Skip to main content

Early prediction of moderate-to-severe condition of inhalation-induced acute respiratory distress syndrome via interpretable machine learning



Several studies have investigated the correlation between physiological parameters and the risk of acute respiratory distress syndrome (ARDS), in addition, etiology-associated heterogeneity in ARDS has become an emerging topic quite recently; however, the intersection between the two, which is early prediction of target conditions in etiology-specific ARDS, has not been well-studied. We aimed to develop and validate a machine-learning model for the early prediction of moderate-to-severe condition of inhalation-induced ARDS.


Clinical expertise was applied with data-driven analysis. Using data from electronic intensive care units (retrospective derivation cohort) and the three most accessible vital signs (i.e. heart rate, temperature, and respiratory rate) together with feature engineering, we applied a random forest approach during the time window of 90 h that ended 6 h prior to the onset of moderate-to-severe respiratory failure (the ratio of partial pressure of arterial oxygen to fraction of inspired oxygen ≤ 200 mmHg).


The trained random forest classifier was validated using two independent validation cohorts, with an area under the curve of 0.9127 (95% confidence interval 0.8713–0.9542) and 0.9026 (95% confidence interval 0.8075–1), respectively. A Stable and Interpretable RUle Set (SIRUS) was used to extract rules from the RF to provide guidelines for clinicians. We identified several predictive factors, including resp_96h_6h_min < 9, resp_96h_6h_mean ≥ 16.1, HR_96h_6h_mean ≥ 102, and temp_96h_6h_max > 100, that could be used for predicting inhalation-induced ARDS (moderate-to-severe condition) 6 h prior to onset in critical care units. (‘xxx_96h_6h_min/mean/max’: the minimum/mean/maximum values of the xxx vital sign collected during a 90 h time window beginning 96 h prior to the onset of ARDS and ending 6 h prior to the onset from every recorded blood gas test).


This newly established random forest‑based interpretable model shows good predictive ability for moderate-to-severe inhalation-induced ARDS and may assist clinicians in decision-making, as well as facilitate the enrolment of patients in prevention programmes to improve their outcomes.

Peer Review reports


Acute respiratory distress syndrome (ARDS) is life-threatening and the major cause of morbidity and mortality in intensive care units (ICUs), with a mortality rate exceeding 40% [1, 2]. Despite the fact that therapies, such as ECMO, were well developed in the modern era, those did not remarkably reduce the mortality of the disease, this is because guidelines provide a uniform recommendation while neglecting the effect of case-mix in etiologies. Etiology-specific patients are more likely to respond to a given therapy [3]. Therefore, early prediction of patients with a high risk for developing moderate-to-severe ARDS and the use of prevention strategies for such patients are of great value in critical care units [4]. However, healthcare providers experience challenges in recognising ARDS patients [5], which may be due to the diversity of causes (e.g. inhalation, trauma, coronavirus disease [COVID-19]) [4]. Additionally, ARDS is rarely present at the time of hospital admission [5, 6], increasing the risk of clinicians not readily recognising, interpreting, and acting upon relevant information [7].

Because of the high mortality and difficulty in disease recognition, understanding the relationship between risk factors and ARDS is of considerable value. For example, serum zinc levels have been recommended as an appropriate biomarker for evaluating the risk of developing inhalation-induced ARDS [8]. Additionally, XGBoost, an open-source software library, has been adopted to identify ARDS based on non-invasive physiological parameters, such as heart rate and respiratory rate [9]. Alternatively, machine-learning has recently been used for the early prediction of ARDS [5]. Although ARDS prediction has attracted the attention of many researchers, previous studies have been focused on all-cause ARDS, without considering the etiology of the disease, and studies on etiology-specific ARDS prediction are lacking. In addition, it is challenging to translate machine-learning methods into clinical practice owing to the following reasons: (1) most models have not been validated using an independent test set and therefore have unknown generalisability [10]; (2) ensemble learning models (e.g. XGBoost), despite providing good prediction performance, lack transparency (interpretability); and (3) many models incorporate too many indices (e.g. laboratory blood tests), increasing their potential complexity and burden in clinical practice [5, 9,10,11].

