Immunocompromise is common in the intensive care unit (ICU) and is strongly associated with adverse outcomes. However, robust quantitative tools for assessing the severity of immunocompromise are lacking. We aimed to develop and validate a machine learning-powered immunocompromise score and its risk stratification based on common immunocompromise conditions and biomarkers in the ICU.

A single-centre retrospective study was carried out in two ICUs of an academic tertiary care center in China. Adult patients who were admitted to the ICU for at least three days were enrolled. Feature selection was performed via the Boruta algorithm. The primary endpoint was 28-day all-cause mortality, whereas secondary endpoints included septic shock, the use of special antimicrobial agents, the peak levels of interleukin-6 (IL-6), etc. Seven machine learning models were developed via 10-fold cross-validation. The predicted probabilities from the optimal model were used to define the Immunocompromise and Severity (ICS) Score. To rapidly obtain patient immunocompromise information, a simplified ICS score (SICS score) was developed based on LASSO-selected features from day 1. Additionally, secondary endpoints were used to validate the rationale of ICS and SICS scores, and 216 patients were included for temporal validation.

A total of 1863 patients were included for Algorithm derivation, with 679 deaths (36.9%). Among the seven machine learning models, the XGBoost model using data from the first 3 ICU days showed the highest performance, defined as ICS score (AUC 0.887; sensitivity 0.896; specificity 0.723; accuracy 0.787), and its performance was superior to that of the APACHE II and SOFA ( P   <  0.001). The LASSO algorithm selected 5 key variables (age, organ failure, immune impairing diseases and treatments, and IL-6  >  100 pg/mL with lymphopenia  <  0.8 × 10 ⁹ /L) on ICU Day 1, for the SICS score (AUC 0.851; sensitivity 0.83; specificity 0.72; accuracy 0.765). The ICS and SICS scores effectively stratified patients into low, moderate, and high-risk groups with similar mortality rates (2.6% vs. 2.8%; 24.2% vs. 24.4%; 69.9% vs. 69.6%). Notably, the ICS score identified a greater proportion of patients in both the low (31.4% vs. 19.6%) and high risk (42.6% vs. 36.6%) groups. Secondary endpoints were significantly correlated with higher risk stratification, supporting the rationality of ICS and SICS scores. An additional 216 patients were used for temporal validation, yielding an AUC of 0.854 for the ICS score and 0.845 for the SICS score.

Derived from common immunocompromising conditions and biomarkers, the ICS and SICS scores and their risk stratification system provide an accurate and timely solution for identifying and stratifying immunocompromise in critically ill patients. This framework may facilitate early clinical decision-making and resource allocation.