Abstract
Purpose
Patients with moderate-to-severe chronic obstructive pulmonary disease (COPD) face a high risk of three-month unplanned readmission after discharge, yet early identification of high-risk patients remains challenging in clinical practice. This study aimed to develop and validate machine learning (ML) models for predicting the risk and providing a reference for the subsequent development and updating of predictive tools.
Methods
This single-center retrospective cohort study involved individuals with moderate-to-severe COPD from Tianjin Chest Hospital (January 2023 to January 2026). Participants were randomly split into a 70% training set and a 30% validation set. Least absolute shrinkage and selection operator (LASSO) regression identified candidate predictors for model development. Six models were constructed: logistic regression (LR), support vector machine (SVM), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Backpropagation Neural Network (BPNN), and Naive Bayes (NB). Model performance was assessed through the area under the receiver operating characteristic (ROC) curve (AUC), sensitivity (SEN), and specificity (SPE). Calibration was evaluated through calibration curves, and clinical utility via decision curve analysis (DCA).
Results
299 patients with moderate-to-severe COPD were enrolled, and 59 (20%) experienced unplanned readmission within three months after discharge. Six ML models were compared, and the XGBoost model exhibited the best performance in the validation cohort, with an AUC of 0.809 (95% CI: 0.690–0.927), SEN of 0.750, and SPE of 0.826. SHapley Additive exPlanations (SHAP) analysis ranked the key predictive factors in descending order: history of hospitalization in the past year, FEV1, history of coronary heart disease, and smoking history.
Conclusion
XGBoost model showed promising performance in internal validation for predicting the risk of unplanned readmission among patients with moderate-to-severe COPD. However, these findings should be interpreted cautiously given the small sample size, single-center design. Future studies should use prospective, multicenter data from diverse geographic regions to externally validate and update the model.
Keywords: COPD, moderate-to-severe, readmission, ML
Introduction
Chronic obstructive pulmonary disease (COPD) is one of the most prevalent chronic respiratory diseases worldwide and is characterized by persistent airflow limitation, recurrent acute exacerbations, and progressive disease deterioration. With ongoing population aging and the continued presence of environmental risk factors, the global burden of COPD is expected to increase substantially. It has been projected that the number of COPD cases will rise by 23.3% by 2050,1 while the cumulative global economic burden is estimated to approach US$40 trillion.2 As per the Global Initiative for Chronic Obstructive Lung Disease (GOLD) guidelines, a post-bronchodilator forced expiratory volume in one second (FEV1)/forced vital capacity (FVC) ratio <0.7 constitutes the diagnostic gold standard for COPD.3 The severity of airflow obstruction is in four grades according to the percentage of FEV1to the predicted value (FEV1%pred), with GOLD grades 2–4 representing moderate to severe COPD.3 Patients with moderate-to-severe COPD constitute a major focus of clinical management, as they have a higher risk of adverse outcomes and readmission. Studies have indicated that lower FEV1 levels are associated with poorer exercise tolerance and airway defense, thereby increasing the risk of COPD readmission.4 In addition to the risk of readmission, declining pulmonary function is also associated with a higher risk of mortality. Evidence indicates that the population with GOLD grade 2 exhibits a hazard ratio (HR) for mortality of 2.1 (95% CI, 1.7–2.6), or a higher HR of 4.2 (95% CI, 2.7–6.5),5 underscoring the significant increase in mortality risk with FEV1 decreases.6 This further highlights the increased risk of readmission and adverse outcomes among patients with moderate-to-severe COPD compared with those with mild COPD.
Patients with moderate-to-severe COPD often experience rapid clinical deterioration during acute exacerbations, frequently requiring hospitalization. Those with severe disease may require ventilatory support and can progress to respiratory failure, impaired consciousness, or intensive care unit (ICU) admission.7,8 Studies have demonstrated that the moderate or severe COPD population, in comparison to mild COPD patients, experiences longer hospital stays and higher mortality during exacerbations.9 Each exacerbation accelerates the decline in pulmonary function, with a subset of patients failing to recover baseline lung function even eight weeks after the event.10 Exacerbations also increase the risk of comorbid conditions like myocardial infarction (MI), pulmonary embolism, and stroke.11,12 Frequent exacerbation-related readmissions further elevate mortality risk. More than one-third of the COPD population is readmitted within 90 days after discharge due to disease deterioration, which both severely diminishes life quality and places a considerable burden on healthcare systems.13,14 Inaccurate risk stratification may lead to high-risk patients being overlooked and missing opportunities for early intervention, while low-risk patients may receive unnecessary interventions, resulting in inefficient use of limited healthcare resources. Accurately identifying patients at high risk of early readmission following discharge is essential for implementing individualized management strategies. Evidence suggests that interventions such as optimization of discharge medications, standardized inhaler education, early post-discharge follow-up, individualized pulmonary rehabilitation, and self-management education can effectively reduce acute exacerbations and unplanned hospital readmissions among high-risk patients.15,16
In recent years, various predictive models and biomarkers have been proposed, including clinical scores developed using traditional statistical methods, machine learning (ML) models based on electronic health records, and biomarker-based prediction models.17–19 Traditional scoring systems used in clinical practice to assess COPD prognosis include the Age, Dyspnea, and Airflow Obstruction (ADO) score,20 the Body Mass Index (BMI), Airflow Obstruction, Dyspnea, and Severe Exacerbations (BODEx) score,21 the Comorbidity, Obstruction, Dyspnea, and Severe Exacerbations (CODEX) index,22 the Dyspnea, Obstruction, Smoking, and Exacerbations (DOSE) score,23 and the Previous admissions, Extended Medical Research Council dyspnea score, Age, Right heart failure, and Left heart failure (PEARL) score.24 However, the ADO, BODEx, CODEX, and DOSE scores were originally developed primarily to predict long-term all-cause mortality and may have limited predictive performance for short-term unplanned readmission. The CODEX index25 was developed to improve prognostic assessment by incorporating comorbidities, airflow obstruction, dyspnea, and severe exacerbations. However, it may be difficult to distinguish whether its prognostic value reflects outcomes associated with COPD exacerbations or the generally poorer prognosis of patients with multiple comorbidities. The PEARL score uses 90-day readmission or death as a composite endpoint, which may limit its ability to specifically identify readmission risk among patients with moderate-to-severe airflow limitation.24 Most existing ML models based on electronic health records and biomarker-based prediction models have been developed in general COPD populations encompassing a broad spectrum of disease severity, from mild to very severe.26–29 Consequently, the predictor weights in these models reflect patterns across different severity levels, potentially limiting their applicability to patients with moderate-to-severe COPD. Vital signs and laboratory parameters may differ substantially between patients with moderate-to-severe COPD and those with mild disease, potentially resulting in differences in the predictors of readmission. Moreover, the incidence of unplanned readmission within three months is higher among patients with moderate-to-severe COPD than among those with mild disease.4,30 Furthermore, the clinical implementation of these models is constrained by interpretability and data availability. Models with limited interpretability may be less readily accepted by clinicians and provide insufficient insight to guide targeted interventions. Furthermore, some existing models rely on complex variables or specialized biomarkers,31 which may increase costs and delay risk assessment, limiting their suitability for timely bedside decision-making. Therefore, developing interpretable prediction models based on routinely available clinical variables is essential for facilitating clinical implementation. From a methodological perspective, traditional logistic regression (LR) may not adequately capture nonlinear relationships between predictors and outcomes. Accordingly, this study developed an LR model alongside multiple ML models and compared their performance in the validation cohort in terms of discrimination and calibration to identify the best-performing model.
