Development and Internal Validation of a Multivariable Prognostic Model for in-Hospital Mortality in Critically Ill Patients with Acute Exacerbation of COPD

Abstract

Purpose

Acute exacerbations of chronic obstructive pulmonary disease (COPD) requiring intensive care unit (ICU) admission are associated with substantial mortality. Although the Acute Physiology and Chronic Health Evaluation II (APACHE II) score is widely used for prognostic assessment, its complexity and limited disease-specific applicability have prompted the development of simpler prognostic models. Therefore, this study aimed to develop and internally validate a multivariable prognostic model for predicting in-hospital mortality in critically ill patients with acute exacerbation of COPD and to compare its performance with APACHE II.

Patients and Methods

This retrospective observational study included 392 adult patients admitted to the ICU with acute exacerbation of COPD between January 2022 and December 2025. Demographic characteristics, laboratory findings, arterial blood gas parameters, Charlson Comorbidity Index, and APACHE II scores were recorded. Independent predictors of mortality were identified using multivariable logistic regression. Model discriminatory performance was assessed using receiver operating characteristic (ROC) curve analysis, while calibration was evaluated using the Hosmer–Lemeshow test and calibration plots. Internal validation was performed using 1000 bootstrap resamples.

Results

A total of 392 patients were included, of whom 94 (23.9%) died during hospitalization. Multivariable logistic regression identified age, male sex, urea, albumin, C-reactive protein (CRP), and Charlson Comorbidity Index as independent predictors of in-hospital mortality. The final prediction model demonstrated excellent discriminatory performance (AUC 0.899, 95% CI 0.865–0.927) and significantly outperformed APACHE II (AUC 0.813, 95% CI 0.771–0.851; DeLong p = 0.0014). Calibration was good according to the Hosmer–Lemeshow test (p = 0.066), and internal validation using 1,000 bootstrap resamples confirmed model stability.

Conclusion

A multivariable prognostic model based on six routinely available admission variables demonstrated excellent discriminatory performance and good calibration for predicting in-hospital mortality in critically ill patients with acute exacerbation of COPD. The model showed higher discriminatory performance than APACHE II in our cohort. Nevertheless, external validation and direct comparison with established COPD-specific prognostic models are required before routine clinical implementation.

Introduction

Chronic obstructive pulmonary disease (COPD) remains one of the leading causes of morbidity and mortality worldwide.¹ Acute exacerbations of COPD account for a substantial proportion of hospitalizations and intensive care unit (ICU) admissions.² COPD exacerbations requiring intensive care are associated with high mortality because of hypercapnic respiratory failure, infectious complications, cardiovascular events, and multiple organ dysfunction.³ Therefore, early identification of high-risk patients is of critical importance for clinical decision-making.

Scoring systems such as the Acute Physiology and Chronic Health Evaluation II (APACHE II) are widely used for mortality risk assessment in critically ill patients.⁴ However, these scoring systems have several limitations in routine clinical practice. Their calculation requires clinical data collected during the first 24 hours after ICU admission, may not adequately reflect early clinical deterioration, and do not sufficiently account for COPD-specific prognostic characteristics.^5,6 Consequently, there is growing interest in the development of simpler prognostic models based on easily obtainable clinical parameters.

Several COPD-specific prognostic tools, including the DECAF and BAP-65 scores, have been developed to predict short-term outcomes in patients with acute exacerbation of COPD. These scores have demonstrated acceptable prognostic performance, particularly among hospitalized patients and those presenting to emergency departments.^6,7 However, they were not specifically developed for critically ill ICU populations and rely on different combinations of clinical variables. Furthermore, their comparative performance in patients requiring intensive care remains limited. Therefore, there remains a need for practical prognostic models specifically developed for critically ill patients using routinely available clinical and laboratory variables obtained at ICU admission.

Previous studies have demonstrated that advanced age, comorbidity burden, elevated inflammatory markers, renal dysfunction, and hypoalbuminemia are associated with poor prognosis during COPD exacerbations.^8,9 In particular, variables such as urea, C-reactive protein (CRP), albumin levels, and the Charlson Comorbidity Index have been reported to be associated with mortality.^10,11 However, studies evaluating these parameters together within a multivariable prognostic model and directly comparing them with the APACHE II score remain limited.

