEV1 was significantly higher in patients exhibiting a rapid decline pattern compared to those without it [1,645 (1,250, 2,090) mL vs. 1,340 (980, 1,770) mL, p < 0.001]. In contrast, follow-up FEV1 was significantly lower in the rapid-decline group [1,330 (990, 1,820) mL vs. 1,640 (1,210, 2,110) mL, p < 0.001]. A similar trend was observed in FEV1% predicted. The annual decline in FEV1 for the rapid decline group was 105.0 mL/year (65.0, 185.0), while the annual rate of decline in FEV1% predicted was 3.7% (1.9, 7.3) (Table S5).

Figure 2.

Five-panel figure showing distributions of annual FEV1 decline, comparisons across clinical phenotypes via box plots, and stacked bar charts of rapid versus non-rapid decline proportions. Panel A shows a histogram with a density curve of annual FEV1 decline (mL/year), peaking near 0, with a dashed red line marking the rapid-decline threshold at −40 mL/year (annual FEV1 decline of 40 mL or more). Panel B shows a similar histogram for the annual decline rate of FEV1% predicted (%/year), also peaking near 0, with a dashed red line marking the threshold at −1.5%/year (annual decline in FEV1% predicted exceeding 1.5%). Panels C and D show box plots comparing annual FEV1 decline and the annual decline rate of FEV1% predicted across the three clinical phenotypes (Class 1, Class 2, Class 3), with Kruskal-Wallis test p < 0.001 for both. Panel E shows stacked bar charts of the percentage of rapid versus non-rapid decliners within each phenotype: Class 1, 42.6% rapid decline; Class 2, 55.8% rapid decline; Class 3, 38.9% rapid decline (Chi-square test p < 0.001).

Open in a new tab

Rapid decliners and clinical phenotype. (A-B) Annual decline in FEV₁ and annual decline rate of FEV₁% predicted. (C) Annual decline in FEV₁ across the three clinical phenotypes; (D) Annual decline rate of FEV₁% predicted across the three clinical phenotypes; (E) Comparison between rapid decliners and non‑rapid decliners within the clinical phenotypes.

As illustrated in Figure 2C and D, phenotype 2 demonstrated the most pronounced annual decline in FEV1, whereas phenotype 1 displayed the least decline. The distribution of the annual rate of decline in FEV1% predicted across the various phenotypes was consistent with the observed annual decline in FEV1. Notably, significant differences were observed in the proportion of rapid decline among the phenotypes. Figure 2E indicates that the proportion of rapid decline in phenotype 2 was significantly greater than that in phenotype 1 (p = 0.003 after Bonferroni correction) and phenotype 3 (p < 0.001 after Bonferroni correction). In contrast, the difference between phenotype 1 and phenotype 3 was not statistically significant (p = 0.996 after Bonferroni correction).

Model development and performance

The mean age of the longitudinal cohort was 66 years (62,70), with 84.3% of participants being male. Compared to the non-rapidly declining group, patients in the rapidly declining group exhibited higher baseline FEV1 (1.64 L vs. 1.33 L, p < 0.001), a higher predicted FEV1 percentage (64.30% vs. 53.85%, p < 0.001), and a greater proportion of individuals classified as phenotype 2 (38.1% vs. 24.7%). Detailed characteristics are presented in Table S6.

Using rapid lung function decline (annual FEV1 decline ≥40 mL or FEV1% predicted decline >1.5% per year) as the dependent variable, 32 predictor variables remained for feature selection after multicollinearity diagnosis. LASSO regression identified 17 variables (Figure 3A and B), while the Boruta algorithm confirmed 8 features (Figure 3C). Following 500 bootstrap stability tests, six core features with a selection frequency exceeding 80% were ultimately retained: region, fuel type, income level, wheezing episodes, hip circumference, and baseline measured FEV1. As illustrated in Figure 3D–I, the CatBoost model exhibited the best performance (AUC = 0.712), followed by logistic regression (AUC = 0.701) and XGBoost (AUC = 0.692). When the threshold for the CatBoost model was set at 0.365, sensitivity was recorded at 95.5%, the negative predictive value was 86.4%, and specificity was 23.3% (Figure 3D–I). A detailed comparison of the performance of the six machine learning models on the test set is provided in Table S7, and the optimal hyperparameter combination for the CatBoost model is displayed in Table S8.

