The Augmented PCM (IL-8, Ang2, sTNFR-1, serum bicarbonate) outperformed the existing IL-6 model for hyperinflammatory phenotype classification in sepsis/ARDS (ROSE AUROC 0.97 vs 0.93; p<0.0001).
Observational (n=4,173)
Yes
Does an augmented parsimonious classifier model including IL-8 and Ang-2 improve hyperinflammatory phenotype classification compared to the existing IL-6-based PCM in sepsis/ARDS patients?
An augmented parsimonious model incorporating IL-8, Ang-2, sTNFR-1, and serum bicarbonate outperforms the existing IL-6-based model for classifying hyperinflammatory phenotypes in sepsis/ARDS.
Effect estimate: AUROC (95% CI 0.96-0.98)
Absolute Event Rate: 0.97% vs 0.93%
p-value: p=< 0.0001
Abstract Rationale Latent Class Analysis (LCA) using 20-30 variables has identified molecular phenotypes with differential treatment responses in sepsis/ARDS. Consequently, a simplified parsimonious classifier model (PCM) using interleukin (IL)-6, soluble tumour necrosis factor (sTNFR-1), and serum bicarbonate is now being applied in prospective trials. This PCM may not fully capture the biology defined by LCA, given the importance of IL-8 and endothelial markers in capturing the Hyperinflammatory pathophysiology. We therefore evaluated whether models including IL-8 and/or angiopoietin-2 (Ang-2), a marker of endothelial injury, outperform PCM. Methods We derived the model combining (N = 1,604;Hyper=512) an observational cohort (EARLI) and clinical trial (FACTT). We used two newer clinical trials, ROSE (N = 1,006;Hyper=397) and CLOVERS (N = 1,563;Hyper=368) for validation. Three- or four-variable logistic regression models using a combination of IL-6, IL-8, sTNFR-1, ICAM-1, Ang-2, and serum bicarbonate were trained, with bicarbonate and at least one interleukin included in all models. We used highest area under receiver operating characteristic curve (AUROC) and lowest Bayesian Information Criterion (BIC) for selecting best models and compared their performance with the IL6/sTNFr1 PCM using precision recall curve (AUPRC), AUROC, sensitivity, specificity, and precision. We used mutual information (MI), a measure of mutual dependence, to quantify overlap between LCA-derived and the new parsimonious models’ derived probabilities. Results When compared directly, IL-8 models outperformed their equivalent IL-6 models with significantly higher AUROCs (p 0.001). Models with ICAM-1 had lowest overall performance metrics (data not shown). When comparing the three best performing new models: (IL-8, Ang2, Serum Bicarbonate), (IL-8, sTNFR-1, Serum Bicarbonate), and (IL-8, Ang2, sTNFR-1, Serum Bicarbonate) against the IL6/sTNFr1 PCM, the IL-8, Ang2, sTNFR-1, Serum Bicarbonate model (termed Augmented PCM) had the highest AUROC in both validation datasets ROSE: 0.97 (95%CI 0.96-0.98); CLOVERS: 0.95 (0.94-0.96). These AUROCs were significantly higher than the IL6/sTNFr1 PCM ROSE: 0.93 (0.91-0.94), CLOVERS: 0.92 (0.90-0.94); p 0.0001 both. Using a probability cutoff of 0.5 to assign phenotypes, the Augmented PCM was more sensitive than the PCM with comparable specificity (Panel A). The LCA-derived phenotype probabilities strongly correlated with those from the Augmented PCM (ROSE: r = 0.92; CLOVERS: r = 0.83; p 0.0001), with MI showing high overlapping alignment of their bimodal probability distributions (Panel B), which were significantly higher than the PCM. Conclusions We developed and validated a parsimonious model incorporating Ang-2, an endothelial marker, with inflammatory biomarkers. Its superior performance over the existing PCM suggests closer alignment with the biology identified by the original LCA models, supporting biological validity and statistical robustness. This abstract is funded by: R35HL177135 (CSC), R01HL173531, R35GM142992 (PS)
Tripathi et al. (Fri,) conducted a observational in Sepsis/ARDS (n=4,173). Augmented Parsimonious Classifier Model (IL-8, Ang-2, sTNFR-1, Serum Bicarbonate) vs. IL6/sTNFr1 Parsimonious Classifier Model was evaluated on Area under receiver operating characteristic curve (AUROC) for hyperinflammatory phenotype classification (ROSE validation dataset) (AUROC, 95% CI 0.96-0.98, p=< 0.0001). The Augmented PCM (IL-8, Ang2, sTNFR-1, serum bicarbonate) outperformed the existing IL-6 model for hyperinflammatory phenotype classification in sepsis/ARDS (ROSE AUROC 0.97 vs 0.93; p<0.0001).