Based on previous evidence, we hypothesized that a multivariable prognostic model based on routinely available demographic, clinical, and laboratory variables obtained at ICU admission could accurately predict in-hospital mortality in critically ill patients with acute exacerbation of COPD and provide additional prognostic value to existing severity scores. Therefore, the aim of this study was to identify clinical and laboratory parameters associated with in-hospital mortality, develop and internally validate a multivariable prognostic model, and compare its discriminatory performance with the APACHE II score.

Materials and Methods

Study Design and Patient Population

This retrospective observational study was conducted in the Intensive Care Unit of Aksaray Training and Research Hospital. Adult patients admitted to the ICU with acute exacerbation of COPD (AECOPD) between January 2022 and December 2025 were eligible for inclusion.

COPD had been diagnosed before ICU admission by pulmonologists based on compatible clinical history, previous spirometry when available, and documented medical records. AECOPD was defined according to the 2025 Global Initiative for Chronic Obstructive Lung Disease (GOLD) report.¹²

Patients aged ≥18 years were eligible for inclusion. Patients with missing clinical data, active malignancy, advanced chronic liver failure, or recurrent ICU admissions were excluded. In patients with multiple ICU admissions, only the first admission was included to ensure statistical independence. Patients with advanced chronic liver failure were excluded because severe hepatic dysfunction may independently influence albumin synthesis, inflammatory biomarkers, and mortality risk.

Data Collection

Demographic characteristics, comorbidities, laboratory findings, arterial blood gas measurements, and ICU severity scores were retrospectively extracted from the hospital electronic medical record system.

Variables recorded at ICU admission included age, sex, laboratory findings, and arterial blood gas measurements. The Charlson Comorbidity Index (CCI) and APACHE II score were calculated using standard definitions. Laboratory parameters included serum urea, creatinine, C-reactive protein (CRP), albumin, and procalcitonin (PCT). Arterial blood gas parameters included pH, PaCO₂, HCO₃⁻, PaO₂, and SaO₂.

Serum PCT measurements were not routinely available for all patients because they were obtained at the discretion of the treating physician according to clinical indications. Therefore, PCT was analyzed descriptively and in univariate analyses only and was not considered for multivariable model development.

All laboratory and arterial blood gas measurements were based on the first samples obtained within 24 hours of ICU admission.

Patient confidentiality was maintained by anonymizing all collected data before analysis, and no personally identifiable information was accessed or reported.

Outcome Measure

The primary outcome of the study was in-hospital mortality. Patients were categorized into survivor and non-survivor groups based on survival status.

Statistical Analysis

Statistical analyses were performed using SPSS Statistics for Windows, Version 27 (IBM Corp., Armonk, NY, USA) and MedCalc Version 23.5.2.

The distribution of continuous variables was assessed using the Kolmogorov–Smirnov test. Non-normally distributed variables were presented as median (interquartile range) and compared using the Mann–Whitney U-test. Categorical variables were expressed as numbers and percentages and compared using the chi-square test.

Logistic regression analysis was performed to identify independent predictors of in-hospital mortality. Candidate predictors were selected based on their clinical relevance and statistical significance in the univariate analyses. Variables with a p value <0.05 in the univariate analysis were considered for multivariable modelling. Variables with a substantial proportion of missing data, such as procalcitonin, were not considered for multivariable model development because their inclusion would have markedly reduced the effective sample size and increased the risk of selection bias. When two or more variables reflected similar biological processes, the variable considered to have greater clinical relevance and contribution to model performance was retained to improve model parsimony and reduce redundancy. All selected variables were entered simultaneously into the multivariable logistic regression model using the Enter method. Multicollinearity was assessed using variance inflation factor (VIF) and tolerance statistics before model construction, and no evidence of problematic multicollinearity was identified.

The discriminative performance of the model was evaluated using receiver operating characteristic (ROC) curve analysis, and the area under the curve (AUC) was calculated.

Internal validation of the final prediction model was performed using 1,000 bootstrap resamples to assess coefficient stability. Model calibration was evaluated using the Hosmer–Lemeshow goodness-of-fit test.