Figure 3.

Nine-panel figure showing LASSO cross-validation and coefficient paths, Boruta feature selection, and ROC curves with AUC values for six machine learning models. Panel A shows the LASSO cross-validation curve of binomial deviance plotted against log(λ), with the number of variables shown on the top axis and vertical dashed lines at lambda.min and lambda.1se. Panel B shows the LASSO regression coefficient profiles plotted against log(λ), with the same vertical dashed lines. Panel C shows a horizontal bar chart of mean importance from the Boruta algorithm, with confirmed features in green and rejected features in red, including the shadowMax, shadowMean, and shadowMin reference attributes. Panels D–I show receiver operating characteristic (ROC) curves of sensitivity against the false positive rate for six machine learning models: CatBoost (AUC 0.712), Logistic Regression (AUC 0.701), Naive Bayes (AUC 0.692), Random Forests (AUC 0.665), Support Vector Machine (AUC 0.687), and XGBoost (AUC 0.692).

Open in a new tab

Model development and performance. (A) LASSO cross‑validation error curve; (B) LASSO regression coefficient profiles; (C) Feature selection using the Boruta algorithm. (D) ROC curves in CatBoost model; (E) ROC curves in Logistic Regression model; (F) ROC curves in Naive Bayes model; (G) ROC curves in Random Forests model; (H) ROC curves in Support Vector Machines model; (I) ROC curves in XGBoost model.

The SHAP global feature importance analysis (Figure 4A) reveals that the baseline FEV1 value is the most significant predictor, with a mean absolute SHAP value of 0.389. This is followed by region (0.185), income level (0.180), hip circumference (0.178), wheezing episodes (0.163), and fuel type (0.100). The scatter plot (Figure 4B) illustrates that low baseline FEV1 measurements, low income, high hip circumference, absence of wheezing history, and the use of non-clean fuels are associated with an increased risk of rapid decline. The waterfall plots for representative individuals (Figures 4C and D) highlight the contribution of each feature to the predicted probability. The calibration curve of the CatBoost model (Figure 4E) indicates strong model calibration, while the decision curve analysis (Figure 4F) demonstrates that the CatBoost model provides a higher net benefit compared to both the ‘intervene always’ and ‘do nothing’ strategies across a threshold probability range of 0.01–0.83, underscoring the model’s clinical utility. At the optimized threshold of 0.365, the model’s net benefit was 0.328, surpassing that of the ‘intervention for all’ strategy (0.29), thereby illustrating that employing the model to guide intervention decisions can prevent unnecessary interventions. The nomogram, which incorporates six simple variables, facilitates the rapid calculation of an individual’s probability of rapid decline (Figure 5).

Figure 4.

Six-panel figure showing SHAP feature importance, a SHAP beeswarm plot, SHAP waterfall plots for two individual predictions, a calibration curve, and decision curve analysis. Panel A is a bar chart of mean absolute SHAP values ranking feature importance, with FEV1_S.y highest (0.389), followed by region, income, hipline, asthma_attack, and fuel. Panel B is a SHAP summary beeswarm plot showing the distribution of SHAP values for each feature across individual patients, with point color indicating feature values from low (purple) to high (orange). Panels C and D are SHAP waterfall plots for individual predictions: Panel C shows a patient who experienced rapid lung function decline (FEV1_S.y = 0.511, final prediction f(x) = 2.47), and Panel D shows a patient without rapid decline (FEV1_S.y = 2.87, f(x) = −3.39); each waterfall plot starts from the base value E[f(x)] = 0.221 and sequentially adds the contribution of each feature to arrive at the final prediction. Panel E is a calibration curve comparing observed and predicted probabilities, showing the apparent, bias-corrected, and ideal curves with good alignment. Panel F is a decision curve analysis plotting standardized net benefit against threshold probability.

