Development and validation of parsimonious algorithms to classify acute respiratory distress syndrome phenotypes: a secondary analysis of randomised controlled trials

Abstract

Background: Using latent class analysis (LCA) in five randomised controlled trial (RCT) cohorts, two distinct phenotypes of acute respiratory distress syndrome (ARDS) have been identified: hypoinflammatory and hyperinflammatory. The phenotypes are associated with differential outcomes and treatment response. The objective of this study was to develop parsimonious models for phenotype identification that could be accurate and feasible to use in the clinical setting.

Methods: In this retrospective study, three RCT cohorts from the National Lung, Heart, and Blood Institute ARDS Network (ARMA, ALVEOLI, and FACTT) were used as the derivation dataset (n=2022), from which the machine learning and logistic regression classifer models were derived, and a fourth (SAILS; n=715) from the same network was used as the validation test set. LCA-derived phenotypes in all of these cohorts served as the reference standard. Machine-learning algorithms (random forest, bootstrapped aggregating, and least absolute shrinkage and selection operator) were used to select a maximum of six important classifier variables, which were then used to develop nested logistic regression models. Only cases with complete biomarker data in the derivation dataset were used for variable selection. The best logistic regression models based on parsimony and predictive accuracy were then evaluated in the validation test set. Finally, the models' prognostic validity was tested in two external ARDS clinical trial datasets (START and HARP-2) by assessing mortality at days 28, 60, and 90 and ventilator-free days to day 28.

Findings: The six most important classifier variables were interleukin (IL)-8, IL-6, protein C, soluble tumour necrosis factor receptor 1, bicarbonate, and vasopressor use. From the nested models, three-variable (IL-8, bicarbonate, and protein C) and four-variable (3-variable plus vasopressor use) models were adjudicated to be the best performing. In the validation test set, both models showed good accuracy (AUC 0·94 [95% CI 0·92-0·95] for the three-variable model and 0·95 [95% CI 0·93-0·96] for the four-variable model) against LCA classifications. As with LCA-derived phenotypes, the hyperinflammatory phenotype as identified by the classifier model was associated with higher mortality at day 90 (87 [39%] of 223 patients vs 112 [23%] of 492 patients; p<0·0001) and fewer ventilator-free days (median 14 days [IQR 0-22] vs 22 days [0-25]; p<0·0001). In the external validation datasets, three-variable models developed in the derivation dataset identified two phenotypes with distinct clinical features and outcomes consistent with previous findings, including differential survival with simvastatin versus placebo in HARP-2 (p=0·023 for survival at 28 days).

Interpretation: ARDS phenotypes can be accurately identified with parsimonious classifier models using three or four variables. Pending the development of real-time testing for key biomarkers and prospective validation, these models could facilitate identification of ARDS phenotypes to enable their application in clinical trials and practice.

Funding: National Institutes of Health.

Figure 1:. Overview of the analysis plan designed a priori for the primary analysis.

The portion of the plan above the dotted line were performed in the derivation dataset (black font) and the portion below the dotted line in the represents the portion of the analysis performed on the validation dataset (blue font).

Figure 2:. Receiver operator characteristics (ROC) two best performing regression models in the validation dataset and the respective model coefficients.

3-variable model: IL-8, bicarbonate, and Protein C; 4-variable model: IL-8, bicarbonate, Protein C, and Vasopressor use. AUC = Area under the curve. (Log = logarithm, e = 2·718281, IL-8 = interleukin-8).

Figure 3. Kaplan-Meier survival curve in HARP-2 stratified by phenotypes assigned using a 3-variable ancillary parsimonious model (interleukin-6, soluble tumour necrosis factor receptor-1, and vasopressor-use) and treatment (simvastatin or placebo).

Class was assigned using a probability cut-off of ≥ 0·5 to assign phenotype. The number of patients censored at the analysis end-point for each phenotype and treatment level are presented in brackets. A. Censored at 28 days; B. Censored at 90 days.

Figure 4. Box-Whisker plot depicting difference in key variables between the hyperinflammatory and hypo-inflammatory phenotypes in START using the 3-variable model (Interleukin-8, bicarbonate, and protein C) with a probability cut-off of ≥ 0·5 to assign phenotype.

A. Interleukin-6 (one value not shown in hypo-inflammatory class due to y-axis censoring for visual interpretation) B. Soluble tumour necrosis factor receptor-1 C. Platelet count D. PaO₂/FiO₂ (P-values are representative of Man-Whitney-U test).