Therefore, this study aimed to develop and internally validate interpretable ML models for predicting unplanned readmission within three months among patients with moderate-to-severe COPD, compare the performance of multiple algorithms, and identify the best-performing model.
Materials and Methods
Populations
This retrospective cohort analysis leveraged clinical data from the moderate-to-severe COPD population hospitalized from January 2023 to January 2026 in the Department of Respiratory and Critical Care Medicine at Tianjin Chest Hospital. Inclusion criteria were: 1) diagnosed with moderate to severe COPD as per the 2025 GOLD guidelines, and 2) complete follow-up records. Exclusion criteria encompassed: 1) patients lost to follow-up, and 2) patients with concomitant chronic pulmonary diseases like bronchiectasis, asthma, or pulmonary neoplasms.
Covariates
Patients’ clinical data were retrospectively collected, including demographic information: sex, age, BMI, smoking history, COPD duration and severity, acute exacerbations and hospitalizations in the past year, and comorbidities, such as hypertension, diabetes mellitus (DM), coronary heart disease (CHD), heart failure (HF), MI, and cerebrovascular infarction. Laboratory parameters encompassed complete blood count, alanine aminotransferase (ALT), aspartate aminotransferase (AST), creatinine, C-reactive protein (CRP), procalcitonin (PCT), B-type natriuretic peptide (BNP), as well as arterial blood gas analysis (eg, PH, PCO2, PO2, lactate). Pulmonary function indices encompassed FEV1, FEV1%pred, FEV1/FVC, and the percentage of carbon monoxide diffusing capacity to the predicted value (DLCO%pred). The following were also recorded: respiratory support interventions such as non-invasive ventilation(NIV), invasive mechanical ventilation (IMV), oxygen therapy (OT), and corticosteroid use, ICU admission, and in-hospital complications.
Diagnosis Criteria for Outcome Variables
Unplanned readmission was defined as a patient’s re-hospitalization within a short period after discharge due to an acute exacerbation of COPD, which was not pre-scheduled.32 Discharged patients were followed up for three months via outpatient visits or telephone, and those meeting criteria for unplanned readmission within three months were identified for analysis.
Handling of Missing Data
Multiple imputation by chained equations (MICE) was performed when the proportion of missing data for a variable was below 20%. The predictor matrix included all other variables in the dataset. Imputation was performed separately in the training and validation sets, with three imputed datasets generated for each set. However, because the proportion of missing data was very low, the results were highly similar across the imputed datasets. For presentation, the imputed dataset yielding the median validation performance among the imputed datasets was selected based on the performance of the best-performing model.
Statistical Analysis
The sample size was estimated based on the events-per-variable (EPV) principle, with a recommended minimum EPV of 10. This meant that the number of events (patients experiencing unplanned readmission) should be at least 10 times the number of variables included in the model. Previous studies have commonly included 4–10 variables in prediction models, as including too many variables increases model complexity and may limit clinical applicability. Assuming that five variables would ultimately be included in the model, at least 50 readmission events would be required. Given a previously reported unplanned readmission rate of approximately 20%, the minimum required sample size was estimated to be approximately 250 patients.
Continuous variables in normal distribution were shown as mean ± standard deviation (SD) and compared employing independent samples t-tests. Non-normally distributed variables were presented as median with interquartile range (IQR) and compared via the rank-sum test. Categorical variables were shown as counts and percentages and compared via the chi-square test.
The dataset was randomized into training and validation sets at a 7:3 ratio. In the training set, least absolute shrinkage and selection operator (LASSO) regression was used for variable selection, with five-fold cross-validation to determine the optimal penalty parameter (λ). The variance inflation factor (VIF) was used to assess multicollinearity, and variables with a VIF > 5 were excluded. In the training cohort, LR, Extreme Gradient Boosting (XGBoost), support vector machine (SVM), Light Gradient Boosting Machine (LightGBM), Backpropagation Neural Network (BPNN), and Naive Bayes (NB) models were developed using LASSO-selected variables. During model development, repeated stratified five-fold cross-validation combined with grid search was used to optimize the hyperparameters of the SVM (penalty parameter C: 0.05 and 5; kernel width σ: 0.01 and 2), XGBoost (number of iterations: 200, 400, 600, and 800; maximum tree depth: 2, 3, and 4; learning rate: 0.01, 0.05, and 0.1; minimum child weight: 0.8, 1.0, and 2.0), LightGBM (number of iterations: 200, 400, 600, and 800; learning rate: 0.01, 0.05, and 0.1; number of leaves: 8, 16, and 24; minimum number of samples per leaf: 5, 10, and 20), and BPNN (number of hidden-layer nodes: 3, 4, and 5; weight decay: 0, 0.001, 0.01, and 0.1). Model performance was rated using the area under the receiver operating characteristic (ROC) curve (AUC). Internal validation was performed using bootstrap resampling. The optimal cutoff probability was determined based on the Youden index. Sensitivity (SEN) and specificity (SPE) were calculated. Decision curve analysis (DCA) assessed clinical applicability, and calibration curves evaluated model calibration. The optimal model was selected based on the highest AUC achieved in the validation set. Furthermore, SHapley Additive exPlanations (SHAP) analysis was carried out to increase model interpretability by ranking the relative importance of the included variables. P <0.05 was considered statistically significant. The following R packages were used in this study: glmnet, pROC, mice, rmda, rms, xgboost, randomForest, e1071, CBCgrps, mRMRe, caret, kernelshap, fastshap, lightgbm, klaR, shapviz, corrplot, neuralnet, class, and missForest.