Open in a new tab

Model interpretability and clinical utility evaluation (A) SHAP feature importance bar plot. Features are ranked by their mean absolute SHAP values, which quantify the average impact of each feature on the prediction of rapid lung function decline. Higher values indicate greater feature importance; (B) SHAP summary beeswarm plot. Each point represents a SHAP value for a specific feature across individual patients. The x-axis shows SHAP values: positive values (right side) increase the predicted probability of rapid lung function decline, while negative values (left side) decrease it. The color gradient indicates feature values, from low (deep purple) to high (orange), illustrating how feature values influence the direction and magnitude of the prediction; (C–D) SHAP waterfall plots for individual predictions. Panel C shows a waterfall plot for a patient who experienced rapid lung function decline, while Panel D shows the same for a patient without rapid decline. Each waterfall plot starts from the base value (the average model prediction) and sequentially adds the contribution of each feature (positive contributions in red, negative contributions in blue) to arrive at the final prediction (f(x)). This visualization demonstrates how specific feature values contribute to the individual prediction outcome. (E) Calibration curve; (F) Decision curve analysis (DCA).

Figure 5.

Nomogram with parallel scales for six predictors (hipline, FEV1_S.y, region, fuel, income, asthma_attack) mapping to Points, Total Points, and the predicted risk of rapid lung function decline. The figure is a nomogram consisting of nine parallel scales: a “Points” scale (0–100) at the top; six predictor scales for hipline (70–130), FEV1_S.y (descending from 3 to 0.4), region (values 1–7), fuel (1–2), income (1–3), and asthma_attack (0–1); a “Total Points” scale (0–220); and a “Risk of Rapid Decline” scale (0.1–0.9) at the bottom. To use the nomogram, the value of each predictor is mapped vertically upward to the Points scale, the points from all six predictors are summed to obtain the Total Points, and the Total Points value is mapped downward to read the individualized predicted risk of rapid lung function decline.

Open in a new tab

Nomogram for predicting the risk of rapid lung function decline. To use the nomogram, locate the value of each predictor on its corresponding axis (hipline, FEV1_S.y, region, fuel, income, asthma, attack). Draw a vertical line upward to the “Points” scale to obtain the points for that predictor. Sum the points from all six predictors to obtain the “Total Points”, then draw a vertical line downward from the Total Points axis to the “Predicted Risk” axis to read the individualised probability of rapid lung function decline.

Region codes: 1, North China; 2, Northeast; 3, Central China; 4, East China; 5, Northwest; 6, South China; 7, Southwest. Fuel: 1, clean fuel; 2, unclean fuel. Income (RMB/year): 1, <50,000; 2, 50,000–200,000; 3, >200,000. Asthma attack: 0, no; 1, yes.

Discussion

In this nationwide multicenter cohort of 38,863 patients with COPD, we identified three distinct clinical phenotypes through unsupervised machine learning, among which the GOLD stage 1, predominantly smoking-male phenotype exhibited the highest proportion of rapid lung function decline. A CatBoost model based on six routinely available indicators achieved clinically useful screening performance (AUC 0.712), providing evidence from the Chinese population for precise COPD stratification and individualized management.

