A Nomogram for Predicting the Mortality of Patients with Acute Respiratory Distress Syndrome

Acute respiratory distress syndrome (ARDS) is an acute lung injury associated with high morbidity and mortality. This study aimed to establish an accurate prediction model for mortality risk in ARDS. 70% of patients from the Medical Information Mart for Intensive Care Database (MIMIC-III) were selected as the training group, and the remaining 30% as the testing group. Patients from a Chinese hospital were used for external validation. Univariate and multivariate regressions were used to screen the independent predictors. The receiver operating characteristic curve (ROC) analysis, the Hosmer–Lemeshow test, and the calibration curve were used for evaluating the performance of the model. Age, hemoglobin, heart failure, renal failure, Simplified Acute Physiology Score II (SAPS II), immune function impairment, total bilirubin (TBIL), and PaO2/FiO2 were identified as independent predictors. The algorithm of the prediction model was: ln (Pr/(1 + Pr)) = −3.147 + 0.037 ∗ age − 0.068 ∗ hemoglobin + 0.522 ∗ heart failure (yes) + 0.487 ∗ renal failure (yes) + 0.029 ∗ SAPS II + 0.697 ∗ immune function impairment (yes) + 0.280 ∗ TBIL (abnormal) − 0.006 ∗ PaO2/FiO2 (Pr represents the probability of death occurring). The AUC of the model was 0.791 (0.766–0.816), and the internal and the external validations both confirmed the good performance of the model. A nomogram for predicting the risk of death in ARDS patients was developed and validated. It may help clinicians early identify ARDS patients with high risk of death and thereby help reduce the mortality and improve the survival of ARDS.


Introduction
Acute respiratory distress syndrome (ARDS) is an acute lung injury characterized by progressive hypoxemia and respiratory distress and is associated with high morbidity and mortality [1][2][3]. In China, the incidence of ARDS was 27% in the ICU, and the mortality rate was as high as 25% to 75% [4]. e incidence is about 10.4% with the overall incidence of postoperative ARDS of about 3% [5]. Herein, it is of great significance to early identify ARDS patients with higher risk of death and to perform early intervention and treatment, which would help reduce the mortality risk of ARDS and improve the poor prognosis.
At present, many studies have extensively studied the risk factors for mortality of ARDS, including age [6][7][8], lower respiratory tract infection [8], immunosuppressive drugs [9,10], multiple organ failures [8,11], and some biomarkers [12,13]. However, due to multiple factors that worked together to determine the final outcome of ARDS patients, developing an effective prediction model would be of great clinical use for risk assessment. Currently, most of the established prediction models were limited by the small sample size, single study population, or lacking external validation [14,15]. erefore, this study aims to establish an accurate prediction model for mortality risk in ARDS based on the Medical Information Mart for Intensive Care Database (MIMIC-III) and perform external validation in a Chinese population.

Study Population.
In the retrospective study, we collected patient data from the Medical Information Mart for Intensive Care Database III version 1.3 (MIMIC-III v1.3) for the development of the prediction model. Inclusion criteria were as follows: (1) patients whose age ≥18 years and (2) patients who were diagnosed with ARDS according to the Berlin definition [1]. e MIMIC-III Database is a large, freely accessible database comprising information related to patients admitted to critical care unit at a large tertiary care hospital. It integrates deidentified, comprehensive clinical data of patients admitted to the Beth Israel Deaconess Medical Center in Boston, Massachusetts, and makes it widely accessible to researchers internationally under a data use agreement.
Also, ARDS patients in the First Affiliated Hospital of Zhengzhou University from June 2014 to December 2020 were enrolled in the study for external validation. Inclusion criteria were as follows: (1) patients whose age ≥18 years and (2) patients who were diagnosed with ARDS according to the Berlin definition [1].
e Ethics Committee of the First Affiliated Hospital of Zhengzhou University suggests that retrospective studies be exempted from ethical review. As the present study was a retrospective study, the Ethics Committee of the First Affiliated Hospital of Zhengzhou University exempted it from the requirement of the ethical review. All identifiable information about the patients has been stripped; the Ethics Committee of the First Affiliated Hospital of Zhengzhou University has waived the requirement for the informed consent in the study. Also, the study was conducted in line with the Declaration of Helsinki.