To overcome the above obstacles, we focused on inhalation-induced ARDS. Smoke inhalation was one of the leading causes of pulmonary ARDS after pneumonia [3]. Owing to the nature of the injury (e.g. smoke from fires, smoke bombs), patients with inhalation-induced ARDS often present in groups, increasing pressure on clinical practitioners. Limited access to health care resources, such as low nurse-to-patient ratios, has made the situation worse; for example, it has caused delays in initiating and providing adequate treatment for deteriorating patients and higher ICU mortality [7]. Thus, this study aimed to develop a simple transparent (interpretable) model to predict early moderate-to-severe inhalation-ARDS with generalisability using data from eICUs [12], as a retrospective derivation cohort, and independently validate the model with Cohort 1 (PLAGH): patients with acute HC/Zinc chloride smoke inhalation lung injury who were admitted to the respiratory department at People's Liberation Army General Hospital (PLAGH) in 2014; Cohort 2: freely available critical care data from the Medical Information Mart for Intensive Care III (MIMIC-III; [13]. Our model is able to reliably predict the onset of moderate-to-severe inhalation-induced ARDS 6 h prior to onset.


This study used a random forest approach for the early prediction of inhalation-induced ARDS. To develop the model, we used a retrospective observational cohort, obtained from the eICU Collaborative Research Database that consists of inpatient ICU encounters at 32 de-identified medical centres between 2014 and 2015. Fifteen patients from respiratory department at PLAGH and nine patient encounters from the MIMIC-III database were used as a validation cohort. eICU and MIMIC-III databases are publicly accessible, and the MIMIC-III publication states that ‘the project was approved by the Institutional Review Boards of Beth Israel Deaconess Medical Center (Boston, MA) and the Massachusetts Institute of Technology (Cambridge, MA)’. Requirement for individual patient consent was waived because the project did not impact clinical care and all protected health information was deidentified. This study is reported following the STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) guidelines. Additional file 1 shows the completed STROBE checklist.

Trained investigators extracted data from electronic patient medical records from the two public cohorts. Selection of patients was based on ICD-9/10 codes and diagnostic key words. Patients with previous ICU admission were excluded. Each patient was cared for at a single medical centre. We applied additional inclusion criteria to focus the scope of our study. Patient stays that did not involve at least one recorded measurement of the variables were excluded [5, 11]. In addition, we included only patient stays with a duration within a specified window (see “Predictor variables” section for details). The final derivation cohort (i.e. eICU) consisted of five unique patients corresponding to 48 unique observations. The final validation cohort (i.e. PLAGH, MIMIC-III) comprised 15 unique patients (177 observations), and three unique patients corresponding to 19 observations. The characteristics of both cohorts are presented in Table 1. Figure 1 outlines the patient selection process and details the number of patients after each selection procedure.

Table 1 Characteristics of the final cohorts
Fig. 1
figure 1

The patient selection process (including the number of patients after each selection procedure)

Outcome variables

The primary outcome of interest in this study was the ratio of partial pressure of arterial oxygen to fraction of inspired oxygen (PaO2/FiO2 ratio) 6 h since investigation. ARDS was diagnosed according to the Berlin definition established in 2012 [14]. Depending on the severity of lung failure according to the PaO2/FiO2 ratio, ARDS is currently classified as mild (200 < PaO2/FiO2 ≤ 300), moderate (100 < PaO2/FiO2 ≤ 200), or severe (PaO2/FiO2 ≤ 100) [14, 15]. In the model development, the numeric ratio was converted to a binary outcome: presence of moderate or severe ARDS (PaO2/FiO2 ≤ 200) versus mild or better ARDS (PaO2/FiO2 > 200).

Predictor variables

For model development, the predictive power of three primary vital signs (body temperature, pulse (heart rate [HR]), and breathing rate (respiratory rate)), were evaluated. These were selected as predictor variables because they are the most accessible, non-invasive physiological parameters that are continually monitored. In addition, measurements for these vital signs generated in different platforms are easily compared as they are not affected by external factors or obtained by means of different medical devices. Thus, if a correlation between the vital signs and ARDS could be established, it would allow a prediction with simplicity and generalisability.

Feature engineering was performed before features were included in the random forest analysis. We adopted a “time-phased” strategy that calculated the minimum, maximum, and mean values of the three vital signs collected during a 90 h time window beginning 96 h prior to the onset of ARDS and ending 6 h prior to the onset from every recorded blood gas test (by which the PaO2/FiO2 ratio could be calculated) [16]. The feature’s spatiotemporal pattern representation is shown in Fig. 2 for three representative patients from each cohort.