The patient selection process is shown in Figure 1.

Figure 1.

Study Flowchart.

The study protocol was approved by the Aksaray University Scientific Research Ethics Committee for Health Sciences (Ethics Committee Approval No. 2025/184; September 25, 2025). Due to the retrospective observational design of the study, informed consent was waived.

Results

The baseline demographic, clinical, and laboratory characteristics of the 392 included patients are summarized in Table 1. The median age was 71 (64–82) years, and the overall in-hospital mortality rate was 23.9% (94/392). Non-survivors were significantly older than survivors (79 vs 69 years, p < 0.001) and were more likely to be male (80.9% vs 65.1%, p = 0.004).

Table 1.

Comparison of Demographic, Clinical, and Laboratory Characteristics According to Survival Status

Variables	Total (n=392)	Survivors (n=298)	Non-Survivors (n=94)	p-value
Age (years)	71 (64–82)	69 (62–79)	79 (72–86)	<0.001
Male sex, n (%)	270 (68.9%)	194 (65.1%)	76 (80.9%)	0.004
pH	7.32 (7.26–7.39)	7.33 (7.27–7.39)	7.30 (7.25–7.40)	0.023
PaCO2 (mmHg)	75 (60–93)	78.5 (64–93)	73 (58–91)	0.144
HCO3− (mmol/L)	32 (27–36)	32 (28–35.9)	31.5 (24.5–37.1)	0.428
PaO2 (mmHg)	60 (53–67)	63 (60–67)	44.5 (43–54.2)	<0.001
SaO2 (%)	89 (81–91)	90 (90–91)	81.5 (80.7–90)	<0.001
Urea (mg/dL)	51 (34–84)	46 (31–62)	79 (59–130)	<0.001
Creatinine (mg/dL)	0.88 (0.68–1.28)	0.85 (0.67–1.05)	0.95 (0.73–1.89)	<0.001
Albumin (g/L)	33 (29–38)	34 (31–38)	30 (27–36)	<0.001
CRP (mg/L)	54 (22.5–93)	50 (20.9–80.5)	68.5 (32–131)	<0.001
PCT (ng/mL)	0.12 (0.06–1.02)	0.09 (0.05–0.20)	2.25 (1.00–3.20)	<0.001
CCI score	2 (1–2)	2 (1–2)	2 (2–2)	<0.001
APACHE II	15 (12–21)	14 (11–19)	23 (18–28)	<0.001
ICU stay (days)	4 (2–9)	3.5 (2–6)	8 (4–16)	<0.001

Abbreviations: CRP, C-reactive protein; PCT, procalcitonin; CCI, Charlson Comorbidity Index; APACHE II, Acute Physiology and Chronic Health Evaluation II; ICU, intensive care unit; PaCO₂, partial pressure of arterial carbon dioxide; HCO₃⁻, bicarbonate; PaO₂, partial pressure of arterial oxygen; SaO₂, arterial oxygen saturation. Data are presented as median (interquartile range) or n (%), as appropriate.

Arterial blood gas analysis showed that median pH was significantly lower in non-survivors (p = 0.023). In addition, PaO₂ (44.5 vs 63 mmHg, p < 0.001) and SaO₂ (81.5% vs 90%, p < 0.001) were significantly lower in non-survivors, whereas PaCO₂ and HCO₃⁻ did not differ significantly between the groups.

Biochemical analysis revealed significantly higher levels of urea, creatinine, CRP, and procalcitonin (PCT) [2.25 (1.00–3.20) vs 0.09 (0.05–0.20) ng/mL, p < 0.001] in the mortality group, whereas albumin levels were significantly lower (all p < 0.001). Both the Charlson Comorbidity Index (CCI), reflecting chronic disease burden, and the APACHE II score, reflecting acute disease severity, were significantly higher in non-survivors than in survivors (both p < 0.001).

The results of the logistic regression analyses are presented in Table 2. Multivariable analysis demonstrated that age, male sex, urea, CRP, albumin, and Charlson Comorbidity Index were independently associated with mortality. Although PCT levels were significantly higher in non-survivors in the group comparison analysis, PCT was not associated with in-hospital mortality in univariate logistic regression analysis. Furthermore, because of the substantial proportion of missing data, PCT was not considered for multivariable model development.