The three phenotypes identified through latent class analysis reveal the unique clinical heterogeneity of the Chinese COPD population. Phenotype 1 warrants particular attention due to its predominantly female and non-smoking characteristics. This phenotype is marked by a heavy symptom burden and is concentrated in rural areas, suggesting that non-smoking risk factors—such as exposure to biofuels, indoor air pollution, or early-life respiratory infections—may play a key role in the onset and progression of the disease within this subgroup. Research has confirmed that exposure to biomass fuels is a significant risk factor for COPD, with individuals exposed exhibiting a 2.58-fold increased risk of developing the disease, a risk that is even higher among women [23]. Currently, over 3 billion people worldwide continue to rely on biomass fuels as their primary energy source for basic household needs, the majority of whom reside in low- and middle-income countries [24]. A cluster study of Chinese COPD patients similarly identified a ‘female biomass fuel exposure’ subtype, thereby confirming the presence of this phenotype within the Chinese population and underscoring its public health significance [25]. Consistent with our findings, longitudinal and imaging studies indicate that biomass smoke–related COPD, which predominantly affects never-smoking women, represents an airway-predominant phenotype with less emphysema, greater air trapping, and a slower rate of FEV1 decline than tobacco-related COPD, suggesting partly distinct underlying pathology [26,27]. Western studies have predominantly focused on smoking-related phenotypes [28], and our findings expand the global COPD phenotype spectrum. When formulating strategies for COPD prevention and control in China, screening and intervention of non-smoking risk factors must be considered.

There were significant differences in the trajectories of lung function decline among the three phenotypes, with Phenotype 2 exhibiting the highest proportion of rapid decline, while Phenotype 3 displayed the lowest. This phenomenon may be attributed to the ‘floor effect’—patients with severely impaired baseline lung function have a limited physiological capacity for further decline [29,30]. The SPIROMICS cohort study also confirmed that a steeper decline in lung function is significantly associated with better baseline lung function [11]. As a phenotype of COPD characterized by relatively well-preserved lung function, Phenotype 2 may represent a ‘window of opportunity’ for interventions aimed at disease progression. Implementing measures such as intensified smoking cessation interventions and optimized pharmacotherapy for this population may maximize the delay in disease progression.

Machine learning algorithms provide distinct advantages over traditional statistical methods when handling large-scale, multivariate data, as they effectively capture non-linear relationships and interactions among variables [31]. This study employed a multi-stage feature selection strategy that combined LASSO regression with the Boruta algorithm, enhancing the robustness and interpretability of the selected feature set through Bootstrap stability validation [32]. The model incorporated six variables: region, fuel type, income level, wheezing episodes, hip circumference, and baseline measured FEV1. The AUC values for the six models ranged from 0.66 to 0.71, with the CatBoost model demonstrating the best performance. A literature review of 33 global COPD screening tools indicated AUC values ranging from 0.65 to 0.79 [33]. Another primary-care screening model achieved AUC 0.72 [34], consistent with our model’s performance. Predicting lung function decline is inherently challenging. The ECLIPSE study revealed considerable individual variation in annual FEV1 decline; this substantial heterogeneity complicates accurate trajectory prediction [35]. Dynamic factors including treatment adherence, comorbidity changes, and acute exacerbations may influence trajectories but cannot be fully captured by baseline data [36,37].

Nevertheless, the model, which has an AUC of 0.71, retains significant clinical value as a screening tool. At threshold 0.365, it achieves a sensitivity of 95.5% and a negative predictive value of 86.4%. This indicates that out of 100 actual patients experiencing rapid deterioration, the model can accurately identify 96 cases, fulfilling the primary objective of a screening tool to minimize missed diagnoses [38]. Although the specificity is relatively low, individuals with a positive screening result can proceed to precise assessment such as intensive pulmonary function monitoring, enabling effective risk stratification. This high-sensitivity, two-stage strategy holds significant practical relevance in resource-limited medical settings [39]. The calibration curve demonstrates good consistency, providing reliable risk estimates [40]. Receiver operating characteristic (ROC) curve analysis confirmed that across threshold probabilities 0.01–0.83, the net benefit of interventions guided by the model surpassed that of both ‘universal intervention’ and ‘no intervention’ strategies, reinforcing its clinical utility. To facilitate clinical implementation, this study also constructed a nomogram to enable rapid estimation of an individual’s risk of rapid decline.

