Dynamic Lymphocyte Recovery Patterns Predict 90-Day Mortality in Sepsis: A Machine Learning-Enhanced Analysis of the MIMIC-IV Cohort

BackgroundSepsis-induced immunosuppression, characterized by lymphopenia, is associated with adverse outcomes. We aimed to identify distinct lymphocyte recovery patterns in patients with sepsis, evaluate their association with mortality, and develop a machine learning model to enhance prediction MethodsThis retrospective cohort study included adult patients with sepsis and initial lymphopenia (Absolute Lymphocyte Count [ALC] < 1.0 x 10L/L) from the Medical Information Mart for Intensive Care (MIMIC)-IV database. We defined three lymphocyte recovery patterns: "Persistent Suppression," "Partial Recovery," and "Complete Recovery." The primary outcome was 90-day all-cause mortality. Multivariable Cox models were used to assess the association between recovery patterns and mortality. An Extreme Gradient Boosting (XGBoost) model was developed to predict 90-day mortality. Results90-day mortality was highest in the Persistent Suppression group (49.1%) versus Partial (41.3%) and Complete Recovery (35.7%) groups (p<0.001). Persistent Suppression remained an independent predictor of mortality (adjusted Hazard Ratio 1.31, 95% CI 1.10-1.55). The XGBoost model achieved superior discrimination (Area Under the Receiver Operating Characteristic Curve [AUROC]=0.767) over traditional scores. SHAP analysis confirmed that dynamic features, including lymphocyte recovery metrics, were key model drivers. The model also demonstrated robust performance across various clinical subgroups (e.g., age, disease severity). ConclusionsThe trajectory of lymphocyte recovery following sepsis onset is an independent prognostic marker. Failure to restore lymphocyte counts is strongly associated with increased long-term mortality. Integrating this dynamic immunological feature into machine learning algorithms significantly enhances predictive accuracy, offering a promising tool for real-time risk stratification.