Table 2.

Logistic Regression Analysis of Independent Predictors of Mortality

Variables	Univariate Analysis OR (95% CI)	p-value	Multivariable Analysis OR (95% CI)	p-value
Age	1.066 (1.042–1.091)	<0.001	1.073 (1.039–1.108)	<0.001
Male sex	2.263 (1.285–3.988)	0.005	4.797 (2.232–10.308)	<0.001
pH	0.651 (0.030–13.972)	0.784	–	–
PaO2	0.958 (0.916–1.002)	0.062	–	–
SaO2	0.673 (0.484–0.937)	0.019	–	–
Urea	1.030 (1.022–1.038)	<0.001	1.028 (1.018–1.037)	<0.001
Creatinine	3.819 (2.352–6.199)	<0.001	–	–
Albumin	0.869 (0.825–0.916)	<0.001	0.884 (0.826–0.947)	<0.001
CRP	1.007 (1.003–1.011)	0.001	1.008 (1.002–1.014)	0.006
Charlson Comorbidity Index	2.662 (1.886–3.757)	<0.001	3.386 (2.091–5.481)	<0.001
PCT	1.023 (0.961–1.090)	0.474	–	–

Abbreviations: OR, odds ratio; CI, confidence interval; CRP, C-reactive protein; PaO₂, partial pressure of arterial oxygen; SaO₂, arterial oxygen saturation; PCT, Procalcitonin. PCT was analyzed in the subset of patients with available measurements and was not considered for multivariable model development because of the substantial proportion of missing data. Odds ratios for continuous variables were calculated per one-unit increase. Variables not retained in the final multivariable model were evaluated in the univariate analysis and excluded according to the predefined variable selection strategy to achieve a parsimonious model and avoid redundancy among clinically related predictors.

Among the independent predictors in the final model, male sex (OR: 4.797, 95% CI: 2.232–10.308) and Charlson Comorbidity Index (OR: 3.386, 95% CI: 2.091–5.481) showed the strongest associations with in-hospital mortality. Each additional year of age was associated with a 7.3% increase in the odds of in-hospital mortality (OR: 1.073, 95% CI: 1.039–1.108), whereas each 1 g/L increase in albumin level was associated with an 11.6% decrease in the odds of in-hospital mortality (OR: 0.884, 95% CI: 0.826–0.947). The Hosmer–Lemeshow goodness-of-fit test demonstrated good model calibration (p = 0.066).

Final Prediction Equation

The final multivariable logistic regression equation was:

Logit(p) =−8.228 + (0.070 × Age) + (1.568 × Male sex) + (0.027 × Urea) − (0.123 × Albumin) + (0.008 × CRP) + (1.220 × Charlson Comorbidity Index), where male sex was coded as 1 for males and 0 for females. The predicted probability of in-hospital mortality was calculated as .

The discriminative performance of the proposed model and the APACHE II score is summarized in Table 3. ROC curve analysis showed that the final model consisting of age, sex, urea, albumin, CRP, and Charlson Comorbidity Index showed good discriminative performance for predicting mortality (AUC: 0.899; 95% CI: 0.865–0.927; p < 0.001). In contrast, the discriminative performance of the APACHE II score was lower (AUC: 0.813; 95% CI: 0.771–0.851; p < 0.001). Comparison of the two models using the DeLong test showed that the AUC of the combined model was significantly higher than that of the APACHE II score (ΔAUC: 0.086; 95% CI: 0.033–0.139; p = 0.0014), as illustrated in Figure 2.

Table 3.

Comparison of the Discriminative Performance of the Proposed in-Hospital Mortality Prediction Model and the APACHE II Score

Model	AUC	95% CI	p-value
Proposed model	0.899	0.865–0.927	<0.001
APACHE II	0.813	0.771–0.851	<0.001
DeLong difference	0.086	0.033–0.139	0.0014

Abbreviation: APACHE II, Acute Physiology and Chronic Health Evaluation II.

Figure 2.