Regional disparities emerged as significant predictors in this study, aligning with findings that regional economic development and healthcare access influence disease recognition and management, affecting decline trajectories [41]. Low-income patients may experience more rapid decline due to inadequate treatment, limited health literacy, and poor self-management [42]. The counterintuitive finding that the absence of wheezing contributes positively to rapid decline may relate to ‘non-wheezing COPD’—a subtype with predominant emphysema and atypical airway symptoms. Such patients often experience delayed diagnosis due to absent warning signs, leading to silent but rapid deterioration [43]. Hip circumference may reflect associations between peripheral adipose tissue metabolic activity and systemic inflammatory status [44]. These findings suggest rapid decline represents a comprehensive manifestation of multidimensional factor interplay, including regional healthcare resources, income, symptoms, and body fat—constituting multi-level intervention targets for precision risk stratification.

However, several limitations should be acknowledged. Firstly, lung function decline was estimated from only two spirometric assessments two years apart, which cannot capture the non-linear trajectories of decline; longer follow-up with more frequent measurements is ongoing. Secondly, symptom scores (mMRC and CAT) and exacerbation history were excluded because of missing rates exceeding 20%; consequently, our phenotypes are screening-oriented phenotypes based on routinely obtainable variables and are not directly equivalent to classical phenotypes such as emphysema-dominant or frequent-exacerbator phenotypes. Thirdly, participants were enrolled through a targeted high-risk screening programme rather than population-based sampling, resulting in an over-representation of males and ever-smokers and too few female smokers for stable subgroup estimates; the generalizability of our findings to the broader COPD population and other ethnic cohorts therefore requires further verification. Fourthly, the model was evaluated only by internal validation. Finally, radiological parameters, inflammatory biomarkers, and changes in medication and exacerbations during follow-up were not available and were not incorporated into the model.

Conclusion

Based on a nationwide multicenter cohort, this study identified three clinical phenotypes in COPD patients through latent class analysis and demonstrated that lung function decline trajectories differed significantly across these phenotypes. The phenotype with relatively well-preserved lung function represented a key target group for disease progression interventions. A CatBoost predictive model using six simple indicators serves as an effective screening tool for risk stratification in high-risk rapid decliners. This study provides crucial Chinese population evidence for precise COPD classification and personalized intervention, optimizing public health resource allocation.

Thanks to all the researchers in every sub-research center who work hard in enrollment and follow-ups for this study.

Supplementary Material

Supplemental Material
IANN_A_2735163_SM7831.docx (48.7KB, docx)
Figure S2.tif
IANN_A_2735163_SM5423.tif (556.6KB, tif)
Figure S1.tif
IANN_A_2735163_SM5420.tif (128.2KB, tif)

Funding Statement

This work was supported by the Noncommunicable Chronic Diseases–National Science and Technology Major Project (grant numbers 2023ZD0506306 and 2023ZD0506003) and the Beijing Nova Program (grant number 20250484849).

Disclosure statement

No potential conflict of interest was reported by the author(s).

Availability of data and materials

The datasets generated and/or analysed during the current study are not publicly available due to patient privacy and confidentiality restrictions. They may be made available from the corresponding author on reasonable request, subject to approval from the relevant institutional ethics committee and in compliance with applicable data protection regulations. The corresponding author serves as the principal investigator of The National COPD Screening Programme in China and possesses full authority and access to all data collected under the programme. Data access and utilization permissions for the present study were obtained through the programme’s established governance and authorization framework.