Development and Validation of the Prediction Model.
Firstly, the 70% of the study patients from the MIMIC-III Database were randomly selected as the training group for the development of the prediction model, and the remaining 30% as the testing group for the internal validation. Data of patients from the hospital were used for external validation. After developing the prediction model, we adopted the receiver operating characteristic curve (ROC) analysis, the Hosmer-Lemeshow test, and the calibration curve to evaluate the performance of the model. Univariate regression analysis was performed using the data of the training group from the MIMIC-III Database. Variables with statistical significance in the univariate analysis were included in the multivariate regression for stepwise screening, to screen the independent predictors and thereby to develop the model. e algorithm of the prediction model is as follows: the dependent variable y is 0 (represents survival) and 1 (represents death); the Pr value is the probability of death event.
en we used the maximum likelihood estimation (MLE) to estimate the coefficients of each variable.

Statistical Analysis.
Normally distributed measurement data were described as mean ± standard deviation (Mean ± SD), and the independent t-test was used for comparison between groups. Nonnormally distributed data were described as median and interquartile range M (Q1, Q3), and the Mann-Whitney U test was used for comparison. Besides, enumeration data were described as number of cases and constituent ratio N (%), and the chi-squared test or Fisher's exact test was used for comparison. We adopted the univariate and multivariate regression analysis to screen some independent predictors, and thereby these predictors were included in the prediction model to establish a prediction equation for assessing the risk of death in ARDS patients.
For visualizing the prediction model, we also plotted a nomogram.
en, the established model performed the internal and external validation, to assess the predicting performance of model. e receiver operating characteristic curve (ROC) analysis, the Hosmer-Lemeshow test, and the calibration curve were used for evaluating the performance of the model. e two-tailed test was carried out for all statistical tests, and P < 0.05 was considered statistically significant. e SAS 9.4 software (SAS Institute Inc., Cary, NC, USA) was used for the screening of independent predictors and the development of the prediction model. R 4.0.2 was used to validate and visualize the model.

Baseline Description.
In the present study, 1,814 patients were randomly selected from the MIMIC-III Database with 1,230 in the training group and 584 in the testing group. e mean age was 62.16 ± 16.93 years. ere were 1,048 (57.77%) males and 766 (42.23%) females. e ARDS of 150 (8.27%) patients was caused by pneumonia, the ARDS of 51 (2.81%) was by sepsis, and the ARDS of the remaining 1,613 (88.92%) was by other causes. Impaired immune function was reported in 544 (29.99%) patients, heart failure in 553 (30.49%) patients, and renal failure in 611 (33.69%) patients. 1,550 (85.45%) patients received ventilation and 264 (14.55%) did not, and the median ventilation time was 7.00 (3.00, 15.00) days. e median SAPS II was 38.00 (29.00, 48.00). e median GCS score was 9.00 (5.00, 14.00). e median SOFA score was 6.00 (4.00, 9.00). As shown in Table 2, there were no significant differences in baseline information and laboratory indicators between the randomly selected training group and the testing group (all P > 0.05).
Variables with statistical significance in the univariate analysis and the factors in the literature that have an impact on the prognosis of ARDS patients (ARDS causes and ventilation time [16]) were further included in the multivariate logistic regression. As shown in en we used the MLE to estimate the coefficients of each variable, and the algorithm of the prediction model was as follows: ln (Pr/(1 + Pr)) � −3.147 + 0.037 * age − 0.068 * hemoglobin + 0.522 * heart failure (yes) + 0.487 * renal failure (yes) + 0.029 * SAPS II + 0.697 * immune function impairment (yes) + 0.280 * TBIL (abnormal) − 0.006 * PaO 2 /FiO 2 (Pr represents the probability of death occurring). For visualizing the prediction model, we also plotted a nomogram (Figure 1). For example, as shown in Figure 2, the patient was 67.2 years old with normal TBIL. e hemoglobin was 7.4 g/dL and PaO 2 /FiO 2 was 22.2. SAPS II was 38. e patient was complicated with renal failure and immune function impairment, but no heart failure was reported. According to the nomogram, the total number of points was 373 and the corresponding predicted probability was 0.728, which indicated a high risk of death and was in line with the actual outcome of the patient.

