Chronic obstructive pulmonary disease (COPD) causes long-term blockage of airflow and is a major cause of illness and death, and it greatly reduces quality of life. Long-lasting inflammation in the airways and changes in airway structure are common, and pulmonary hypertension (PH) is another important factor that affects outcome (1, 2). The link between COPD and PH comes from several processes. These include changes in the lung blood vessels, inflammatory responses driven by cytokine signals, and the control of genetic risk. Lung tissue (parenchymal) damage relates to how often PH occurs and how severe it is, but there is a lot of variation between patients: some people with only mild changes in the lungs develop severe PH, while others with advanced disease have only small increases in pulmonary artery pressure (2–4). This mismatch between lung tissue damage and blood vessel changes suggests that changes inside the vessels, inflammatory mediators, and genetic factors may all work together to cause PH in COPD, even when parenchymal disease is not very severe (5, 6).

Because of these differences between patients and their effects on outcome, it is very important to tell apart pulmonary hypertension and non-pulmonary hypertension in people with chronic obstructive pulmonary disease. This is important because it affects treatment plans, decisions about who needs urgent care, and how often patients should be checked (7). Early detection enables timely, targeted interventions, which are associated with a reduction in hospitalizations and an improvement in quality of life (8). Precise differentiation between COPD with and without PH also supports phenotype-informed treatment strategies (9). Despite its clinical importance, PH is still frequently missed or recognized late in patients with COPD.

Early detection of PH in COPD is hampered by nonspecific symptoms such as exertional dyspnea and chest tightness, which are difficult to attribute to vascular versus airway pathology and often delay further investigation. Delayed diagnosis is associated with worse outcomes (10, 11). Right heart catheterization remains the diagnostic gold standard, but its invasiveness and resource requirements limit use in routine practice (12–14). Echocardiography provides noninvasive assessment of right-heart structure and pressure surrogates but cannot deliver continuous monitoring and may be constrained by hyperinflation-related acoustic windows (15, 16). In low- and middle-income countries, shortages of trained personnel and equipment further contribute to underdiagnosis and fragmented longitudinal management (17). Several inexpensive and widely available biomarkers—such as mean platelet volume (MPV), red blood cell distribution width (RDW), brain natriuretic peptide (BNP), the pulmonary artery–to–aortic diameter ratio, and the neutrophil-to-lymphocyte ratio (NLR)—have been associated with PH risk in COPD (18–21). But when these biomarkers are used alone, they do not have enough power to clearly separate patients into different risk groups. Because of this problem, researchers are studying machine learning (ML) methods that bring together different kinds of data and improve diagnostic accuracy.

ML methods can use many types of data in COPD at the same time, such as clinical features, pulmonary function tests, vascular indicators from CT, and blood biomarkers. By using these different inputs together, ML models can find complex risk patterns. With these high-dimensional data, ML models can diagnose disease more accurately and more completely than methods that rely on fixed rules or a single measurement, and they can give quantitative support for decisions in clinical practice (22, 23). Recent studies show that ML can increase the detection rate of pulmonary hypertension in patients with chronic lung disease, but reliable and easy-to-explain tools for pulmonary hypertension related to COPD are still not available.

ML-based methods may help provide individualized risk assessment and improve the accuracy of detecting PH in patients with COPD. But there are still questions about whether these methods work well in many different clinical settings, how clear their decisions are, and how easily they can be added to daily clinical work (24, 25). Solving these problems is important for safe and wide use in clinical practice. Accordingly, this study aimed to develop and internally validate an interpretable ML-based risk-prediction model for PH in COPD. The model is designed to (a) estimate individualized PH risk, (b) clarify feature-level relationships between COPD and PH, and (c) support phenotype-informed therapeutic planning. The intended end-users are respiratory and cardiovascular specialists, and the intended use is early in-hospital risk screening and referral stratification among hospitalized patients with COPD, to help determine whether further advanced examinations, such as echocardiography or right heart catheterization are warranted.