Comparison of the receiver operating characteristic (ROC) curves of the developed multivariable logistic regression model and the APACHE II score for prediction of in-hospital mortality. The developed model showed superior discriminatory performance with an AUC of 0.899, compared with 0.813 for the APACHE II score. The difference in AUCs was statistically significant based on DeLong’s test (p = 0.0014).

Internal validation of the model was performed using 1000 bootstrap resamples. Bootstrap analysis confirmed that all variables included in the model maintained their significance and that model coefficients remained stable across repeated samples. In addition, the low bias values obtained after bootstrap resampling supported the absence of substantial overfitting.

Calibration analysis demonstrated good agreement between the predicted and observed probabilities of in-hospital mortality (Figure 3), consistent with the non-significant Hosmer–Lemeshow goodness-of-fit test (p = 0.066).

Figure 3.

Calibration plot of the final multivariable prognostic model. Patients were categorized into deciles according to their predicted risk of in-hospital mortality. The dashed diagonal line represents perfect agreement between predicted and observed probabilities of in-hospital mortality (ideal calibration), whereas the black circles represent the observed probability of in-hospital mortality within each decile of predicted risk.

The optimal cut-off value of the model was determined as 0.20, yielding a sensitivity of 86.2% and a specificity of 77.9% for predicting mortality. Using the same cohort, the optimal cut-off value for the APACHE II score yielded a sensitivity of 78.7% and a specificity of 71.5%. The proposed model achieved higher sensitivity and specificity than APACHE II at the respective optimal cut-off values.

Discussion

In this study, we developed and internally validated a multivariable prognostic model for predicting in-hospital mortality in critically ill patients with AECOPD. To ensure statistical independence, only the first ICU admission for each patient was included. The final model, incorporating age, sex, urea, albumin, CRP, and CCI, demonstrated better discriminatory performance than APACHE II. In addition, both calibration analysis and bootstrap validation confirmed the robustness and stability of the model.

Several prognostic models have been developed to facilitate early risk stratification in patients with acute exacerbation of COPD, including disease-specific tools such as DECAF and BAP-65, as well as general critical illness scores such as APACHE II. A recent systematic review and pooled analysis of 37 studies concluded that DECAF demonstrated the best overall discriminatory performance for predicting in-hospital mortality among existing AECOPD-specific models; however, the authors also reported a high risk of bias across all available prediction models and emphasized the need for further model development and external validation.¹³ Importantly, most of these models were derived and validated in general hospitalized AECOPD populations rather than exclusively in critically ill patients requiring intensive care. Our model was specifically developed in an ICU cohort using routinely available admission variables and demonstrated excellent discriminatory performance (AUC 0.899), with higher discriminatory performance than APACHE II in our cohort. Our findings should therefore be viewed as complementary rather than competitive with existing COPD-specific prognostic tools, and external validation in independent multicenter ICU cohorts is essential before routine clinical implementation.

Age remained an independent predictor of in-hospital mortality, with each additional year associated with a 7.3% increase in the odds of death. This finding is consistent with previous evidence identifying age as one of the strongest predictors of adverse outcomes in COPD. A systematic review and meta-analysis by Owusuaa et al demonstrated that increasing age was consistently associated with higher mortality across different COPD populations.¹⁴ Likewise, a recent systematic review of prognostic models for AECOPD identified age as the most frequently incorporated predictor in models predicting in-hospital mortality, underscoring its central role in risk stratification.¹³ The association observed in our study is biologically plausible, as advancing age is accompanied by reduced physiological reserve, greater frailty, impaired immune responses, and a higher burden of comorbidities, all of which may contribute to poorer outcomes following severe exacerbations.

Male sex was independently associated with increased in-hospital mortality in our multivariable model. Previous studies have also suggested that male sex may be associated with worse outcomes in COPD; however, the strength and consistency of this association remain uncertain. A recent systematic review and meta-analysis identified male sex as an independent predictor of mortality, although the reported effect size was substantially smaller than that observed in our cohort.¹⁴ Conversely, a large contemporary cohort of hospitalized patients with AECOPD found no significant difference in in-hospital mortality between men and women after hospitalization.¹⁵ The relatively strong association observed in our study may therefore reflect characteristics specific to our critically ill ICU population, including differences in disease severity, cumulative smoking exposure, comorbidity burden, and referral patterns. Consequently, although male sex may contribute to risk stratification in critically ill patients with AECOPD, this finding should be interpreted cautiously and requires external validation in independent multicenter ICU cohorts.