Fig. 2
figure 2

The feature’s spatiotemporal pattern representation for three representative patients from each cohort

Model development

We adopted a random forest classifier, an ensemble learning method for classification, to predict the binary diagnosis outcome of ARDS. In the random forest analysis, each tree was constructed using a different bootstrap sample from the original training data [17]. The parameter “Number of trees” was set to 5,000 to ensure that every input row would be predicted at least a few times; all other parameters were set to default. During the initial model training process, nine predictor variables during the 90-h time window were included: minimum, maximum, and mean HR; minimum, maximum, and mean respiratory rate; and minimum, maximum, and mean temperature. The importance of each variable was ranked according to the mean decrease in Gini values. The top four features (i.e. resp_96h_6h_mean, HR_96h_6h_mean, resp_96h_6h_min, and temp_96h_6h_max) were finally selected because their importance, measured according to the mean decrease in Gini value, was higher than the average mean decreases in Gini for all candidate features. In this way, we could halve the number of features in the model for a small cost in accuracy.

Seven prediction models, including the random forest, were developed for comparison: multivariate logistics regression (MLR), Lasso regression, random forest and XGBoost were developed for both four-variable and all-variable sets as appropriate. Optimal parameters of ML algorithms (e.g., XGBoost) were obtained through cross validation on training set (i.e., eICU). Model performances were compared on two independent validation sets (i.e. PLAGH, MIMIC).

Statistical analysis

Cohort characteristics were expressed as mean (range) and proportions (n, %), as dictated by the data type. We used the analysis of variance (ANOVA) to compare means of continuous variables of more than two groups, and χ2 test to compare the frequencies of categorical variables between groups. Gini-based feature selection is implemented in the randomForest R package [18].

Receiver operating characteristic (ROC) curves were generated to quantify the predictive accuracy of the models, and the area under the curve (AUC) was used to assess the discriminatory ability of the models. Cut-off threshold for optimal point was determined by performing ROC curve analysis in the retrospective cohort [6].

Data were analysed using R V.4.0.1. Two-sided p values < 0.05 were considered statistically significant.


Performance on independent test set

The binary ARDS diagnosis prediction performance by our final model (i.e. RF with selected variables) is summarised using a ROC curve in Fig. 3. Model performance comparison was shown in Additional file 2: Table S1. The random forest with the selected four features performed better than other models (e.g., MLR, Lasso, XGB) The final model (i.e., Four-Var RF) demonstrated an area under the ROC curve (AUROC) of 0.9127 (95% confidence interval [CI] 0.8713–0.9542) and 0.9026 (95% confidence interval [CI] 0.8075–1) for early ARDS prediction with the independent validation sets (PLAGH, MIMIC-III) 6 h prior to the onset of moderate or severe ARDS. The corresponding ROC curve is shown in Fig. 3. The final model was internally validated and the bootstrap estimate of AUC was 0.8391 (500 times).

Fig. 3
figure 3

Receiver characteristic operating (ROC) curves of the application of the model on the independent validation sets (PLAGH, MIMIC III

We also measured the prediction performance of the model using other evaluation metrics (i.e. sensitivity, specificity, and accuracy. The results are presented in Table 2.

Table 2 Prediction performance of the model using different evaluation metrics

Model interpretation

The lack of interpretability of the ensemble learning methods is a strong limitation for applications involving clinical decision-making. To overcome this issue, we employed an interpretable random forest model, Stable and Interpretable RUle Set (SIRUS) [19], as implemented in the R package SIRUS [20], to extract rules from a tree ensemble based on their frequency of appearance. As stated in the original method article “the most frequent rules, which represent robust and strong patterns in the data, are ultimately linearly combined to form predictions” [21]. SIRUS outputs a simple set of 10 rules (with all parameters set to default). To generate the prediction for each specific instance, SIRUS considers whether the conditions for each rule are satisfied to assign one of the two possible likelihood output values. In the next step, the 10-rule outputs were averaged to provide the predicted probability of the onset of moderate or severe ARDS. The 10-rule set is presented in Table 3. Overall, the rules, generated by the interpretable model, provide transparent and concise information regarding how the prediction is made, thus overcoming the aforementioned challenges hindering translational research of ARDS.

