# Unravelling the complex inflammatory landscape of COVID-19 infection: a pathway to biomarkers identification in infection-associated delirium in the ICU

## TL;DR

This study identifies immune markers linked to delirium in ICU patients with severe COVID-19, using machine learning to improve prediction accuracy.

## Contribution

The novel use of high-dimensional immune profiling and synthetic data generation to identify and validate biomarkers for delirium in ICU patients with COVID-19.

## Key findings

- Delirious patients showed elevated CXCL1, CCL11, CXCL13, HGF, and VEGF-A, and reduced IL-1α, IL-21, and IL-22.
- Immune cell changes included increased exhausted B cells and decreased CXCR3+ CD4+ T cells and HLA-DR+ activated T cells.
- A machine learning model with synthetic data achieved 95% AUC in predicting delirium using 12 immune markers.

## Abstract

Delirium is a serious complication in patients with COVID-19-related acute respiratory distress syndrome (ARDS) admitted to the intensive care unit (ICU). Although numerous clinical risk factors have been identified, the immunologic pathways underlying delirium remain unclear. In this retrospective cohort study, we investigated high-dimensional immune signatures in ICU patients to delineate peripheral immune markers associated with delirium. We also explored machine learning (ML) approaches to enhance biomarker discovery and strengthen predictive modelling through synthetic data generation.

We studied a cohort of 62 COVID-19 ARDS patients admitted to the ICU at Lausanne University Hospital, Switzerland. The primary analysis compared patients within this cohort who developed delirium (n = 39) to those who remained delirium-free (n = 23). As a baseline for disease severity, we also compared the ICU cohort to 55 non-ICU COVID-19 patients and 450 healthy individuals. We performed high-dimensional immunophenotyping of cytokines, chemokines, and growth factors using multiplex beads assay, along with immune cell profiling via mass cytometry (CyTOF). Ridge regression has been employed to build classification models. We also generated synthetic samples using beta-variational autoencoders to improve sample size and subsequently model stability.

Delirious patients exhibited a distinctive immune signature, including elevated CXCL1, CCL11, CXCL13, HGF, and VEGF-A, coupled with reduced IL-1α, IL-21, and IL-22. Alterations in immune cell populations featured increased exhausted B cells and decreases in CXCR3 + CD4 + T cells, IgM + unswitched memory B cells, and HLA-DR + activated T cells. Leveraging these high-dimensional data, we trained ridge regression models to predict delirium. Incorporating synthetic data helped stabilize the models with a best-performing model achieving an area under the curve (AUC) of 0.95, with high sensitivity (93%) and specificity (86%), based on 12 identified markers.

Our findings demonstrate a distinct immune profile linked to ICU delirium and illustrate how ML can enhance biomarker discovery. Further prospective validation may refine these markers and guide precision-targeted interventions for mitigating delirium in critically ill populations.

The online version contains supplementary material available at 10.1186/s40635-025-00825-w.