Evaluating Protein Selection in Critical Care Proteomics: A Unified Validation Framework for ARDS in Sepsis

Sammanfattning

Identifying proteomic patterns associated with acute respiratory distress syndrome (ARDS) in sepsis patients is a central challenge in critical care research. Standard univariate approaches may miss non-linear associations and higher-order interactions that reflect the true complexity of the syndrome, yet systematic evaluation of alternative methods is lacking. This thesis introduces a unified evaluation framework for protein selection methods, combining simulation validation on synthetic data with planted signals of known type, classifier validation on held-out patient data, and pathway enrichment analysis. The framework is applied to a cohort of 512 adult sepsis patients from the SWECRIT multicentre Swedish ICU registry, of whom 73 developed moderate or severe ARDS, using proteomic measurements from the SomaScan 11K Assay. Three selection methods are evaluated: Welch’s t-test, mutual information with permutationbased significance, and random forest feature importance via SHAP values. The t-test is restricted by construction to mean-shift regulation, but identifies a small set of stable proteins associated with mitochondrial electron transport. Mutual information can in principle detect arbitrary univariate signals, but is power-limited at the cohort sample size and recovers few proteins at moderate effect sizes; nevertheless, its top-ranked proteins yield the strongest biological enrichment signal, with ketone body metabolism as the most robustly supported pathway. RF+SHAP also targets arbitrary univariate signals and produces the strongest classifier performance, with mean AUC up to 0.667 and consistent improvement over random selection, but lacks a principled mechanism for FDR control. No method recovers pure pairwise interaction signals, confirming that all approaches operate at the univariate level. The divergence between classifier and enrichment performance across methods demonstrates that complementary validation strategies are necessary to fully characterise a selection method.