Development and validation of tools for predicting the risk of death and ICU admission of non-HIV-infected patients with Pneumocystis jirovecii pneumonia

Abstract

Introduction: The mortality rate of non-HIV-infected Pneumocystis jirovecii pneumonia (PCP) is high. This research aimed to develop and validate two clinical tools for predicting the risk of death and intensive care unit (ICU) admission in non-HIV-infected patients with PCP to reduce mortality.

Methods: A retrospective study was conducted at Peking Union Medical College Hospital between 2012 and 2021. All proven and probable non-HIV-infected patients with PCP were included. The least absolute shrinkage and selection operator method and multivariable logistic regression analysis were used to select the high-risk prognostic parameters. In the validation, the receiver operating characteristic curve and concordance index were used to quantify the discrimination performance. Calibration curves were constructed to assess the predictive consistency compared with the actual observations. A likelihood ratio test was used to compare the tool and CURB-65 score.

Results: In total, 508 patients were enrolled in the study. The tool for predicting death included eight factors: age, chronic lung disease, respiratory rate, blood urea nitrogen (BUN), lactate dehydrogenase (LDH), cytomegalovirus infection, shock, and invasive mechanical ventilation. The tool for predicting ICU admission composed of the following factors: respiratory rate, dyspnea, lung moist rales, LDH, BUN, C-reactive protein/albumin ratio, and pleural effusion. In external validation, the two clinical models performed well, showing good AUCs (0.915 and 0.880) and fit calibration plots. Compared with the CURB-65 score, our tool was more informative and had a higher predictive ability (AUC: 0.880 vs. 0.557) for predicting the risk of ICU admission.

Conclusion: In conclusion, we developed and validated tools to predict death and ICU admission risks of non-HIV patients with PCP. Based on the information from the tools, clinicians can tailor appropriate therapy plans and use appropriate monitoring levels for high-risk patients, eventually reducing the mortality of those with PCP.

Keywords: ICU admission; Pneumocystis jirovecii pneumonia (PCP); clinical tool; death risk; non-HIV.

Figure 1

Predictor selection and prognostic nomogram development for predicting death risk of patients with PCP. (A) Identification of optimal parameters (lambda) in the LASSO model using minimum criteria and 5-fold cross-validation. Dotted vertical lines are drawn at the selected values using the minimum criteria and the 1 standard error of the minimum criteria (1-SE criteria); (B) LASSO coefficient profiles of 48 features; (C) predictor scores were found on the uppermost point scale that matched with the value of each variable. The sum of these scores is located on the total points axis, and a line is drawn straight down to get the percent probability of death. LASSO, least absolute shrinkage and selection operator; bpm, beat per minute; BUN, blood urea nitrogen; LDH, lactate dehydrogenase; CMV, cytomegalovirus; IMV, invasive mechanical ventilation.

Figure 2

Discrimination performance and predictive consistency of the tool for predicting death risk in internal validation. (A) ROC curve and AUC of the tool for predicting probability of death; (B) calibration curves of the tool predicting probability of death. The dotted line represents a perfect prediction by an ideal model, and the solid line represents the performance of the nomogram in the validation. A closer fit of the lines represents a better prediction. ROC, receiver operating characteristic; AUC, area under curve; ICU, intensive care unit.

Figure 3

Discrimination performance and predictive consistency of tools for predicting death in the external validation. (A) ROC curve and AUC of the tool for predicting probability of death; (B) calibration curves of the tool predicting probability of death. The dotted line represents a perfect prediction by an ideal model and the solid line represents the performance of the nomogram in the validation. A closer fit of the lines represents a better prediction. ROC, receiver operating characteristic; AUC, area under curve; ICU, intensive care unit.

Figure 4

Predictor selection and prognostic nomogram development for predicting ICU admission of patients with PCP. (A) Identification of optimal parameters (lambda) in the LASSO model using minimum criteria and 5-fold cross-validation. Dotted vertical lines are drawn at the selected values using the minimum criteria and the 1 standard error of the minimum criteria (1-SE criteria); (B) LASSO coefficient profiles of 41 features; (C) prognostic nomogram formulated for predicting ICU admission of patients with PCP. LASSO, least absolute shrinkage and selection operator; ICU, intensive care unit; PCP, Pneumocystis jirovecii pneumonia; bpm, beat per minute; BUN, blood urea nitrogen; CAR, C-reactive protein/albumin ratio; LDH, lactate dehydrogenase.

Figure 5

Predictive power assessment of the tool for predicting ICU admission in the validation. (A) ROC curve and AUC of predicting ICU admission. (B) calibration curves for predicting the probability of ICU admission. The dotted line represents a perfect prediction by an ideal model, and the solid line represents the performance of the nomogram in the validation. ICU, intensive care unit; ROC, receiver operating characteristic; AUC, area under curve.

Figure 6

Comparison of the predictive power between our tool and CURB-65 score. (A) ROC curve and AUC of predicting ICU admission in the external validation; (B) ROC curve and AUC of CURB-65 score for predicting ICU admission in our patients. ICU, intensive care unit; ROC, receiver operating characteristic; AUC, area under curve.