References

  • 1.Celli B, Fabbri L, Criner G, et al. Definition and nomenclature of chronic obstructive pulmonary disease: time for its revision. Am J Respir Crit Care Med. 2022;206(11):1317–1325. doi: 10.1164/rccm.202204-0671PP. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.De Oca MM, Perez-Padilla R, Celli B, et al. The global burden of COPD: epidemiology and effect of prevention strategies. Lancet Respir Med. 2025;13(8):709–724. doi: 10.1016/S2213-2600(24)00339-4. [DOI] [PubMed] [Google Scholar]
  • 3.Zhang Y, Gai X, Chu H, et al. Prevalence of non-smoking chronic obstructive pulmonary disease and its risk factors in China: a systematic review and meta-analysis. BMC Public Health. 2024;24(1):3010. doi: 10.1186/s12889-024-20170-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Zhong N, Wang C, Yao W, et al. Prevalence of chronic obstructive pulmonary disease in China: a large, population-based survey. Am J Respir Crit Care Med. 2007;176(8):753–760. doi: 10.1164/rccm.200612-1749OC. [DOI] [PubMed] [Google Scholar]
  • 5.Singh D, Stockley R, Anzueto A, et al. GOLD Science Committee recommendations for the use of pre- and post-bronchodilator spirometry for the diagnosis of COPD. Eur Respir J. 2025;65(2):2401603. doi: 10.1183/13993003.01603-2024. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Cazzola M, McDonald VM, Stolz D, et al. Emerging therapeutics in COPD: mapping innovation to treatable traits. Lung. 2025;203(1):92. doi: 10.1007/s00408-025-00844-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Xie C, Wang K, Yang K, et al. Toward precision medicine in COPD: phenotypes, endotypes, biomarkers, and treatable traits. Respir Res. 2025;26(1):274. doi: 10.1186/s12931-025-03356-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Appleton LK, Hanania NA, Adrish M.. Personalized COPD care: the Future of precision-based therapies. J Clin Med. 2024;13(21):6339. doi: 10.3390/jcm13216339. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Hou X, Deng Z, Zhou Y, et al. Latent class analysis to identify novel phenotypes in exacerbations of COPD: a retrospective, multicenter cohort study. MedComm (2020). 2025;6(11):e70444. doi: 10.1002/mco2.70444. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Fan J, Cong S, Zhang Y, et al. Associations of pre-COPD indicators with lung function decline and their longitudinal transitions. Pulmonology. 2025;31(1):2486881. doi: 10.1080/25310429.2025.2486881. [DOI] [PubMed] [Google Scholar]
  • 11.Buhr RG, Jackson NJ, Fazio JC, et al. Characteristics associated with lung function trajectories: an analysis of the SPIROMICS cohort. Ann Am Thorac Soc. 2025;22(8):1165–1175. doi: 10.1513/AnnalsATS.202405-500OC. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Rehman AU, Shah S, Abbas G, et al. Assessment of risk factors responsible for rapid deterioration of lung function over a period of one year in patients with chronic obstructive pulmonary disease. Sci Rep. 2021;11(1):13578. doi: 10.1038/s41598-021-92968-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Wang J, Ställberg B, Hårdstedt M, et al. Lung function decline in COPD - relations to changes in symptom burden, inflammation, and comorbidities. COPD. 2025;22(1):2525433. doi: 10.1080/15412555.2025.2525433. [DOI] [PubMed] [Google Scholar]
  • 14.Kim SJ, Lee J, Park YS, et al. Age-related annual decline of lung function in patients with COPD. Int J Chron Obstruct Pulmon Dis. 2016;11:51–60. doi: 10.2147/COPD.S95028. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Bell AJ, Ram S, Labaki WW, et al. Temporal exploration of chronic obstructive pulmonary disease phenotypes: insights from the COPD gene and SPIROMICS cohorts. Am J Respir Crit Care Med. 2025;211(4):569–576. doi: 10.1164/rccm.202401-0127OC. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Yang K, Chen R.. Association between cooking fuel exposure and respiratory health: longitudinal evidence from the China Health and Retirement Longitudinal Study (CHARLS). Ecotoxicol Environ Saf. 2024;275:116247. doi: 10.1016/j.ecoenv.2024.116247. [DOI] [PubMed] [Google Scholar]