Results
Baseline Characteristics of Patients
Among the included variables, one observation was missing for each of the following variables: CRP, PCT, BNP, pH, PCO2, and PO2. Three observations were missing for the percentage of the predicted diffusing capacity of the lung for carbon monoxide measured by the single-breath method (DLCOSB). A total of 299 moderate-to-severe COPD patients were ultimately included, among whom 59 patients (20%) experienced unplanned readmission within three months. Of the 299 patients, 53 had moderate COPD, 125 had severe COPD, and 121 had extremely severe COPD. Finally, the random-number seed for the random split of the data was set to 3, and patients were separated into a training set (n=210) and a validation set (n=89). Their mean age was 70, and 74% were male (Table 1).
Table 1.
Distribution of Predictive Factors in the Training and Validation Sets
| Variables | Training Set (n=210) | Validation Set (n=89) | Total (n=299) | P |
|---|---|---|---|---|
| Age | 70 (65, 76) | 68 (63, 74) | 70 (64, 75) | 0.039 |
| Sex | 0.511 | |||
| Female | 52 (25) | 26 (29) | 78 (26) | |
| Male | 158 (75) | 63 (71) | 221 (74) | |
| BMI | 23.91 (21.49, 27) | 23.88 (20.81, 27.04) | 23.88 (21.46, 27.04) | 0.571 |
| Marital | 0.184 | |||
| Married | 204 (97) | 89 (100) | 293 (98) | |
| Others | 6 (3) | 0 (0) | 6 (2) | |
| Smoking | 0.866 | |||
| No | 46 (22) | 21 (24) | 67 (22) | |
| Yes | 164 (78) | 68 (76) | 232 (78) | |
| Course | 10 (5, 15) | 10 (5, 15) | 10 (5, 15) | 0.563 |
| GOLD | 0.897 | |||
| Moderate | 38 (18) | 15 (17) | 53 (18) | |
| Severe | 86 (41) | 39 (44) | 125 (42) | |
| Very Severe | 86 (41) | 35 (39) | 121 (40) | |
| Hospitalization | 0.086 | |||
| No | 130 (62) | 65 (73) | 195 (65) | |
| Yes | 80 (38) | 24 (27) | 104 (35) | |
| AECOPD | 0.584 | |||
| No | 33 (16) | 17 (19) | 50 (17) | |
| Yes | 177 (84) | 72 (81) | 249 (83) | |
| DM | 0.852 | |||
| No | 185 (88) | 77 (87) | 262 (88) | |
| Yes | 25 (12) | 12 (13) | 37 (12) | |
| Hypertension | 0.683 | |||
| No | 125 (60) | 50 (56) | 175 (59) | |
| Yes | 85 (40) | 39 (44) | 124 (41) | |
| CHD | 0.932 | |||
| No | 149 (71) | 62 (70) | 211 (71) | |
| Yes | 61 (29) | 27 (30) | 88 (29) | |
| HFailure | 0.329 | |||
| No | 80 (38) | 40 (45) | 120 (40) | |
| Yes | 130 (62) | 49 (55) | 179 (60) | |
| MI | 0.636 | |||
| No | 207 (99) | 87 (98) | 294 (98) | |
| Yes | 3 (1) | 2 (2) | 5 (2) | |
| Stroke | 0.170 | |||
| No | 205 (98) | 84 (94) | 289 (97) | |
| Yes | 5 (2) | 5 (6) | 10 (3) | |
| FEV1_P | 31.9 (26.71, 45.59) | 33.6 (26.1, 44.62) | 32.5 (26.38, 45.22) | 0.855 |
| FEV1 | 0.84 (0.64, 1.13) | 0.85 (0.7, 1.14) | 0.85 (0.65, 1.13) | 0.330 |
| FEV1_FVC | 42.79 (35.88, 52.37) | 42.44 (36.34, 53.23) | 42.48 (35.94, 52.62) | 0.890 |
| DLCOSB | 50.98 (39.24, 64.44) | 55.34 (40.3, 70) | 52.07 (39.35, 66.75) | 0.106 |
| WBC | 8 (6.46, 9.7) | 7.77 (5.85, 10.91) | 7.98 (6.3, 10.14) | 0.891 |
| Neutrophil | 5.94 (4.44, 7.83) | 6.07 (4.31, 8.68) | 6.02 (4.43, 8.17) | 0.552 |
| NEUT% | 78.05 (67.45, 83.7) | 74.8 (66.7, 82.9) | 76.4 (67.2, 83.7) | 0.572 |
| EOS | 0.06 (0.01, 0.15) | 0.04 (0.01, 0.13) | 0.05 (0.01, 0.14) | 0.562 |
| EOS% | 0.3 (0.1, 1.5) | 0.4 (0.1, 1.6) | 0.3 (0.1, 1.5) | 0.943 |
| RBC | 4.3 (3.97, 4.73) | 4.33 (4.04, 4.67) | 4.31 (4, 4.72) | 0.938 |
| Hb, Mean ± SD | 134.4 ± 16.65 | 132.39 ± 16.33 | 133.81 ± 16.56 | 0.334 |
| CRP | 7.8 (3.38, 23.77) | 7.8 (3.75, 18.3) | 7.8 (3.4, 21.7) | 0.854 |
| PCT | 0.05 (0.05, 0.05) | 0.05 (0.05, 0.05) | 0.05 (0.05, 0.05) | 0.355 |
| BNP | 38.31 (12.86, 100.25) | 34.07 (10.04, 89.57) | 37.44 (11.99, 99.51) | 0.507 |
| PH | 7.39 (7.36, 7.41) | 7.4 (7.37, 7.42) | 7.39 (7.36, 7.42) | 0.089 |
| PCO2 | 44.8 (39.55, 54.1) | 41.8 (38.1, 54.3) | 44.3 (38.9, 54.2) | 0.159 |
| PO2 | 83.04 (72.25, 99.5) | 81.8 (68.5, 111) | 83 (71.25, 101.5) | 0.857 |
