Development and Validation of Risk Prediction Model for In-hospital Mortality Among Patients Hospitalized With Acute Exacerbation Chronic Obstructive Pulmonary Disease Between 2015 and 2019

Abstract

Background: In patients with chronic obstructive pulmonary disease (COPD), acute exacerbations affect patients' health and can lead to death. This study was aimed to develop a prediction model for in-hospital mortality in patients with acute exacerbations of COPD (AECOPD). Method: A retrospective study was performed in patients hospitalized for AECOPD between 2015 and 2019. Patients admitted between 2015 and 2017 were included to develop model and individuals admitted in the following 2 years were included for external validation. We analyzed variables that were readily available in clinical practice. Given that death was a rare outcome in this study, we fitted Firth penalized logistic regression. C statistic and calibration plot quantified the model performance. Optimism-corrected C statistic and slope were estimated by bootstrapping. Accordingly, the prediction model was adjusted and then transformed into risk score. Result: Between 2015 and 2017, 1,096 eligible patients were analyzed, with a mean age of 73 years and 67.8% male. The in-hospital mortality was 2.6%. Compared to survivors, non-survivors were older, more admitted from emergency, more frequently concomitant with respiratory failure, pneumothorax, hypoxic-hypercarbic encephalopathy, and had longer length of stay (LOS). Four variables were included into the final model: age, respiratory failure, pneumothorax, and LOS. In internal validation, C statistic was 0.9147, and the calibration slope was 1.0254. Their optimism-corrected values were 0.90887 and 0.9282, respectively, indicating satisfactory discrimination and calibration. When externally validated in 700 AECOPD patients during 2018 and 2019, the model demonstrated good discrimination with a C statistic of 0.8176. Calibration plot illustrated a varying discordance between predicted and observed mortality. It demonstrated good calibration in low-risk patients with predicted mortality rate ≤10% (P = 0.3253) but overestimated mortality in patients with predicted rate >10% (P < 0.0001). The risk score of 20 was regarded as a threshold with an optimal Youden index of 0.7154. Conclusion: A simple prediction model for AECOPD in-hospital mortality has been developed and externally validated. Based on available data in clinical setting, the model could serve as an easily used instrument for clinical decision-making. Complications emerged as strong predictors, underscoring an important role of disease management in improving patients' prognoses during exacerbation episodes.

Keywords: acute exacerbation of COPD; development; in-hospital mortality; prediction model; validation.

Figure 1

Discimination and calibration of in-hospital mortality prediction model in internal validation using data collected between 2015 and 2017. (A) AUC of prediabetes prediction model: the prediction model demonstrated excellent discrimination. The apparent C statistics and optimism-corrected C statistics were 0.9147 and 0.90887, respectively. After recalibration, C statistics remained the same and the estimate (95% CI) was still 0.9147 (0.8850, 0.9444). (B) Calibration plot of prediction model: calibration slope was 1.0254 (95% CI: 0.7276, 1.3233) and intercept was 0.0326 (95% CI: −0.7834, 0.8486). Hosmer–Lemeshow test was insignificant (chi-square statistics = 2.6249, df = 8, P = 0.9556).

Figure 2

Discrimination and calibration of in-hospital mortality prediction model in external validation using data collected between 2018 and 2019. (A) AUC of in-hospital mortality prediction model: C statistic was 0.8176 (95% CI: 0.7487, 0.8865) in patients hospitalized for AECOPD during 2018 and 2019. (B) Calibration plot of prediction model: calibration slope was 0.5986 (95% CI: 0.2409, 0.9563) and intercept was −1.4804 (95% CI: −2.8037, −0.1571). Hosmer–Lemeshow test indicated poor calibration of our prediction model in AECOPD inpatients between 2018 and 2019. When stratified by predicted mortality, the prediction model was well-calibrated in low-risk patients with predicted mortality rate ≤10%. The Hosmer–Lemeshow test was insignificant (chi-square = 9.2052, P = 0.3253). Whereas, the death risk was overestimated in patients with predicted mortality rate greater than 10% (chi-square = 281.1386, P < 0.0001). The overall Hosmer–Lemeshow test result was significant (chi-square = 17.0856, P = 0.0292).

Figure 3

Observed, predicted in-hospital mortality along with risk score and distribution of risk score in AECOPD patients between 2015 and 2017. Model predicted mortality was calculated using our final prediction model: Log(mortality/(1-mortality)) = −9.1935 + 0.0418*age +3.0752*Respiratory failure + 1.7674*Pneumothorax + 0.0243*Length of stay Risk score was the total risk points, which was the sum of each predictor's risk point based on individual values.