Fig. 2.

Summary of the primary clustering analyses in the three trials. ALVEOLI = Assessment of Low Tidal Volume and Elevated End-Expiratory Pressure to Obviate Lung Injury (N = 549), FACTT = Fluids and Catheters Treatment Trial (N = 1000), SAILS = Statins for Acutely Injured Lungs from Sepsis (N = 745). LCA = Latent class analysis, PAM = partitioning around medoids, HC = Hierarchical clustering, MOB = model based recursive partitioning, CF = Causal forest, XL-RF = X-learner with Random Forest (RF); XL-BART = Bayesian Additive Regression Trees (BART). Panel 2a: Optimal number of clusters and proportions of patients in each clusters per algorithm (The colours are representative of the size of the clusters). Panel 2b: Top 10 variable importance for each clustering algorithm. sTNFR-1 = Soluble tumour-necrosis factor receptor-1, IL-8 = Interleukin-8, WBC = White blood cell count, ICAM-1 = Intercellular adhesion molecule-1, VE = Minute ventilation, SP-D = Surfactant protein-D, IL-6 = Interleukin-6, PAI-1 = Plasminogen activator inhibitor-1, HR = Heart rate, P:F ratio = PaO₂/FiO₂, White (race), VT = Tidal volume, UO = Urine output, BMI = Body mass index, PEEP = = Positive end-expiratory pressure, vWF = Von Willebrand Factor, Pplat = Plateau pressure, SBP = Systolic blood pressure, Pmean = Mean airway pressure. Panel 2c: Odds ratio for heterogeneity of treatment effect in clusters for each algorithm (odds ratio > 1 was associated with harm, p-value represents the significance of the coefficient of the interaction term of randomised intervention and clusters in a logistic regression model with mortality at day 90 as the dependent variable; P-values were generated for the interaction term using the ANOVA likelihood ratio test).