Table 3 Rules for calculating the likelihood of moderate to severe ARDS developing in 6 h


In this study, we developed a machine-learning model for the early prediction of inhalation-induced ARDS (moderate-to-severe condition) at 6 h prior to onset in critical care units. The model demonstrated good prediction performance and clinical interpretability on eICU data. Specifically, the model achieved a high AUROC with two completely independent external validation set (i.e. PLAGH, MIMIC-III). All features used in the model were derived from clinical variables that are routinely/continuously collected in ICUs, which can be simply and non-invasively obtained. This clinically applicable model can be easily applied to ICU patients to assist clinical decision-making and thus, holds great potential for improving patient outcomes.

Importantly, the model provided a convenient way of calculating the likelihood of the target event, i.e. the development of inhalation-induced ARDS (moderate-to-severe condition). To the best of our knowledge, no previous study has used the SIRUS in a clinical setting. The model improved prediction performance based on the random forest approach while maintaining transparency. More precisely, the rule set and how each of the rules contributed to the prediction was clearly shown. Instead of a decision-making path (where rules work interactively), those rules work independently towards a probability, making mechanistic exploration and/or explanation more realistic. It is noteworthy that we conducted experiments with both linear (e.g., MLR, Lasso) and non-linear (e.g., RF, XGBoost) algorithms, and chosen RF based on a data-driven approach, because the discrimination ability of RF is better than other algorithms (the comparison of the prediction performance of different ML algorithms in validation sets was in Additional file 2: Table S1).

To generate the predicted likelihood of the presence of moderate to severe acute respiratory distress syndrome (ARDS), the model considers whether the conditions for each rule are satisfied to assign one of the two possible output values. For example: resp_96h_6h_min = 5:0 (minimum value of respiratory rate of all available respiratory rate measured during 96–6 h prior to the target time point), then the first rule is satisfied, returning p (1) = 0.864. Next, the 10 rule outputs are averaged to provide the predicted probability of ARDS (moderate or severe) onset.

There are several vital signs typically examined by medical professionals, such as temperature, HR, and respiratory rate. Additionally, although blood pressure is not considered a vital sign, it is often measured alongside the vital signs. However, the three vital signs are generally more accessible and accurate than blood pressure, which is more prone to influence from external factors [22]. Moreover, model simplicity can lead to higher feasibility of implementation, particularly in low-income countries that lack ICU capacity. Furthermore, because of infrequent vital sign monitoring and a lack of standardised management practices in resource-limited settings, basic monitoring measures need to be better used [23]. Thus, we included only three simple, non-invasive, routinely monitored vital signs in our model.

The risk factors identified in our study are consistent with those from previous studies and were validated using an independent validation set. In 2020, Liu et al. compared the most commonly used early warning scoring systems for hospitalised patients with and without infection at risk for in-hospital mortality and transfer to the ICU and found that the National Early Warning Score (NEWS) exhibited the highest discrimination for mortality (followed by the Modified Early Warning Score-MEWS) [24]. The main findings of the present study agreed with the role of the NEWS [25]. For example, a respiratory rate ≤ 9 breaths/min increases the risk of developing critical or moderate-to-severe ARDS according to both the NEWS and our model. To further compare the discrimination capability of our model to that of commonly used risk scoring systems, we calculated MEWS and Systemic Inflammatory Response Syndrome (SIRS) for each data point of the included patients in the 90-h time window and the real time of the recorded blood gas test both. Discrimination capability of the scores based on AUCs was compared in Additional file 2: Table S2. Our model had a better performance in early warning of moderate to severe ARDS given the 90-h window performance and it also had a better reproducibility and generalizability as its AUC is stable on two independent validation sets both.

Rather than providing a binary outcome, our model is able to calculate the likelihood of developing moderate-to-severe ARDS. This was achieved by aggregating the vital sign measurements collected during hospital stay to yield the score used in our model, allowing the output to reflect risk for the target event. In addition, our prediction leaves a definite time window prior to the event, making potential intervention realistic. Such time-phase determination is suitable for clinical practice as clinical practitioners generally prefer having a suitable time window that is neither too far nor too near to onset, in order to leave sufficient time for intervention. The exact time window for feature engineering can be treated as a hyperparameter, which may vary slightly across different cohorts (e.g. patient characteristics) [16]. However, parameter tuning was outside the scope of this study.

In comparison with other similar studies using machine-learning models for the early prediction of ARDS, our model showed improved prediction performance. The prediction performance measured using the AUROC ranged from 0.75 to 0.87 in previous studies [5, 26, 27], whereas our model achieved an AUROC of being greater than 0.90 with two an independent validation sets. In addition, the predictors used in our model were further simplified [5]. A potential explanation may be that the retrospective cohort included in this study were etiology-specific ARDS patients (ARDS patients with inhalation injury). The improved modelling performance suggests that the etiology of ARDS could be used to identify a more homogeneous subset of ARDS for prediction enrichment which is an important component of etiology-associated heterogeneity in ARDS [3].