| BL | 1.2 (1, 1.6) | 1.2 (1, 1.5) | 1.2 (1, 1.6) | 0.517 |
| Creatinine | 66.8 (56, 76.86) | 61 (55, 75) | 65.88 (55.31, 76.33) | 0.055 |
| ALT | 16.6 (12.53, 22.77) | 15.3 (12, 21.7) | 16.1 (12.15, 22.4) | 0.406 |
| AST | 16.15 (13.6, 20.98) | 15.8 (13, 20.6) | 16 (13.4, 20.85) | 0.417 |
| IMV | 1.000 | |||
| No | 208 (99) | 89 (100) | 297 (99) | |
| Yes | 2 (1) | 0 (0) | 2 (1) | |
| NIV | 0.583 | |||
| No | 188 (90) | 77 (87) | 265 (89) | |
| Yes | 22 (10) | 12 (13) | 34 (11) | |
| OT | 1.000 | |||
| Yes | 210 (100) | 89 (100) | 299 (100) | |
| Glucocorticoids | 0.611 | |||
| No | 17 (8) | 5 (6) | 22 (7) | |
| Yes | 193 (92) | 84 (94) | 277 (93) | |
| ICU | 0.327 | |||
| No | 205 (98) | 89 (100) | 294 (98) | |
| Yes | 5 (2) | 0 (0) | 5 (2) | |
| AE | 0.760 | |||
| No | 197 (94) | 85 (96) | 282 (94) | |
| Yes | 13 (6) | 4 (4) | 17 (6) | |
| Time | 11 (8, 14) | 11 (9, 14) | 11 (8, 14) | 0.449 |
| Group | 0.538 | |||
| No | 171 (81) | 69 (78) | 240 (80) | |
| Yes | 39 (19) | 20 (22) | 59 (20) |
Abbreviations: AE, adverse event; AECOPD, acute exacerbations of chronic obstructive pulmonary disease; ALT, alanine aminotransferase; AST, aspartate aminotransferase; BMI, body mass index; BNP, B-type natriuretic peptide; CHD, coronary heart disease; CRP, C-reactive protein; DLCOSB, the percentage of carbon monoxide diffusing capacity to the predicted value; DM, diabetes mellitus; Eos, eosinophil; Eos%, Eosinophil percentage; FEV1%pred, the percentage of FEV1 to the predicted value; FEV1, forced expiratory volume in one second; FEV1/FVC, forced expiratory volume in one second/forced vital capacity; GOLD, Global Initiative for Chronic Obstructive Lung Disease; HF, heart failure; ICU, intensive care unit; IMV, invasive mechanical ventilation; Lac, lactate; MI, myocardial infarction; NEUT%, neutrophilic granulocyte percentage; NIV, non-invasive ventilation; OT, oxygen therapy; PCO2, Partial Pressure of Carbon Dioxide; PCT, procalcitonin; PH, pH value; PO2, oxygen partial pressure; RBC, red blood cell; WBC, white blood cell.
Variable Selection Process
LASSO regression was applied to the training set to select predictive variables. Five-fold cross-validation was used to determine the final λ. The random seed was set to 3 before cross-validation. The optimal penalty parameter determined by LASSO regression was λ=0.035. Four variables were ultimately selected as predictors of three-month unplanned readmission in patients with moderate-to-severe COPD: hospitalization within the past year (VIF=1.048), FEV1 (VIF=1.042), history of CHD (VIF=1.009), and smoking history (VIF=1.012) (Figure 1).
Figure 1.
Variable selection process for predicting readmission in moderate-to-severe COPD using LASSO regression.
Notes: Before variable selection using LASSO regression, the glmnet function in R was used. Its default setting, standardize=TRUE, standardizes predictor variables before model fitting; the resulting coefficients are returned on the original scale of the variables. The categorical variables were coded as follows: sex (0=female, 1=male), marital status (0=married, 1=other), smoking history (0=no, 1=yes), Global Initiative for Chronic Obstructive Lung Disease (GOLD) grade (1=moderate, 2=severe, 3=very severe), hospitalization history (0=no, 1=yes), acute exacerbation of COPD (AECOPD; 0=no, 1=yes), diabetes mellitus (DM; 0=no, 1=yes), hypertension (0=no, 1=yes), coronary heart disease (CHD; 0=no, 1=yes), heart failure (0=no, 1=yes), myocardial infarction (MI; 0=no, 1=yes), stroke (0=no, 1=yes), invasive ventilation (0=no, 1=yes), noninvasive ventilation (0=no, 1=yes), oxygen therapy (0=no, 1=yes), glucocorticoid use (0=no, 1=yes), intensive care unit (ICU) admission (0=no, 1=yes), and acute exacerbation (AE; 0=no, 1=yes).