Assessment and Validation of the Prediction Model.
According to the ROC analysis, the AUC value of the training group was 0.791 (0.766-0.816), and the AUC was 0.780 (0.743-0.816) in the testing group (Table 5), all suggesting the good discrimination of the model. e Hosmer-Lemeshow test (χ 2 � 49.123, P � 0.107), the ROC curves, and the calibration curves all indicated the good     Journal of Healthcare Engineering discrimination and calibration of the model (Figure 3). e Youden index suggested the cutoff value of 0.458. In the external validation, the AUC was 0.758 (0.756-0.761) ( Table 5). e Hosmer-Lemeshow test (χ 2 � 7.256, P � 0.509) and the calibration curves both suggested the good performance of the model in Chinese patients (Figure 3).

Discussion
In the present study, the prediction model based on eight predictors, age, heart failure, renal failure, immune function impairment, hemoglobin, TBIL, PaO 2 /FiO 2 , and SAPS II, was developed with good discrimination and calibration. e internal validation and external validation both confirmed the good performance of the model as reflected by the ROC analysis, the Hosmer-Lemeshow test, and the calibration curve. is may help clinicians predict the individual risk of death in ARDS patients. Respiratory system dysfunction is often characterized by hypoxemia and impairment of gas exchange with the most developed form as ARDS [17,18]. In the model, with the increase of hemoglobin and the oxygenation index of PaO 2 / FiO 2 , the risk of death was decreased. Villar et al. reported similar findings that patients with more severe lung disease tend to have lower PaO 2 /FiO 2 [15]. Our model also found that an older age was associated with an increased risk of death in ARDS patients. is was consistent with previous studies [6,7,12]. e body may experience functional  degeneration such as immune function impairment with the increase of age, leading to the decline of respiratory capacity and antibacterial capacity. In addition, older patients with ARDS may be complicated by other systemic diseases. e results of this study showed that both heart failure and renal failure independently increased the risk of death from ARDS. is was consistent with previous findings that multiple organ failures were responsible for death in ARDS patients [18,19]. Moran et al. found that although the proportion of severe ARDS patients who died of respiratory failure alone decreased, the number of deaths from multiple organ failures increased year by year [20]. Herein, in clinical treatment, attention should be paid not only to the elderly patients but also to the deterioration of ARDS caused by other systemic failures.
To our knowledge, there are few studies that have established prediction models for assessing the risk of death in ARDS patients [14,15,21]. e model developed by Gajic O et al. was well calibrated, but it required data of organ functions three days after intubation [21]. Villar et al. developed a risk model categorizing continuous variables into tertiles [15]. However, tertiles may not be appropriate for some variables have intricate dependencies and associations with outcome. In the study, based on a relatively large sample size, we incorporated demographic, clinical, and laboratory variables that were available in clinical use and all collected at the admission, allowing for early recognition of ARDS patients at a high risk of death. After univariate and multivariate logistic regressions, eight predictors were finally included in the model. e model was well discriminated as reflected by an AUC of 0.791 in the training set and 0.780 in the testing set and as confirmed by the Hosmer-Lemeshow test and the calibration curve. Also, we performed an external validation using data from a Chinese hospital, and the results indicated the good predictive ability of the model in Chinese patients. In addition, we plotted a nomogram for visualizing the model, which was more convenient for clinicians to predict the mortality risk of individual patients.
Several limitations should be considered in the study. First of all, data in our study were collected from the MIMIC-III Database and our hospital. To keep the uniformity of the variables in the datasets, the selection of variables was limited in some way. Also, the accuracy and the specificity were relatively poor and the sample size in the external validation set was relatively small. In the future, a prospective study with a larger sample size is preferred for validating our model.

Conclusion
In the present study, a nomogram for predicting the risk of death in ARDS patients was developed and validated. e model incorporated eight predictors that were available in clinical use. It may help clinicians early identify ARDS patients with high risk of death, which could make timely treatment therapies and interventions for reducing the mortality and improving the survival of ARDS patients.

Data Availability
e data used to support the findings of this study are available from the corresponding author upon request.

Conflicts of Interest
e authors declare no conflicts of interest.