This single-center, retrospective, cross-sectional observational study was conducted at Beijing Chaoyang Hospital. We analyzed inpatients with COPD, with and without PH, admitted to the Department of Respiratory and Critical Care Medicine between January 2014 and September 2024. Clinical and laboratory data were obtained from the hospital electronic medical record system. The healthcare setting, eligibility criteria, outcome definition, and predictor measurements were identical for participants later allocated to the training and test sets.

Eligible participants were adults (aged 18–95 years) with a diagnosis of COPD confirmed by pulmonary function testing in accordance with Global Initiative for Chronic Obstructive Lung Disease (GOLD) guidelines.

We excluded patients with: (a) idiopathic pulmonary arterial hypertension (PAH), pulmonary thromboembolism (PTE), or PH associated with congenital heart disease; (b) severe cardiovascular or cerebrovascular disease, or hepatic or renal insufficiency; or (c) active malignancy. For patients with multiple hospitalizations during the study period, only the first eligible admission was considered to avoid duplicate observations.

All consecutive patients meeting the criteria during the accrual period were included, yielding a total of 523 hospitalized patients with COPD, of whom 176 (33.6%) met diagnostic criteria for COPD-PH (i.e., outcome events). According to commonly cited standards for multivariable prediction models, an events-per-predictor (EPV) ratio of at least 10 is often recommended to reduce overfitting and enhance parameter stability. In the present study, the final CatBoost model included 18 predictors and 176 outcome events, corresponding to an EPV of approximately 9.8, which is close to this conventional target and exceeds the minimum EPV threshold of 5 that some methodological work and guidelines still consider acceptable. This EPV suggests that the available data were adequate to support model development and internal validation.

Guided by expert consensus and prior evidence, 39 candidate predictors were prespecified and grouped into six domains: demographics, personal history, medical history, laboratory indices, lung function, and echocardiographic parameters (Table 1). TRV was excluded a priori from candidate predictors because it formed part of the echocardiographic outcome definition (TRV > 2.8 m/s), thereby avoiding incorporation bias.

After random shuffling, the 523 patients were split into a training set (n = 423) and an independent test set (n = 100) in an 8:2 ratio. Model selection and stability were assessed using stratified five-fold cross-validation within the training set. The optimal model was then refitted on the full training set and evaluated on the test set. Decision thresholds for all performance metrics were fixed during cross-validation and were not adjusted on the test set.

Variables with less than 30% missing values were filled in using the k-nearest neighbors (kNN) method. Missing values for each variable are shown in Supplementary Table 1 and in Figure 1. In Figure 1, each column shows one variable, and white cells show missing values. Variables with 30% or more missing values were not used in model development.

To deal with class imbalance, we used synthetic oversampling of the minority class (SMOTE) on the training data in each cross-validation fold before model fitting. This method lowers bias toward the majority class and keeps the original feature space.

The initial 39 candidate predictors underwent multicollinearity testing, visualized using a correlation matrix to generate a heatmap (Figure 2). Features with an absolute pairwise Pearson correlation coefficient (|r|) > 0.80 were removed to limit redundancy and reduce estimator instability. The |r| = 0.80 threshold was chosen as a pragmatic, commonly used cutoff in applied epidemiology and machine-learning studies to flag near-duplicate variables while preserving clinically distinct information, thereby balancing model stability and information retention (28).

Recursive feature elimination (RFE) was then applied to the pruned feature set. All correlation-based screening and RFE procedures were performed within each training fold and applied only to the corresponding held-out fold to avoid information leakage. Based on these procedures, a final set of 18 predictors was selected and subsequently used for model training in the full training set and performance evaluation in the independent test set.

Before model fitting, continuous predictors were standardized using z-score transformation, and binary or categorical variables were encoded as dummy (one-hot) indicators for all algorithms except CatBoost, which used its native handling of categorical features. A unified machine-learning pipeline was applied to the retained predictors using eight algorithms: Categorical Boosting (CatBoost), Random Forest (RF), Gradient Boosting Machine (GBM), Adaptive Boosting (AdaBoost), Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), k-Nearest Neighbors (kNN), and Multilayer Perceptron (MLP).