  • 17.Wang J, Yu L, Yang Z, et al. Development of chronic obstructive pulmonary disease after a tuberculosis episode in a large, population-based cohort from Eastern China. Int J Epidemiol. 2025;54(2):dyae174. doi: 10.1093/ije/dyae174. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Lei J, Huang K, Wu S, et al. Heterogeneities and impact profiles of early chronic obstructive pulmonary disease status: findings from the China Pulmonary Health Study. Lancet Reg Health West Pac. 2024;45:101021. doi: 10.1016/j.lanwpc.2024.101021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Lei J, Huang K, Pan J, et al. The national COPD screening programme in China: rationale and design. ERJ Open Res. 2023;9(2):00597-2022. doi: 10.1183/23120541.00597-2022. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Fang L, Gao P, Bao H, et al. Chronic obstructive pulmonary disease in China: a nationwide prevalence study. Lancet Respir Med. 2018;6(6):421–430. doi: 10.1016/S2213-2600(18)30103-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Vestbo J, Edwards LD, Scanlon PD, et al. Changes in forced expiratory volume in 1 second overtime in COPD. N Engl J Med. 2011;365(13):1184–1192. doi: 10.1056/NEJMoa1105482. [DOI] [PubMed] [Google Scholar]
  • 22.Wang JM, Labaki WW, Murray S, et al. Machine learning for screening of at-risk, mild and moderate COPD patients at risk of FEV1 decline: results from COPDGene and SPIROMICS. Front Physiol. 2023;14:1144192. doi: 10.3389/fphys.2023.1144192. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Zhang X, Zhu X, Wang X, et al. . Association of Exposure to Biomass Fuels with Occurrence of Chronic Obstructive Pulmonary Disease in Rural Western China: A Real-World Nested Case-Control Study. Int J Chron Obstruct Pulmon Dis. 2023;18:2207–2224. doi: 10.2147/COPD.S417600. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.Pillarisetti A, Daouda M, Gould CF, et al. Household energy use and health in low-income and middle-income countries. Lancet Glob Health. 2026;14(4):e612–e625. doi: 10.1016/S2214-109X(26)00002-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Wang C, Yu F, Cao Z, et al. Exploring COPD patient clusters and associations with health-related quality of life using a machine learning approach: a nationwide cross-sectional study. Engineering. 2025;50:220–228. doi: 10.1016/j.eng.2025.05.005. [DOI] [Google Scholar]
  • 26.Ramírez-Venegas A, Sansores RH, Quintana-Carrillo RH, et al. FEV1 decline in patients with chronic obstructive pulmonary disease associated with biomass exposure. Am J Respir Crit Care Med. 2014;190(9):996–1002. doi: 10.1164/rccm.201404-0720OC. [DOI] [PubMed] [Google Scholar]
  • 27.Camp PG, Ramirez-Venegas A, Sansores RH, et al. COPD phenotypes in biomass smoke- versus tobacco smoke-exposed Mexican women. Eur Respir J. 2014;43(3):725–734. doi: 10.1183/09031936.00206112. [DOI] [PubMed] [Google Scholar]
  • 28.Celli BR, Christenson SA, Rabe KF, et al. Current smoker: A clinical chronic obstructive pulmonary disease phenotype affecting disease progression and response to therapy. Am J Respir Crit Care Med. 2025;211(5):729–736. doi: 10.1164/rccm.202407-1379CI. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Backman H, Blomberg A, Lundquist A, et al. Lung function trajectories and associated mortality among adults with and without airway obstruction. Am J Respir Crit Care Med. 2023;208(10):1063–1074. doi: 10.1164/rccm.202211-2166OC. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Bertels X, Ross JC, Faner R, et al. Clinical relevance of lung function trajectory clusters in middle-aged and older adults. ERJ Open Res. 2024;10(1):00793-2023. doi: 10.1183/23120541.00793-2023. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Chen W, Jiang Y, Noble WS, et al. Error-controlled non-additive interaction discovery in machine learning models. Nat Mach Intell. 2025;7(9):1541–1554. doi: 10.1038/s42256-025-01086-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Huang J, Liu W.. Comparison of machine learning models for predicting stroke risk in hypertensive patients: lasso regression model, random forest model, Boruta algorithm model, and Boruta algorithm combined with Lasso regression model. Medicine (Baltimore). 2025;104(22):e42690. doi: 10.1097/MD.0000000000042690. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Han MK, Steenrod AW, Bacci ED, et al. Identifying patients with undiagnosed COPD in primary care settings: insight from screening tools and epidemiologic studies. Chronic Obstr Pulm Dis. 2015;2(2):103–121. doi: 10.15326/jcopdf.2.2.2014.0152. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Zhang B, Sun D, Niu H, et al. Development of a prediction model to identify undiagnosed chronic obstructive pulmonary disease patients in primary care settings in China. Chin Med J (Engl). 2023;136(6):676–682. doi: 10.1097/CM9.0000000000002448. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Singh D, Criner GJ, Naya I, et al. Measuring disease activity in COPD: is clinically important deterioration the answer? Respir Res. 2020;21(1):134. doi: 10.1186/s12931-020-01387-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Meeraus WH, Mullerova H, El Baou C, et al. Predicting re-exacerbation timing and understanding prolonged exacerbations: An analysis of patients with COPD in the ECLIPSE cohort. Int J Chron Obstruct Pulmon Dis. 2021;16:225–244. doi: 10.2147/COPD.S279315. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Suri R, Non A, Bailey J, et al. The long-term effects of cost-related nonadherence on COPD outcomes and progression in the COPDgene study cohort. Chronic Obstr Pulm Dis. 2026;13(1):1–7. doi: 10.15326/jcopdf.2025.0689. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.Hyndman ME, Paproski RJ, Kinnaird A, et al. Development of an effective predictive screening tool for prostate cancer using the ClarityDX machine learning platform. NPJ Digit Med. 2024;7(1):163. doi: 10.1038/s41746-024-01167-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Mohan S, Cárdenas MK, Ricciardi F, et al. Cost-accuracy analysis of chronic obstructive pulmonary disease screening in low- and middle-income countries. Am J Respir Crit Care Med. 2022;206(3):353–356. doi: 10.1164/rccm.202201-0071LE. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Austin PC, Harrell FE, Van Klaveren D.. Graphical calibration curves and the integrated calibration index (ICI) for survival models. Stat Med. 2020;39(21):2714–2742. doi: 10.1002/sim.8570. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41.Wang C, Qi W, Yang T, et al. The care cascade of chronic obstructive pulmonary disease in China: a cross-sectional study of individual-level data at enrolment into the national “Happy Breathing” Programme. EClinicalMedicine. 2024;74:102597. doi: 10.1016/j.eclinm.2024.102597. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42.Ying Y, Khunthason S, Apidechkul T, et al. Influencing factors of good quality of life among chronic obstructive pulmonary disease patients living in Zhejiang Province, China. Sci Rep. 2024;14(1):8687. doi: 10.1038/s41598-024-59289-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43.Castaldi PJ, Xu Z, Young KA, et al. Heterogeneity and progression of chronic obstructive pulmonary disease: emphysema-predominant and non-emphysema-predominant disease. Am J Epidemiol. 2023;192(10):1647–1658. doi: 10.1093/aje/kwad114. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Zhang S-J, Qin X-Z, Zhou J, et al. Adipocyte dysfunction promotes lung inflammation and aberrant repair: a potential target of COPD. Front Endocrinol (Lausanne). 2023;14:1204744. doi: 10.3389/fendo.2023.1204744. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplemental Material
IANN_A_2735163_SM7831.docx (48.7KB, docx)
Figure S2.tif
IANN_A_2735163_SM5423.tif (556.6KB, tif)
Figure S1.tif
IANN_A_2735163_SM5420.tif (128.2KB, tif)

Data Availability Statement

The datasets generated and/or analysed during the current study are not publicly available due to patient privacy and confidentiality restrictions. They may be made available from the corresponding author on reasonable request, subject to approval from the relevant institutional ethics committee and in compliance with applicable data protection regulations. The corresponding author serves as the principal investigator of The National COPD Screening Programme in China and possesses full authority and access to all data collected under the programme. Data access and utilization permissions for the present study were obtained through the programme’s established governance and authorization framework.

Articles from Annals of Medicine are provided here courtesy of Taylor & Francis

RESOURCES