Although the prediction ability of the model was validated using two completely unrelated data sets, our study had some limitations. Similar to the limitations in previous data challenges “the analysis was performed in a retrospective setting. The generalisability and stability of the proposed model need to be evaluated systematically in prospective settings” [16]. Moreover, we included first admissions only, which might be a potential source of selection bias. We also acknowledge the limitation that the sample size was small. In order to overcome the issue, we extended the limited number of patients to hundreds of data points at the first place to ensure the credibility and power of model training, model validation and model performance evaluation, for example, the derivation set contained 48 data points with each class label being balanced; Secondly, the patients from different cohorts (i.e., eICU, MIMIC, PLAGH) had diversity in patient profiles, the model developed and validated on cohort with diversity had reproducibility and generalizability to new patients thereby; Last but not least, our final model was validated internally (e.g., bootstrap, LOOCV) and on two independent validation cohorts which were geographically distinct and demographically distinct from the derivation cohort. We believe that studies with a small number of subjects can quickly address research questions in a relatively short space of time; thus, a more efficient allocation of resources (e.g. subjects, time, financial costs) will be achieved by first testing a new research hypothesis in a small number of subjects [28].


In this study, an interpretable machine-learning model with key features (derived from three non-invasively measured vital signs) was successfully established for predicting moderate-to-severe condition of inhalation-induced ARDS 6 h prior to onset. The prediction model is intended to improve communication between nursing staff and junior doctors and “flag” patients who need to be given immediate priority [29]. Inhalation-induced ARDS typically affects individuals such as fire fighters and soldiers, who are often in otherwise good condition without chronic diseases, and providing critical care expertise at an early stage is extremely important for improving patient outcomes.

In future research, we hope to include more patients, while expanding to other etiology-specific ARDS, such as COVID-19, as the current ARDS predictions for patients with COVID-19 are either risk factor analysis [30], real-time risk scoring systems [31], and/or utilising relatively complex predictors, such as the neutrophil-to-lymphocyte ratio [32]. We hope to accomplish this aim by integrating this early prediction model in our own ICU risk management system to evaluate its effectiveness.

In conclusion, as clinical instances accumulate and clinical records become more comprehensive, this study forms a basis for evaluating the effectiveness of personalised intervention (e.g. vital sign-directed therapy) [33]. etiology-specific medicine is a critical component of precision, personalised healthcare and will be the core of effective care in the anticipatable future.

Availability of data and materials

The data that support the findings of this study are available from Chinese PLA General Hospital, but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of Chinese PLA General Hospital.



Acute respiratory distress syndrome


Area under the curve


Confidence interval


Heart rate


Intensive care units


Modified Early Warning Score


Medical Information Mart for Intensive Care III


National Early Warning Score


Systemic Inflammatory Response Syndrome


Random forest


Receiver operating characteristic


Stable and Interpretable Rule Set


  1. Villar J, Blanco J, Añón J, Santos-Bouza A, Blanch L, Ambrós A, et al. The ALIEN study: incidence and outcome of acute respiratory distress syndrome in the era of lung protective ventilation. Intensive Care Med. 2011;37:1942.

    Article  Google Scholar 

  2. Reyna MA, Josef CS, Jeter R, Shashikumar SP, Sharma A. Early prediction of sepsis from clinical data: the PhysioNet/Computing in Cardiology Challenge 2019. Crit Care Med. 2019;48(2):1.

    Google Scholar 

  3. Ruan S-Y, Huang C-T, Chien Y-C, Huang C-K, Chien J-Y, Kuo L-C, et al. Etiology-associated heterogeneity in acute respiratory distress syndrome: a retrospective cohort study. BMC Pulm Med. 2021;21(1):1–9.

    Article  Google Scholar 

  4. Luo J, Yu H, Hu Y-H, Liu D, Wang Y-W, Wang M-Y, et al. Early identification of patients at risk for acute respiratory distress syndrome among severe pneumonia: a retrospective cohort study. J Thorac Dis. 2017;9(10):3979.