Within this pipeline, data preprocessing and grid-search hyperparameter optimization were carried out prior to final model fitting; algorithm-specific hyperparameters are summarized in Supplementary Table 2.

Model performance was evaluated using stratified five-fold cross-validation in the training set and subsequently in the independent test set. Discrimination was summarized using the receiver operating characteristic (ROC) curve and area under the curve (AUC). For threshold-based performance, we reported accuracy, sensitivity, specificity, F1 score, positive predictive value (PPV), and negative predictive value (NPV).

For each model, performance across the five validation folds was summarized as the mean and 95% confidence interval (95% CI), providing estimates of central tendency and sampling variability. Each model produced an individual-level predicted probability of COPD-PH.

We used decision curve analysis (DCA) to measure the possible clinical benefits over a range of reasonable intervention thresholds. In addition, we measured the marginal contribution of each predictor variable to the model output by calculating SHAP values (Shapley Additive Explanations). We did this at the cohort level and at the individual patient level. This helped make the model easier to understand in clinical practice.

All statistical analyses were done in R (version 4.4.1) and Python (version 3.12.4). We used R mainly for data management and standard statistical analyses. We used the packages readxl (for data import), dplyr (for data handling), boot (for resampling-based inference), and effsize (for effect size estimation). We used Python to compute SHAP values and plots to explain features in the prediction models.

We compared basic features between the training set and the test set with the Mann–Whitney U test. For categorical variables we used Fisher’s exact test. We set the significance level at α = 0.05. By checking similarity in key covariates between the training and test sets, we assessed possible selection bias and supported the model performance estimates.

We then used Cohen’s d for paired samples to measure the practical size of differences in predictive performance between models. We estimated its 95% confidence interval with 1,000 non-parametric resampling runs. We interpreted the effect size with common cut points: |d| < 0.2 (no practical difference), 0.2 ≤ |d| < 0.5 (small), 0.5 ≤ |d| < 0.8 (moderate), and |d| ≥ 0.8 (large). A larger absolute value means a stronger practical effect (29).

After application of the inclusion and exclusion criteria, 523 hospitalized patients with COPD were included in the analysis. Supplementary Table 3 presents baseline characteristics for the training and test sets, and Table 2 compares patients with COPD-PH and those with COPD alone.

Correlation analysis between the 39 candidate predictors and COPD-PH status (Figure 3) highlighted right-heart dimensional indices and central pulmonary arterial measures as the most informative features—specifically right ventricular end-diastolic diameter (RVD), right atrial transverse diameter (RA TD), pulmonary artery diameter (PA), right atrial longitudinal diameter (RA LD), and the pulmonary artery-to-aorta diameter ratio (PA/AO).

Across 40 baseline variables, 36 (90.0%) showed no statistically significant difference between the training and test sets (p ≥ 0.05; Supplementary Table 3). The four variables with statistically significant differences were smoking index, NT-proBNP, RA TD, and RVD. For RA TD and RVD, the median differences were both 1 mm (approximately 3% relative difference), with interquartile range (IQR) overlap coefficients of 0.67 and 0.75, respectively, indicating highly overlapping distributions and minimal differences in magnitude. NT-proBNP showed an approximately 12% relative difference in medians and smoking index about 40%; nonetheless, IQR overlap remained substantial (0.52 and 0.80, respectively).

Of the 523 included patients, 176 (33.6%) had COPD-PH and 347 (66.4%) had COPD alone. After random allocation, the training set comprised 423 patients (143 COPD-PH, 280 COPD alone), and the independent test set comprised 100 patients (33 COPD-PH, 67 COPD alone). Overall, the distributions of key predictors and outcomes were similar between the training and test sets, with only small differences in a few variables (e.g., smoking index, NT-proBNP, RA TD, and RVD), supporting the comparability of the development and evaluation datasets.