Charlson Comorbidity Index (CCI) was one of the strongest independent predictors of in-hospital mortality in our model. This finding is consistent with previous evidence demonstrating that comorbidity burden is a major determinant of adverse outcomes following acute exacerbations of COPD. In a multicenter cohort of 2,391 patients hospitalized for AECOPD, Spece et al reported that each one-point increase in CCI was independently associated with a higher risk of 30-day readmission or death (adjusted OR 1.24, 95% CI 1.18–1.30).¹⁶ The stronger association observed in our study (OR 3.386) is likely attributable to differences in the study population. Our model was derived exclusively from critically ill ICU patients, who generally have substantially greater disease severity than patients included in general hospitalized AECOPD cohorts. Therefore, the higher odds ratio should not be interpreted as evidence of a stronger biological effect but rather as reflecting the characteristics of a critically ill ICU population. Biologically, a higher CCI reflects the cumulative burden of chronic diseases, reduced physiological reserve, and increased susceptibility to complications during critical illness. Overall, our findings support the prognostic importance of comorbidity burden; however, the magnitude of this association should be interpreted cautiously until confirmed in independent multicenter ICU cohorts.

Serum albumin was independently associated with lower in-hospital mortality in our multivariable model, with each 1 g/L increase corresponding to an 11.6% reduction in the odds of death. Our findings are consistent with previous studies identifying hypoalbuminemia as an important marker of poor prognosis in patients with COPD. In a large retrospective cohort of 3,398 critically ill patients with COPD from the MIMIC-IV database, Ling et al demonstrated that lower serum albumin levels were independently associated with higher in-hospital mortality, with the strongest association observed when albumin concentrations were below 30 g/L.¹⁷ Similarly, Chen et al incorporated serum albumin into the newly developed AAAAN prognostic score for hospitalized patients with AECOPD, confirming its independent contribution to predicting in-hospital mortality.¹⁸ The prognostic value of albumin is biologically plausible because it reflects both nutritional status and systemic inflammation. In critically ill patients, hypoalbuminemia may also indicate increased capillary permeability, impaired hepatic protein synthesis, and greater disease severity. Overall, these findings support the inclusion of albumin as a readily available biomarker that complements clinical assessment for early mortality risk stratification in critically ill patients with AECOPD.

Urea was also independently associated with in-hospital mortality in our multivariable model. Our finding is consistent with accumulating evidence supporting blood urea nitrogen (BUN) as a simple yet robust prognostic biomarker in patients with AECOPD. In a retrospective cohort of 842 patients presenting to the emergency department with AECOPD, Chen et al demonstrated that higher admission BUN levels were independently associated with increased in-hospital mortality after adjustment for multiple clinical and laboratory variables.¹⁹ Similarly, in a study of 1,201 critically ill patients with AECOPD from the MIMIC-IV database, elevated BUN remained an independent predictor of in-hospital mortality even after propensity score matching.²⁰ The prognostic significance of urea is biologically plausible, as elevated levels may reflect impaired renal perfusion, neurohormonal activation, dehydration, increased protein catabolism, or concomitant cardiovascular dysfunction, all of which are common in severe COPD exacerbations. The consistent association observed across emergency department, hospitalized, and ICU cohorts supports the inclusion of urea as a readily available biomarker for early mortality risk stratification in critically ill patients with AECOPD.