ML Model Construction
Based on the selected predictors, six ML models, including LR, XGBoost, SVM, LightGBM, BPNN, and NB, were developed. In the training cohort, the LR model achieved an AUC of 0.696 (95% CI: 0.605–0.787) and the calibrated AUC was 0.657 (95% CI: 0.577–0.740), with an optimal cutoff probability of 0.115, a SEN of 0.692, and a SPE of 0.661. The SVM model yielded an AUC of 0.785 (95% CI: 0.703–0.866) and the calibrated AUC was 0.768 (95% CI: 0.660–0.884), with a cutoff probability of 0.186, a SEN of 0.744, and a SPE of 0.690. The XGBoost model demonstrated the highest discriminative ability, with an AUC of 0.817 (95% CI: 0.748–0.886) and a corrected AUC of 0.803 (95% CI: 0.740–0.866), a cutoff probability of 0.155, a SEN of 0.795, and a SPE of 0.655. The LightGBM model achieved an AUC of 0.735 (95% CI: 0.645–0.825) and a corrected AUC of 0.706 (95% CI: 0.637–0.772), with a cutoff probability of 0.186, a SEN of 0.590, and a SPE of 0.795. The BPNN model yielded an AUC of 0.805 (95% CI: 0.727–0.883) and a corrected AUC of 0.700 (95% CI: 0.631–0.766), with a cutoff probability of 0.364, a SEN of 0.692, and a SPE of 0.854. The NB model achieved an AUC of 0.696 (95% CI: 0.603–0.789) and a corrected AUC of 0.650 (95% CI: 0.568–0.732), with a cutoff probability of 0.247, a SEN of 0.667, and a SPE of 0.713 (Table 2 and Figure 2). Calibration plots demonstrated that the LightGBM, LR, and BPNN models exhibited good agreement between predicted and observed probabilities (Figure 3). DCA demonstrated that all models exhibited a net benefit in the training set across threshold probabilities (Figure 4).
Table 2.
Discriminative Performance of the ML Models in the Training Sets
| Model | Cutoff | ROC (95% CI) | Corrected ROC (95% CI) | SEN | SPE | Positive Predictive Value | Negative Predictive Value | F1 Score | Accuracy | Calibration Slope | Calibration Intercept |
|---|---|---|---|---|---|---|---|---|---|---|---|
| LR | 0.115 | 0.696 (0.605–0.787) | 0.657 (0.577–0.740) | 0.692 | 0.661 | 0.318 | 0.904 | 0.436 | 0.667 | 0.964 | 0.022 |
| SVM | 0.186 | 0.785 (0.703–0.866) | 0.768 (0.660–0.884) | 0.744 | 0.690 | 0.354 | 0.922 | 0.480 | 0.7 | 0.826 | 0.118 |
| XGB | 0.155 | 0.817 (0.748–0.886) | 0.803 (0.740–0.866) | 0.795 | 0.655 | 0.344 | 0.933 | 0.480 | 0.681 | 1.062 | −0.025 |
| LightGBM | 0.186 | 0.735 (0.645–0.825) | 0.706 (0.637–0.772) | 0.590 | 0.795 | 0.397 | 0.895 | 0.475 | 0.757 | 0.791 | 0.082 |
| BPNN | 0.364 | 0.805 (0.727–0.883) | 0.700 (0.631–0.766) | 0.692 | 0.854 | 0.519 | 0.924 | 0.593 | 0.824 | 0.93 | 0.026 |
| NB | 0.247 | 0.696 (0.603–0.789) | 0.650 (0.568–0.732) | 0.667 | 0.713 | 0.347 | 0.904 | 0.457 | 0.705 | 0.944 | 0.039 |
Notes: Corrected ROC (95% CI) indicates the results of internal validation based on 1000 bootstrap resamples.
Figure 2.
ROC curves of the ML models for predicting readmission in the moderate-to-severe COPD population in the (A) training and (B) validation sets.
Figure 3.
Calibration curves of the ML models in the training set. (A) SVM; (B) XGBoost; (C) LR; (D) LightGBM; (E) BPNN; (F) NB.
Figure 4.
DCA of the ML models in the training set. (A) SVM; (B) XGBoost; (C) LR; (D) LightGBM; (E) BPNN; (F) NB.
Model Validation
In the validation cohort, the corresponding results were as follows. The LR model achieved an AUC of 0.730 (95% CI: 0.584–0.875), with a cutoff probability of 0.115, a SEN of 0.650, and a SPE of 0.812. The SVM model achieved an AUC of 0.604 (95% CI: 0.430–0.778), with a cutoff probability of 0.186, a SEN of 0.500, and a SPE of 0.580. The XGBoost model achieved an AUC of 0.809 (95% CI: 0.690–0.927), with a cutoff probability of 0.155, a SEN of 0.750, and a SPE of 0.826. The LightGBM model achieved an AUC of 0.726 (95% CI: 0.589–0.864), with a cutoff probability of 0.186, a SEN of 0.600, and a SPE of 0.768. The BPNN model achieved an AUC of 0.679 (95% CI: 0.508–0.850), with a cutoff probability of 0.364, a SEN of 0.400, and a SPE of 0.899. The NB model achieved an AUC of 0.760 (95% CI: 0.624–0.897), with a cutoff probability of 0.247, a SEN of 0.650, and a SPE of 0.855. The XGBoost model showed better predictive value (Table 3 and Figure 2). The calibration curves showed that the LightGBM, XGBoost, and BPNN models exhibited relatively good calibration (Figure 5). DCA further demonstrated that all models yielded a net clinical benefit across threshold probabilities in the validation set (Figure 6).
Table 3.
Discriminative Performance of the ML Models in the Validation Sets
| Model | Cutoff | ROC (95% CI) | SEN | SPE | Positive Predictive Value | Negative Predictive Value | F1 Score | Accuracy | Calibration Slope | Calibration Intercept |
|---|---|---|---|---|---|---|---|---|---|---|
| LR | 0.115 | 0.730 (0.584–0.875) | 0.650 | 0.812 | 0.500 | 0.889 | 0.565 | 0.775 | 1.206 | −0.156 |
| SVM | 0.186 | 0.604 (0.430–0.778) | 0.500 | 0.580 | 0.256 | 0.800 | 0.339 | 0.562 | 2.534 | −0.290 |
| XGB | 0.155 | 0.809 (0.690–0.927) | 0.750 | 0.826 | 0.556 | 0.919 | 0.639 | 0.809 | 0.915 | 0.057 |
| LightGBM | 0.186 | 0.726 (0.589–0.864) | 0.600 | 0.768 | 0.429 | 0.869 | 0.500 | 0.730 | 0.987 | 0.026 |
| BPNN | 0.364 | 0.679 (0.508–0.850) | 0.400 | 0.899 | 0.533 | 0.838 | 0.457 | 0.787 | 0.686 | 0.175 |
| NB | 0.247 | 0.760 (0.624–0.897) | 0.650 | 0.855 | 0.565 | 0.894 | 0.605 | 0.809 | 1.274 | −0.172 |
Figure 5.