Among the 176 COPD-PH patients, 133 (75.6%) were diagnosed by echocardiography and 43 (24.4%) were confirmed by RHC. Baseline characteristics were partly different between the RHC-diagnosed and echo-diagnosed subgroups (Supplementary Table 4). In general, the RHC-diagnosed subgroup showed features consistent with greater disease severity and right-heart/pulmonary-artery remodeling (e.g., lower PaO₂ and DLCO and higher NT-proBNP and uric acid, together with larger right-heart or pulmonary-artery dimensions), which likely reflects catheterization referral patterns rather than contradictory diagnostic definitions.

The flow of participants through the study, including numbers screened, excluded, and finally included with and without COPD-PH, is shown in Figure 4.

Eight candidate models were compared using seven core performance metrics. Overall cross-validated performance in the training set is summarized in Table 3, and fold-specific results are provided in Supplementary Table 5. The ROC curves from five-fold cross-validation in the training set are shown in Figure 5, illustrating the consistency of discriminative performance across validation folds.

CatBoost demonstrated the best overall internal performance among the eight algorithms. Its fold-specific ROC curves are shown in Figure 6 and indicate stable discrimination across validation folds. In cross-validation, CatBoost achieved an average AUC of 0.9337, accuracy of 0.8696, and F1 score of 0.8732.

Across 196 pairwise model comparisons, effect size analysis using Cohen’s d showed large effects (|d| ≥ 0.8) versus traditional baselines (kNN, MLP), corresponding to 15–31% relative improvements in AUC and accuracy. In contrast, effect sizes versus XGBoost were mostly small or negligible (|d| < 0.5), suggesting broadly comparable predictive performance, although CatBoost was more efficient and easier to tune in practice. Among 49 valid pairwise comparisons favoring CatBoost, 81.6% showed medium-to-large and 75.5% large effects, supporting its selection as the primary model for subsequent test-set evaluation.

On the independent test set, using the decision threshold fixed during cross-validation, CatBoost maintained superior out-of-sample performance (Figure 7). Test-set results for all eight models are summarized in Table 4. CatBoost achieved an AUC of 0.848, accuracy of 0.830, F1 score of 0.746, sensitivity of 0.758, and specificity of 0.866, with balanced PPV and NPV. Random Forest ranked second (AUC 0.815, accuracy 0.820, F1 0.719), while logistic regression (AUC 0.806, F1 0.700) and GBM (AUC 0.804, F1 0.684) showed moderate performance; XGBoost was comparatively weaker (AUC 0.756, F1 0.683).

Decision curve analysis in the test set showed that, across an approximate threshold probability range of 0.10–0.60, the net benefit of the CatBoost model consistently exceeded that of the other top-performing models (Figure 8), indicating greater potential clinical decision-making value. Taken together, these findings support CatBoost as the preferred model, with the best overall balance of discrimination and clinical utility for subsequent application.

We used SHAP to interrogate the CatBoost model and quantify feature-level contributions at both cohort and individual levels. In the global importance plot (Figure 9), right ventricular diameter (RVD) and pulmonary artery diameter were the dominant structural predictors. Arterial blood gases (PaCO₂, PaO₂) reflected gas-exchange impairment, while right atrial transverse and longitudinal diameters (RA TD, RA LD) captured right-sided cardiac remodeling. NT-proBNP and the neutrophil-to-lymphocyte ratio (NLR) indexed cardiac strain and systemic inflammation. Spirometric indices (FEV₁/FVC, FEV₁% predicted), hematologic markers (white blood cell and lymphocyte counts, hemoglobin, RDW-CV), and additional clinical or imaging variables (creatinine, BMI, PA/AO) contributed additional, though more modest, discriminatory information.

