A Simple-To-Use Nomogram to Predict Survival After Acute Respiratory Distress Syndrome

Background: The aim of this study to construct and validate a simple-to-use nomogram to predict the survival of patients with acute respiratory distress syndrome. Methods: A total of 197 patients with acute respiratory distress syndrome were selected from the Dryad Digital Repository. All eligible individuals were randomly stratied into the training set (n=133) and the testing set (n=64) as 2: 1 ratio. LASSO regression analysis was used to select the optimal predictors, and receiver operating characteristic and calibration curves were used to evaluate accuracy and discrimination of the model. Clinical usefulness of the nomogram was also assessed using decision curve analysis and Kaplan–Meier analysis. Results: Age, albumin, platelet count, Acute Physiology and Chronic Health Evaluation II score, PaO 2 /FiO 2 , lactate dehydrogenase, high-resolution computed tomography score, and syndrome etiology were identied as independent prognostic factors on LASSO regression analysis; these factors were integrated for the construction of the nomogram. Results of calibration plots, decision curve analysis, and receiver operating characteristic analysis showed that this model has good predictive ability of patient survival in acute respiratory distress syndrome. Moreover, a signicant difference in the 28-day survival was shown between the patients stratied into different risk groups (P < 0.001). Conclusions: We satisfactorily constructed a simple-to-use nomogram based on eight relevant factors to predict survival and prognosis of patients with acute respiratory distress syndrome. This model can aid personalized treatment and clinical decision-making.


Introduction
Acute respiratory distress syndrome (ARDS) is a clinically and pathophysiologically complex syndrome characterized by rapid progression and devastating hypoxemic respiratory failure [1]. Many risk factors, such as sepsis, pneumonia, pancreatitis, and major trauma, are associated with the development of ARDS [2]. Although there has been some progress in ARDS treatment in the last several decades, the prognosis of patients with ARDS are still not satisfactory. The in-hospital mortality rate of ARDS patients remains between 34% and 60% [3]. At present, the treatment of ARDS predominantly includes mechanical ventilation therapy [4]. Therefore, identi cation of novel and effective treatment strategies is crucial for patients with ARDS. Moreover, a simple-to-use clinical prediction model is also required to provide adequate care to patients with ARDS.
The severity of ARDS is often assessed using the PaO 2 /FiO 2 ratio, although this variable has a low-tomoderate prognostic value [5]. Recently, several biomarkers including in ammation cytokines, epithelial or endothelial damage, and coagulation have been established to evaluated prognosis and therapeutic response of patients with ARDS. For example, a meta-analysis reported that elevated plasma levels of angiopoietin-2 strongly correlate with diagnosis and mortality in populations at high risk of ARDS [6].
Moreover, various clinical biomarkers including lung in ammatory mediators (soluble suppression of tumorigenicity-2 and interleukin-6) [7] and products of epithelial and endothelial injury (the soluble form of the receptor for advanced glycation end products) [8,9] were developed to monitor pathophysiologic changes and outcomes of ARDS. Unfortunately, although few lung-speci c biomarkers have been validated to assess ARDS; however, none of them have been applied into clinical practice. Currently, there is no favorable prognosis prediction model for ARDS.
Nomograms (visualized graphs of a predictive model) are widely applied for prognosis and prediction of various diseases [10,11]. To date, no nomogram has been developed to predict the prognosis of ARDS patients. Therefore, a re ned model is needed to predict the prognosis of ARDS and guide clinical treatment. In this study, we aimed to construct a nomogram to predict the 28-day survival of patients with ARDS using several clinical parameters that are routinely used and readily available. This simple-to-use nomogram might serve as an early warning and prediction system for patients with ARDS.

Patients
A total of 197 patients with ARDS were extracted from the Dryad Digital Repository (http://www.datadryad.org/), which was shared by Anan et al [12]. All ARDS patients were diagnosed according to the Berlin de nition [5]. Patients with chronic interstitial lung disease (idiopathic pulmonary brosis), vasculitis or alveolar haemorrhage, hypersensitivity pneumonitis were excluded. All eligible patients were randomly strati ed into two groups in a 2:1 ratio (training set and validation set, respectively). The extracted clinical data included age, gender, white cell count (WBC), C-reactive protein, lactate dehydrogenase (LDH), albumin (Alb), platelet count (PLT), PEEP, APACHE II score, SOFA score, high-resolution computed tomography (HRCT) score, McCabe score, PaO 2 /FiO 2 , survival time, and survival status. Institutional ethical approval was not necessary because all the data were obtained from an online database.