    Article  PubMed  PubMed Central  Google Scholar 

  5. Le S, Pellegrini E, Green-Saxena A, Summers C, Hoffman J, Calvert J, et al. Supervised machine learning for the early prediction of acute respiratory distress syndrome (ARDS). J Crit Care. 2020;60:96–102.

    Article  PubMed  Google Scholar 

  6. Trillo-Alvarez C, Cartin-Ceba R, Kor DJ, Kojicic M, Kashyap R, Thakur S, et al. Acute lung injury prediction score: derivation and validation in a population-based sample. Eur Respir J. 2011;37(3):604–9.

    Article  CAS  PubMed  Google Scholar 

  7. Hyland SL, Faltys M, Hüser M, Lyu X, Merz TM. Early prediction of circulatory failure in the intensive care unit using machine learning. Nat Med. 2020;26(3):364–73.

    Article  CAS  PubMed  Google Scholar 

  8. Xie F, Zhang X, Xie L. Prognostic value of serum zinc levels in patients with acute HC/zinc chloride smoke inhalation. Medicine. 2017;96(39):e8156.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  9. Yang P, Wu T, Yu M, Chen F, Wang C, Yuan J, et al. A new method for identifying the acute respiratory distress syndrome disease based on noninvasive physiological parameters. PLoS ONE. 2020;15(2):e0226962.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  10. Ding X-F, Li J-B, Liang H-Y, Wang Z-Y, Jiao T-T, Liu Z, et al. Predictive model for acute respiratory distress syndrome events in ICU patients in China using machine learning algorithms: a secondary analysis of a cohort study. J Transl Med. 2019;17(1):1–10.

    Article  CAS  Google Scholar 

  11. Le S, Pellegrini E, Green-Saxena A, Summers C, Hoffman J, Calvert J, et al. Supervised machine learning for the early prediction of acute respiratory distress syndrome (ARDS). medRxiv. 2020.

  12. Pollard TJ, Johnson AE, Raffa JD, Celi LA, Mark RG, Badawi O. The eICU Collaborative Research Database, a freely available multi-center database for critical care research. Sci Data. 2018;5(1):1–13.

    Article  Google Scholar 

  13. Johnson AE, Pollard TJ, Shen L, Li-Wei HL, Feng M, Ghassemi M, et al. MIMIC-III, a freely accessible critical care database. Sci Data. 2016;3(1):1–9.

    Article  Google Scholar 

  14. Force ADT, Ranieri V, Rubenfeld G, Thompson B, Ferguson N, Caldwell E. Acute respiratory distress syndrome. JAMA. 2012;307(23):2526–33.

    Google Scholar 

  15. Hernández-Beeftink T, Guillen-Guio B, Villar J, Flores C. Genomics and the acute respiratory distress syndrome: current and future directions. Int J Mol Sci. 2019;20(16):4004.

    Article  PubMed Central  Google Scholar 

  16. Li X, Xu X, Xie F, Xu X, Sun Y, Liu X, et al. A time-phased machine learning model for real-time prediction of sepsis in critical care. Crit Care Med. 2020;48(10):e884–8.

    Article  PubMed  Google Scholar 

  17. Yang F, Lu W-H, Luo L-K, Li T. Margin optimization based pruning for random forest. Neurocomputing. 2012;94:54–63.

    Article  Google Scholar 

  18. RColorBrewer S, Liaw MA. Package ‘randomForest.’ Berkeley: University of California, Berkeley; 2018.

    Google Scholar 

  19. Bénard C, Biau G, Veiga SD, Scornet E. SIRUS: stable and interpretable RUle set for classification. Electron J Stat. 2021;15(1):427–505.

    Article  Google Scholar 

  20. Bénard C, Biau G, Da Veiga S, Scornet E. SIRUS: stable and interpretable RUle set. Electron J Stat. 2020;15:427–505.

    Google Scholar 

  21. Bénard C, Biau G, Da Veiga S, Scornet E. Interpretable random forests via rule extraction. arXiv preprint arXiv:2004.14841 (2020).

  22. Sapra A, Malik A, Bhandari P. Vital sign assessment—PMID: 319859942019.

  23. Cummings MJ, Wamala JF, Bakamutumaho B, Davis JL. Vital signs: the first step in prevention and management of critical illness in resource-limited settings. Intensive Care Med. 2016;42(9):1–2.