Calibration curves of the ML models in the validation set. (A) SVM; (B) XGBoost; (C) LR; (D) LightGBM; (E) BPNN; (F) NB.
Figure 6.
DCA of the ML models in the validation set. (A) SVM; (B) XGBoost; (C) LR; (D) LightGBM; (E) BPNN; (F) NB.
SHAP Analysis of the Optimal Model
SHAP analysis based on XGBoost, the optimal model, ranked variable importance as follows: hospitalization history within the past year, FEV1, history of CHD, and smoking history (Figures 7 and 8).
Figure 7.
SHAP beeswarm plot of the XGBoost model.
Figure 8.
SHAP feature importance plot of the XGBoost model.
Discussion
This study developed an ML model to predict unplanned readmission within three months among the moderate-to-severe COPD population. Variables were initially screened in the training set using LASSO regression, ultimately identifying four independent predictors: hospitalization over the past year, FEV1, history of CHD, and smoking history, which were incorporated into the model. LR, XGBoost, SVM, LightGBM, BPNN, and NB models were constructed using these four variables in both sets, among which XGBoost displayed superior performance, with the AUC being 0.809 (95% CI: 0.690–0.927) in the validation set. The best threshold was 0.155, yielding an SEN of 0.750 and an SPE of 0.826. The validation performance of other models, including SVM and BPNN, appeared unstable, with a marked decline in performance from the training set to the validation set. In addition, XGBoost had a relatively wide confidence interval, warranting cautious interpretation of its predictive performance.
The four predictors identified by LASSO regression, including hospitalization within the previous year, FEV1, history of CHD, and smoking history, are all supported by well-established pathophysiological mechanisms and clinical evidence. Previous hospitalization is the strongest predictor of COPD readmission.33,34 Recurrent hospitalizations due to acute exacerbations perpetuate airway inflammation, accelerate progressive lung function decline, and are frequently accompanied by additional high-risk conditions such as pulmonary hypertension and cardiac dysfunction, thereby creating a vicious cycle of exacerbation, hospitalization, and subsequent exacerbation. FEV1 remains the gold standard for assessing the severity of airflow limitation in COPD. Lower FEV1 is associated with poorer exercise tolerance and impaired airway defense, rendering patients more susceptible to acute exacerbations even in response to minor triggers, and has been consistently associated with an increased risk of COPD readmission.4 Beyond reflecting the severity of airflow obstruction, reduced FEV1 may also increase the likelihood of readmission indirectly through its effects on systemic inflammation, exercise capacity, and the development of comorbidities.35,36 A meta-analysis conducted in Asian populations reported that patients with FEV1 <50% and at least two acute exacerbations during the previous year had a 90-day readmission rate as high as 31%, further supporting the predictive value of FEV1 for short-term readmission.4 Multiple studies have consistently demonstrated that patients with COPD complicated by CHD have substantially higher risks of readmission and mortality following hospitalization for acute exacerbations.37–39 COPD and CHD share complex pathophysiological interactions,38 including systemic inflammation, endothelial dysfunction, increased platelet reactivity, atherosclerosis, and hypoxemia, all of which contribute to the development and progression of both diseases. These mechanisms not only aggravate pathological injury in each condition but also allow the systemic inflammatory response accompanying acute COPD exacerbations to trigger rupture of unstable atherosclerotic plaques, potentially resulting in acute MI. This vicious cycle contributes to greater disease instability and substantially increases the risks of readmission and death among patients with concomitant CHD. Therefore, clinical management should extend beyond preventing COPD readmission alone and should emphasize the early identification and management of cardiovascular comorbidities such as CHD. Long-term smoking is a modifiable risk factor for COPD development and persistent airway inflammation. Harmful constituents of tobacco smoke impair mucociliary clearance, induce airway epithelial apoptosis and remodeling, progressively worsen ventilatory dysfunction, and significantly increase the risks of acute exacerbation and short-term readmission.40
With the rapid development of artificial intelligence, various ML algorithms have been increasingly applied to predict COPD readmission risk. Although several studies involving advanced ML methods, federated learning, and intelligent clinical decision support systems have reported relatively high AUC values,28,29 these approaches generally rely on numerous laboratory or multidimensional clinical variables, limiting their applicability in routine clinical practice. Most existing prediction models for COPD readmission have been developed in mixed populations including patients with mild, moderate, and severe disease, with few studies specifically optimizing prediction for patients with moderate-to-severe COPD. Previous studies have shown that the risk of unplanned readmission increases with COPD severity, and more than 85% of patients hospitalized for acute exacerbations belong to the GOLD grade 3 or 4 moderate-to-severe subgroup.30 These patients bear the greatest burden of early recurrence and repeated hospitalization, yet dedicated tools for predicting short-term readmission risk in this high-risk population remain lacking. Accordingly, the present study specifically focused on patients with moderate-to-severe COPD. This study selected four clinically relevant predictors associated with COPD severity and adverse outcomes to develop a parsimonious prediction model. As an exploratory study with internal validation, it provides insights to inform the further development and refinement of risk assessment tools.
Limitations of the Study
Although ML was employed to develop a predictive tool for three-month unplanned readmission among moderate-to-severe COPD patients, some limitations exist. The sample size was limited, and the number of patients experiencing unplanned readmission within three months was relatively small. In addition, the single-center design limited the generalizability of the findings, and external validation has not yet been performed. Therefore, the results should be interpreted cautiously. Future prospective multicenter studies involving populations from diverse geographic regions are needed to externally validate and update the model.
Conclusion
In summary, hospitalization over the past year, FEV1, history of CHD, and smoking history are predictors for three-month unplanned readmission among moderate-to-severe COPD patients. A systematic comparison of six ML models developed using routinely available, interpretable clinical features identified XGBoost as the model with the most promising discriminatory performance. The model was tailored to the target population and incorporated a concise set of readily available predictors. However, given the small sample size, single-center design, and class imbalance, the findings of this exploratory prediction study based on internal validation should be interpreted cautiously. Future studies involving multicenter cohorts from diverse geographic regions are needed to externally validate and update the model. If further validated, the model may help clinicians identify patients at high risk of readmission, prioritize intensive follow-up, and develop tailored pulmonary rehabilitation programs to reduce the risk of readmission within three months.