Development of the nomogram
To obtain the subset of predictors, the LASSO regression analysis was used to select the optimal predictors from the risk factors in the training cohort. The "glmnet" package was used to perform the LASSO regression analysis [13,14]. Finally, using the selected predictors from the LASSO regression, a nomogram was developed using the "rms," "Fsurvival," and "foreign" R packages. A dynamic nomogram was constructed using "DynNom" and "shiny" packages.

Validation of the nomogram
To validate the constructed nomogram, the corresponding calibration map and receiver operating characteristic (ROC) analysis were performed in the training and validation sets to assess the prognostic accuracy of the nomogram by using the "rms," "survival," "foreign," "pROC," "wesanderson," and "openxlsx" R packages. In addition, decision curve analysis (DCA) was performed to quantify the clinical applicability of the nomogram.

Statistical analysis
The raw data were expressed as mean ± standard deviation when normally distributed, while expressed as median (interquartile range) when non-normally distributed . Differences between two groups were analyzed using chi-square tests for categorical variables and t-tests for continuous variables. The Kaplan-Meier method and the log-rank test were used to estimate survival. All statistical analyses were performed using R software (Version 3.6.2; http://www.Rproject.org). A two-sided P value < 0.05 was considered to indicate statistical signi cance.

Baseline characteristics
In total, 197 eligible ARDS patients with integrated information were randomly strati ed into two independent cohorts (training set, n = 133; validation set, n = 64). Patients' baseline clinical characteristics are shown in Table 1

Construction of the nomogram
A total of 14 parameters were used for LASSO regression, and eight parameters were selected as the optimal predictors by LASSO ( Figure 1A, 1B). The eight retained variables were then used to construct the predictive model. The risk-score for each individual was calculated based on the model coe cients combined with the corresponding value of the identi ed eight clinical parameters. Thereafter, the patients were classi ed into low-and high-risk clusters in both cohorts according to the median risk-score. Figures  1C, 1D show the risk-score distribution and the survival status of individual in the high-and low-risk cluster. The variables including Age, Alb, PLT, APACHE II score, PaO 2 /FiO 2 , LDH, HRCT were incorporated into the nomogram ( Figure 2). In addition, we developed a dynamic nomogram to predict prognosis of ARDS patients (https://tangyl.shinyapps.io/ARDS1/, Figure 3).

Performanceof the nomogram
The estimated 28-day survival probabilities could be obtained by drawing a perpendicular line from the total point axis to the outcome axis. The Kaplan-Meier survival curves revealed signi cantly poor overall survival in the high-risk group (p=4.7e-8; Figure 4A). Thereafter, we performed ROC analysis to assess the discriminability of the model. The area under the ROC curve (AUC) indicative of the 28-day survival prediction was 77.4% ( Figure 4B), which implied an e cacious performance of the nomogram to predict prognosis. The calibration plots based on the training set showed that the nomogram could accurately predict the 28-day survival ( Figure 4C). The results of DCA also exhibited that the nomogram could help clinicians to obtain maximum bene t when making clinical decisions ( Figure 4D).
To further study the predictive value of each parameter included in the nomogram, we performed ROC analysis for each of them ( Figure 5). The AUC values of all parameters were lower than that of the complete nomogram model. These results demonstrated that the nomogram had superior predictive performance and clinical value than any single factor.

Performance validation of the nomogram
To verify the reliability of the constructed novel nomogram, risk-scores were calculated in the validation set with the same formula that was used for calculating the risk-scores of patients in the training set. In the validation set, the distribution of risk-scores and the survival status ( Figure 6A, 6B) had a trend similar to that in the training set between high-and low-risk groups. Also, survival analysis indicated that low-risk patients had signi cantly favour prognosis than high-risk patients ( Figure 6C). ROC curves were used to assess the prognostic value of the risk-scores; the analysis results suggested that risk-scores could accurately predict the survival rate in patients ( Figure 6D). The calibration plot in the validation set also showed that the nomogram could accurately predict the 28-day survival ( Figure 6E).