Figure 10 illustrates case-level feature attributions for representative patients across GOLD 1–4 airflow limitation categories. In the GOLD 1 exemplar, enlarged right atrial diameters (RA TD 37 mm, RA LD 47 mm) and a white blood cell count of 5.7 × 10⁹/L have positive SHAP contributions, shifting the prediction toward higher PH risk, whereas a supranormal FVC % predicted (133%) and a relatively small main pulmonary artery diameter (24 mm) have negative SHAP contributions that attenuate the overall risk estimate.

Figure 11 (SHAP dependence plots) highlights clinically interpretable regions in which the model’s predicted PH risk increases, such as pulmonary artery diameter > 25 mm, PaCO₂ > 50 mm Hg, and RVD > 30 mm. These values are not intended as prespecified clinical cut-offs but rather as data-driven inflection points that may inform triage decisions and generate hypotheses for future studies.

Interpretability analyses using subgroup SHAP summaries suggested broadly consistent global feature-importance patterns between the RHC-diagnosed and echo-diagnosed subgroups (Figure 12). Because the RHC-diagnosed subgroup was small, subgroup-specific discrimination metrics were considered exploratory and were not emphasized; we therefore focused on the consistency of model explanations using subgroup SHAP summaries.

We developed and internally validated a machine-learning model to estimate PH risk in patients with COPD. Among the eight candidate algorithms, CatBoost provided the strongest overall discrimination and was therefore selected as the final model for downstream evaluation and interpretation.

In clinical risk stratification, CatBoost offers multiple practical advantages: it supports the handling of categorical variables within electronic health records, effectively mitigates overfitting through its symmetric tree structure and ordered boosting mechanism (30), and integrates multi-modal data sources without requiring complex feature engineering (31). Furthermore, its built-in missing value handling strategy enhances model robustness in real-world datasets containing incomplete observations, establishing CatBoost as a pragmatic and reliable choice for clinical decision support applications (32).

Based on SHAP feature ranking of the CatBoost model, the five most important predictors were right ventricular diameter, pulmonary artery diameter, arterial partial pressure of carbon dioxide, right atrial transverse diameter, and age. Together, these variables give information about demographic characteristics, heart structure, and breathing function in patients with chronic obstructive pulmonary disease and pulmonary hypertension (COPD-PH).

This study has several limitations. First, the single-center, retrospective design may introduce selection bias, information bias, and unmeasured confounding. Second, we did not perform external validation; therefore, transportability to other case-mix profiles, imaging protocols, and laboratory platforms remains uncertain and should be tested prospectively. Third, PH was primarily ascertained by echocardiography; although appropriate for screening, it may misclassify cases compared with RHC, the hemodynamic gold standard. The RHC-diagnosed subgroup was small, limiting the precision of subgroup-specific estimates; thus, subgroup findings were considered descriptive/exploratory and warrant confirmation in larger cohorts with broader RHC verification. Finally, some clinically relevant candidate predictors—most notably DLCO—were excluded because of substantial missingness under our prespecified ≥30% threshold, which may have limited model generalizability; future studies with standardized diffusion testing should evaluate the incremental value of DLCO.

We plan to improve the model so that it can be built into the electronic medical record (EMR). In this setting, the system will automatically pull routinely collected variables and will ask clinicians to add any missing key predictors before it estimates risk. When non-critical predictors are missing or seem unlikely, the system should mark these values for checking and either ask for confirmation or continue using only the other confirmed predictors. The main users will be pulmonology and cardiology specialists and residents who manage COPD-PH. They will not need special machine learning skills, only a clear understanding of the COPD-PH risk probabilities and the related risk groups given by the model.

Future work should focus on multicenter collaborative datasets and prospective external validation. This will test how well the model works in different patient groups, imaging protocols, and laboratory platforms. Adding more cases confirmed by right heart catheterization would also improve diagnostic accuracy for pulmonary hypertension. In addition, future studies should examine how model-based risk stratification changes daily clinical care, including diagnostic test requests, referral patterns, treatment choices, and patient-centered outcomes. Ideally, the model should be tested in prospective trials that follow real clinical pathways in different healthcare systems.