Funding Statement
This study was supported by Tianjin Key Medical Discipline Construction Project, China (Grant No.: TJYXZDXK-3-032C).
Data Sharing Statement
The generated and analyzed datasets are available from the corresponding author upon reasonable request.
Ethics Approval and Informed Consent
This study was approved by the Ethics Review Committee of Chest Hospital, Tianjin University [Approval No.: 2026LW-051] and conducted in compliance with the principles of the Declaration of Helsinki. The committee waived the requirement for individual informed consent due to the retrospective study design. All personal identity information was fully protected, and all data were anonymized.
Author Contributions
All authors have made substantial contributions to the reported work, whether in the conception, study design, execution, data acquisition, analysis or interpretation, or all of these areas, have participated in drafting, revising or critically reviewing the article, have given final approval of the version to be published, have agreed on the journal to which the article is being submitted and have been responsible for all aspects of the work.
Disclosure
The authors declare that they have no competing interests.
References
- 1.Boers E, Barrett M, Su JG, et al. Global burden of chronic obstructive pulmonary disease through 2050. JAMA Netw Open. 2023;6:e2346598. doi: 10.1001/jamanetworkopen.2023.46598 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Boers E, Allen A, Barrett M, et al. Forecasting the global economic and health burden of COPD from 2025 through 2050. Chest. 2025;168:880–18. doi: 10.1016/j.chest.2025.03.029 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Global Initiative for Chronic Obstructive Lung Disease. 2025. Available from: https://goldcopd.org/2025-gold-report. Accessed October 19, 2026.
- 4.Lin P, Shen C, Li Q, et al. A systematic review and meta-analysis of chronic obstructive pulmonary disease in Asia: risk factors for readmission and readmission rate. BMC Pulm Med. 2024;24:388. doi: 10.1186/s12890-024-03203-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Cadham CJ, Oh H, Han MK, et al. The prevalence and mortality risks of PRISm and COPD in the United States from NHANES 2007-2012. Respir Res. 2024;25:208. doi: 10.1186/s12931-024-02841-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Liu SF, Lin KC, Chin CH, et al. Factors influencing short-term re-admission and one-year mortality in patients with chronic obstructive pulmonary disease. Respirology. 2007;12:560–565. doi: 10.1111/j.1440-1843.2007.01110.x [DOI] [PubMed] [Google Scholar]
- 7.Yalcinsoy M, Salturk C, Oztas S, et al. Can patients with moderate to severe acute respiratory failure from COPD be treated safely with noninvasive mechanical ventilation on the ward? Int J Chron Obstruct Pulmon Dis. 2016;11:1151–1160. doi: 10.2147/COPD.S104801 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Hoogendoorn M, Feenstra TL, Hoogenveen RT, et al. Association between lung function and exacerbation frequency in patients with COPD. Int J Chron Obstruct Pulmon Dis. 2010;5:435–444. doi: 10.2147/COPD.S13826 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Crisafulli E, Sartori G, Huerta A, et al. Association between rome classification among hospitalized patients with COPD exacerbations and short-term and intermediate-term outcomes. Chest. 2023;164:1422–1433. doi: 10.1016/j.chest.2023.07.021 [DOI] [PubMed] [Google Scholar]
- 10.Watz H, Tetzlaff K, Magnussen H, et al. Spirometric changes during exacerbations of COPD: a post hoc analysis of the WISDOM trial. Respir Res. 2018;19:251. doi: 10.1186/s12931-018-0944-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Cao YQ, Dong LX, Cao J. Pulmonary embolism in patients with acute exacerbation of chronic obstructive pulmonary disease. Chin Med J. 2018;131:1732–1737. doi: 10.4103/0366-6999.235865 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Donaldson GC, Hurst JR, Smith CJ, et al. Increased risk of myocardial infarction and stroke following exacerbation of COPD. Chest. 2010;137:1091–1097. doi: 10.1378/chest.09-2029 [DOI] [PubMed] [Google Scholar]
- 13.Roberts CM, Lowe D, Bucknall CE, et al. Clinical audit indicators of outcome following admission to hospital with acute exacerbation of chronic obstructive pulmonary disease. Thorax. 2002;57:137–141. doi: 10.1136/thorax.57.2.137 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Halpern MT, Stanford RH, Borker R. The burden of COPD in the U.S.A.: results from the Confronting COPD survey. Respir Med. 2003;97 Suppl C:S81–89. doi: 10.1016/S0954-6111(03)80028-8 [DOI] [PubMed] [Google Scholar]
- 15.Sharpe I, Bowman M, Kim A, et al. Strategies to Prevent Readmissions to Hospital for COPD: a Systematic Review. COPD. 2021;18:456–468. doi: 10.1080/15412555.2021.1955338 [DOI] [PubMed] [Google Scholar]
- 16.Silver PC, Kollef MH, Clinkscale D, et al. A respiratory therapist disease management program for subjects hospitalized with COPD. Respir Care. 2017;62:1–9. doi: 10.4187/respcare.05030 [DOI] [PubMed] [Google Scholar]
- 17.Bonomo M, Hermsen MG, Kaskovich S, et al. Using machine learning to predict likelihood and cause of readmission after hospitalization for chronic obstructive pulmonary disease exacerbation. Int J Chron Obstruct Pulmon Dis. 2022;17:2701–2709. doi: 10.2147/COPD.S379700 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Şahin F, Koşar AF, Aslan AF, et al. Serum biomarkers in patients with stable and acute exacerbation of chronic obstructive pulmonary disease: a comparative study. J Med Biochem. 2019;38:503–511. doi: 10.2478/jomb-2018-0050 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Niraula P, Upreti M, Kadariya S, et al. AI/ML driven prediction of COPD exacerbations and readmissions: a systematic review and meta-analysis. Front Digit Health. 2025;7:1641356. doi: 10.3389/fdgth.2025.1641356 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Puhan MA, Garcia-Aymerich J, Frey M, et al. Expansion of the prognostic assessment of patients with chronic obstructive pulmonary disease: the updated BODE index and the ADO index. Lancet. 2009;374:704–711. doi: 10.1016/S0140-6736(09)61301-5 [DOI] [PubMed] [Google Scholar]
