Cytokine-based models for efficient differentiation between infection and cytokine release syndrome in patients with hematological malignancies

Abstract

Although the efficacy of chimeric antigen receptor (CAR)-T cell therapy has been widely demonstrated, its clinical application is hampered by the complexity and fatality of its side effects. Cytokine release syndrome (CRS) is the most common toxicity following CAR-T cell infusion, and its symptoms substantially overlap with those of infection. Whereas, current diagnostic techniques for infections are time-consuming and not highly sensitive. Thus, we are aiming to develop feasible and efficient models to optimize the differential diagnosis in clinical practice. This study included 191 febrile patients from our center, including 85 with CRS-related fever and 106 with infectious fever. By leveraging the serum cytokine profile at the peak of fever, we generated differential models using a classification tree algorithm and a stepwise logistic regression analysis, respectively. The first model utilized three cytokines (IFN-β, CXCL1, and CXCL10) and demonstrated high sensitivity (90% training, 100% validation) and specificity (98.44% training, 90.48% validation) levels. The five-cytokine model (CXCL10, CCL19, IL-4, VEGF, and CCL20) also showed high sensitivity (91.67% training, 95.65% validation) and specificity (98.44% training, 100% validation). These feasible and accurate differentiation models may prompt early diagnosis of infections during immune therapy, allowing for early and appropriate intervention.