    Google Scholar 

  24. Liu VX, Lu Y, Carey KA, Gilbert ER, Afshar M, Akel M, et al. Comparison of early warning scoring systems for hospitalized patients with and without infection at risk for in-hospital mortality and transfer to the intensive care unit. JAMA Netw Open. 2020;3(5):e205191-e.

    Article  Google Scholar 

  25. Physicians RCo. National Early Warning Score (NEWS) 2: standardising the assessment of acute-illness severity in the NHS. Updated report of a working party. London: RCP2017; 2017.

    Google Scholar 

  26. Taoum A, Mourad-Chehade F, Amoud H. Early-warning of ARDS using novelty detection and data fusion. Comput Biol Med. 2018;102:191–9.

    Article  PubMed  Google Scholar 

  27. Zhou M, Sharma R, Zhu H, Li Z, Fan X. Rapid breath analysis for acute respiratory distress syndrome diagnostics using a portable two-dimensional gas chromatography device. Anal Bioanal Chem. 2019;411(24):6435–47.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  28. Hackshaw A. Small studies: strengths and limitations. Eur Respir J. 2008;32(5):1141–3.

    Article  CAS  PubMed  Google Scholar 

  29. Gardner-Thorpe J, Love N, Wrightson J, Walsh S, Keeling N. The value of Modified Early Warning Score (MEWS) in surgical in-patients: a prospective observational study. Ann R Coll Surg Engl. 2006;88(6):571–5.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  30. Xu W, Sun N-N, Gao H-N, Chen Z-Y, Yang Y, Ju B, et al. Risk factors analysis of COVID-19 patients with ARDS and prediction based on machine learning. Sci Rep. 2021;11(1):1–12.

    Article  Google Scholar 

  31. Liang M, He M, Tang J, He X, Liu Z, Feng S, et al. Novel risk scoring system for predicting acute respiratory distress syndrome among hospitalized patients with coronavirus disease 2019 in Wuhan, China. BMC Infect Dis. 2020;20(1):1–10.

    Article  Google Scholar 

  32. Ma A, Cheng J, Yang J, Dong M, Liao X, Kang Y. Neutrophil-to-lymphocyte ratio as a predictive biomarker for moderate-severe ARDS in severe COVID-19 patients. Crit Care. 2020;24(1):1–4.

    Article  Google Scholar 

  33. Hvarfner A, Blixt J, Schell CO, Castegren M, Baker T. Vital signs directed therapy for the critically ill: improved adherence to the treatment protocol 2 years after implementation in an intensive care unit in Tanzania. Emerg Med Int. 2020;2020:4819805.

    Article  PubMed  PubMed Central  Google Scholar 

Download references


We would like to thank Editage [] for editing and reviewing this manuscript for English language.


This study was supported by the following funds and projects: 1. National Natural Science Foundation of China (61976223); 2. Military Program (BLB20J002); 3. National Key Research Program of China (2019YFC0121703); 4. Chinese PLA General Hospital Program (CX19030, 2017MBD-011, 2018FC-WJFWZX-2-04).

Author information

Authors and Affiliations



FX and GX conceived the idea and supervised the study. JW contributed to the acquisition of data; CL, XL analysed data and performed statistical data analysis; JW, CL, LX, XL, KX contributed to the drafting of the manuscript; All authors reviewed and approved the final version. All authors read and approved the final manuscript.

Corresponding authors

Correspondence to Guotong Xie or Fei Xie.

Ethics declarations

Ethics approval and consent to participate

This study was performed in line with the principles of the Declaration of Helsinki. The Ethics Committee of Chinese PLA General Hospital approved the study and use of data (Approval No. of Ethics Committee: S2020-532-03). The Institutional Review Board (IRB) of Chinese PLA General Hospital waived the informed consent.

Consent for publication

Not applicable.

Competing interests

Drs. Li’s, and G. Xie’s institutions received funding from Chinese PLA General Hospital Program (Grant No. 2017MBD-011). The remaining authors have no competing interests to disclose.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Additional file 1.

Reporting items for observational research checklist - the STROBE Statement.

Additional file 2.

Prediction performances.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit The Creative Commons Public Domain Dedication waiver ( applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Wu, J., Liu, C., Xie, L. et al. Early prediction of moderate-to-severe condition of inhalation-induced acute respiratory distress syndrome via interpretable machine learning. BMC Pulm Med 22, 193 (2022).

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI:


  • Early prediction
  • Etiology-specific
  • ARDS (acute respiratory distress syndrome)
  • Critical care
  • Interpretable machine learning