- 21.Soler-Cataluña JJ, Martínez-García MA, Sánchez LS, et al. Severe exacerbations and BODE index: two independent risk factors for death in male COPD patients. Respir Med. 2009;103:692–699. doi: 10.1016/j.rmed.2008.12.005 [DOI] [PubMed] [Google Scholar]
- 22.Almagro P, Soriano JB, Cabrera FJ, et al. Short- and medium-term prognosis in patients hospitalized for COPD exacerbation: the CODEX index. Chest. 2014;145:972–980. doi: 10.1378/chest.13-1328 [DOI] [PubMed] [Google Scholar]
- 23.Jones RC, Donaldson GC, Chavannes NH, et al. Derivation and validation of a composite index of severity in chronic obstructive pulmonary disease: the DOSE Index. Am J Respir Crit Care Med. 2009;180:1189–1195. doi: 10.1164/rccm.200902-0271OC [DOI] [PubMed] [Google Scholar]
- 24.Echevarria C, Steer J, Heslop-Marshall K, et al. The PEARL score predicts 90-day readmission or death after hospitalisation for acute exacerbation of COPD. Thorax. 2017;72:686–693. doi: 10.1136/thoraxjnl-2016-209298 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Corlateanu A, Plahotniuc A, Corlateanu O, et al. Multidimensional indices in the assessment of chronic obstructive pulmonary disease. Respir Med. 2021;185:106519. doi: 10.1016/j.rmed.2021.106519 [DOI] [PubMed] [Google Scholar]
- 26.Goto T, Jo T, Matsui H, et al. Machine learning-based prediction models for 30-day readmission after hospitalization for chronic obstructive pulmonary disease. COPD. 2019;16:338–343. doi: 10.1080/15412555.2019.1688278 [DOI] [PubMed] [Google Scholar]
- 27.Lopez K, Li H, Lipkin-Moore Z, et al. Deep learning prediction of hospital readmissions for asthma and COPD. Respir Res. 2023;24:311. doi: 10.1186/s12931-023-02628-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Alvarez-Romero C, Martinez-Garcia A, Ternero Vega J, et al. Predicting 30-day readmission risk for patients with chronic obstructive pulmonary disease through a federated machine learning architecture on Findable, Accessible, Interoperable, and Reusable (FAIR) data: development and validation study. JMIR Med Inform. 2022;10:e35307. doi: 10.2196/35307 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.López-Canay J, Casal-Guisande M, Pinheira A, et al. Predicting COPD readmission: an intelligent clinical decision support system. Diagnostics. 2025;15:318–340. doi: 10.3390/diagnostics15030318 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Wong AW, Gan WQ, Burns J, et al. Acute exacerbation of chronic obstructive pulmonary disease: influence of social factors in determining length of hospital stay and readmission rates. Can Respir J. 2008;15:361–364. doi: 10.1155/2008/569496 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Peng J, Yu Q, Fan S, et al. High blood eosinophil and YKL-40 levels, as well as low CXCL9 levels, are associated with increased readmission in patients with acute exacerbation of chronic obstructive pulmonary disease. Int J Chron Obstruct Pulmon Dis. 2021;16:795–806. doi: 10.2147/COPD.S294968 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.van Walraven C, Dhalla IA, Bell C, et al. Derivation and validation of an index to predict early death or unplanned readmission after discharge from hospital to the community. CMAJ. 2010;182:551–557. doi: 10.1503/cmaj.091117 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Fernández-García S, Represas-Represas C, Ruano-Raviña A, et al. Social and clinical predictors of short- and long-term readmission after a severe exacerbation of copd. PLoS One. 2020;15:e0229257. doi: 10.1371/journal.pone.0229257 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Njoku CM, Alqahtani JS, Wimmer BC, et al. Risk factors and associated outcomes of hospital readmission in COPD: a systematic review. Respir Med. 2020;173:105988. doi: 10.1016/j.rmed.2020.105988 [DOI] [PubMed] [Google Scholar]
- 35.Wu YK, Su WL, Yang MC, et al. Characterization associated with the frequent severe exacerbator phenotype in COPD patients. Int J Chron Obstruct Pulmon Dis. 2021;16:2475–2485. doi: 10.2147/COPD.S317177 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Roberts CM, Stone RA, Lowe D, et al. Co-morbidities and 90-day outcomes in hospitalized COPD exacerbations. COPD. 2011;8:354–361. doi: 10.3109/15412555.2011.600362 [DOI] [PubMed] [Google Scholar]
- 37.Duffy S, Barnett S, Civic B, et al. Risk of death by comorbidity prompting rehospitalization following the initial COPD hospitalization. Chronic Obstr Pulm Dis. 2015;2:17–22. doi: 10.15326/jcopdf.2.1.2014.0129 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Campo G, Pavasini R, Malagù M, et al. Chronic obstructive pulmonary disease and ischemic heart disease comorbidity: overview of mechanisms and clinical management. Cardiovasc Drugs Ther. 2015;29:147–157. doi: 10.1007/s10557-014-6569-y [DOI] [PubMed] [Google Scholar]
- 39.Nantsupawat T, Limsuwat C, Nugent K. Factors affecting chronic obstructive pulmonary disease early rehospitalization. Chron Respir Dis. 2012;9:93–98. doi: 10.1177/1479972312438703 [DOI] [PubMed] [Google Scholar]
- 40.De Cunto G, Cavarra E, Bartalesi B, et al. Innate immunity and cell surface receptors in the pathogenesis of COPD: insights from mouse smoking models. Int J Chron Obstruct Pulmon Dis. 2020;15:1143–1154. doi: 10.2147/COPD.S246219 [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.
Data Availability Statement
The generated and analyzed datasets are available from the corresponding author upon reasonable request.