C-reactive protein (CRP) was also independently associated with in-hospital mortality in our multivariable model, highlighting the role of systemic inflammation in determining the prognosis of acute exacerbations of COPD. Although the effect size for each 1 mg/L increase in CRP was modest (OR 1.008), this finding is expected because CRP is a continuous biomarker with a broad clinical range. A recent prospective study of hospitalized patients with AECOPD demonstrated that elevated CRP remained an independent predictor of clinically important adverse outcomes, including ICU admission, in-hospital mortality, and readmission, after multivariable adjustment.²¹ Consistent with our findings, previous evidence suggests that CRP reflects the intensity of systemic inflammation and is associated with increased disease severity and poor clinical outcomes in COPD.²² Taken together, these findings support the inclusion of CRP as a readily available biomarker that complements clinical assessment for early mortality risk stratification in critically ill patients with AECOPD. Although serum procalcitonin levels were significantly higher in non-survivors in the univariate analysis, this biomarker was not retained in the final multivariable model. Elevated PCT should be interpreted cautiously because it may reflect bacterial infection as well as non-infectious systemic inflammatory responses in critically ill patients. Furthermore, microbiological data were not systematically available; therefore, the relationship between infective etiology and mortality could not be evaluated.

This study has several strengths. First, the model was developed exclusively in critically ill patients admitted to the ICU, a population that has been underrepresented in previous prognostic studies of AECOPD. Second, the model relies solely on routinely available admission variables, facilitating bedside implementation without additional cost or specialized testing. Third, internal validation using bootstrap resampling demonstrated good model stability.

This study has several limitations. First, this was a retrospective single-center study, which may limit the generalizability of our findings. In addition, because only patients admitted to the ICU were included, the proposed model may not be generalizable to patients with acute exacerbation of COPD managed in emergency departments, hospital wards, or outpatient settings. Second, although the model demonstrated good discriminatory performance and calibration and underwent internal validation using bootstrap resampling, external validation in independent multicenter ICU cohorts is still required before routine clinical implementation. Third, because of the retrospective study design, several potentially important clinical variables, including frailty, functional status, smoking history, treatment-related factors, and detailed COPD-specific characteristics (eg, GOLD stage, pulmonary function test results, phenotype, maintenance inhaled therapy, long-term oxygen therapy, home non-invasive ventilation use, and asthma–COPD overlap [ACO]), were not systematically documented in the available medical records and therefore could not be evaluated. Pre-admission serum albumin measurements were also not systematically available. Likewise, detailed microbiological data, including culture results and causative pathogens, could not be evaluated because they were not consistently documented. Although these variables may influence prognosis, the primary aim of the present study was to develop a pragmatic prognostic model based on routinely available admission variables that are readily accessible at the time of ICU admission. Finally, direct comparison of our model with established COPD-specific prognostic tools such as DECAF and BAP-65 was not possible because some variables required to calculate these scores were not consistently available in our retrospective dataset. Future prospective multicenter studies are warranted to externally validate the model and directly compare its performance with established COPD-specific prognostic tools.

Conclusion

In conclusion, the proposed multivariable prognostic model demonstrated good discriminatory performance and calibration for predicting mortality in critically ill patients with acute exacerbation of COPD. The model showed superior predictive performance compared with the APACHE II score within our cohort and was based on readily available clinical and laboratory parameters. Further multicenter prospective studies with external validation are needed to confirm the generalizability and clinical utility of the model.

Funding Statement

The authors received no specific funding for this work.

Abbreviations

AUC, area under the curve; APACHE II, Acute Physiology and Chronic Health Evaluation II; CCI, Charlson Comorbidity Index; CI, confidence interval; COPD, chronic obstructive pulmonary disease; CRP, C-reactive protein; ICU, intensive care unit; OR, odds ratio; ROC, receiver operating characteristic.

Data Sharing Statement

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Ethics Approval and Informed Consent

The study protocol was approved by the Aksaray University Scientific Research Ethics Committee for Health Sciences (Ethics Committee Approval No. 2025/184; September 25, 2025). The study was conducted in accordance with the principles of the Declaration of Helsinki. Due to the retrospective observational design of the study, informed consent was waived.

Author Contributions

Ahmet Fırat conceived and designed the study, collected and verified the data, performed the statistical analyses, interpreted the results, and drafted the manuscript. Mert Akbulut contributed to data collection, data interpretation, and critical revision of the manuscript. Başak Toptaş Fırat contributed to data collection, data interpretation, and critical revision of the manuscript. Ayşegül Aykaç İbişoğlu contributed to study design, data interpretation, and critical revision of the manuscript. All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors report no conflicts of interest in this work.