Discussion
ARDS, one of the main critical diseases encountered in intensive care units, is a clinically and pathophysiologically complex syndrome of acute lung in ammation. Despite substantial progress in respiratory support strategies for critically ill patients, including the incorporation of a small tidal volume [15], high positive end-expiratory pressure [16], prone position ventilation [17], lung recruitment [18], use of neuromuscular blockers [19], high-frequency oscillatory ventilation [20,21], and extracorporeal membrane oxygenation [22,23], the mortality rate among patients with ARDS remains unacceptably high [24].
However, to our knowledge, no study has previously developed a nomogram to predict the prognosis of patients with ARDS.
Herein, we rst developed a nomogram using simple and easily available variables to evaluate the 28-day survival probabilities of ARDS patients whose information were extracted from a online database. Thereafter, we tested the performance of the nomogram in training and validation corhorts. Eight risk factors were identi ed in this model: age, Alb, PLT, APACHE II score, PaO 2 /FiO 2 , LDH, CT score, and ARDS etiologies. Additionally, our results showed that APACHE II score, PaO 2 /FiO 2 , and CT score could, albeit less accurately, predict the survival probability of ARDS patients compared to our novel model. These results suggest that the nomogram could be used as a cost-effective tool to predict the prognosis of ARDS and assist with clinical decision-making.
In 2012, the Berlin ARDS Society de ned the severity of ARDS according to the oxygenation index [5]. The oxygenation index (PaO 2 /FiO 2 ) was helpful to categorize ARDS patients with different severity, and the mortality was reported to be higher in more severe stages of ARDS (mild, moderate, or severe) ( [5,25]. However, these severity categories have a low-to-moderate prognostic value to predict respiratory failure [26]. Kamo and colleagues [27] reported that the severity strati cation of the Berlin ARDS criteria may have a low capacity to differentiate between mild and moderate ARDS. In this study, the results of ROC curve analysis also indicated that the oxygenation index had low prognostic power (AUC, 55.3204%), which was consistent with previous studies.
CT or other lung imaging techniques have been uesd as diagnostic tools to optimize lung assessment and ventilator management in patients with ARDS; however, it is still controversial whether CT ndings can predict ARDS outcomes [28][29][30]. HRCT scores have been reported to correlate with the pathological stage of diffuse alveolar damage [31]. Ichikado and colleagues [32] noted that HRCT score was one of the independent predictors of death and ventilator dependency in ARDS patients. Simultaneously, HRCT score was also found to be associated with multiorgan failure and ventilator-associated complications [32]. In the present study, to increase model accuracy, HRCT score was incorporated into the nomogram.
To evaluate the performance of HRCT score as a prognostic biomarker for the survival of ARDS patients, we performed ROC analysis. Our results showed that the model t was signi cantly better than that of the one-factor HRCT model.
APACHE II score can be used as indicators to evaluate the prognosis among critically ill patients; it has been used worldwide to measure ICU performance [33]. The APACHE II score is calculated based on acute physiological parameters and chronic health conditions, all of which have signi cant effects on the predictive prognosis of ICU patients [34]. Hwang and colleagues [35] revealed that APACHE II score was a mortality predictor for ARDS patients, but that the accuracy was not high (AUC, 62.3%). Lesur and colleagues [36] reported that APACHE II score may be less predictive value when applied for ARDS patients, and that in those patients, it might be less accurate than other indicators, such as age.
Certain drugs have also been reported to have the potential to cause ARDS. It has been proved that molecular targeted therapy, such as methotrexate and certain herbal medicines, can cause severe respiratory failure or ARDS [37][38][39]. However, only few studies have focused on the prognostic role of different etiologies of ARDS. In the present study, our results indicated that there is a lower risk of death if ARDS is caused by drugs. However, these discrepancies may be partly related to differences in the dose and duration of drug treatments.
Our study has some limitations. Firstly, the nomogram model was developed mainly based on the eight variables. As these factors were unstable throughout the whole follow-up period, which may partly in uence the precision of the model. Secondly, only 197 patients were included in this study; further studies with bigger sample sizes are needed. Thirdly, the lack of external validation may limit the extrapolation of the nomogram.
To summarize, we identi ed eight variables and developed a novel nomogram to predict prognosis in patients with ARDS. These results may help to further improve clinical decision-making and individualized treatment of ARDS patients. Also, this nomogram could distinguish patients with high-risk of ARDS, and further help to perform a careful follow-up among those patients.