AUTHORS
Zekai Mao, Yuting Gong2, Huilin Pan3, Yafang Guo4, Xingcheng Yang5, Yang Zhou6, Li Zhu7, Xiaoxi Zhou8, Xiaoya Hu9, Jia Wei, MD, PhD10
AFFILIATIONS
1 Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, Hubei, China
2 Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, Hubei, China
3 Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China
4 Shanxi Bethune Hospital, Shanxi Academy of Medical Sciences, Third Hospital of Shanxi Medical University, Tongji Shanxi Hospital, Wuhan, China
5 Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, Hubei, China
6 Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, China
7 Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, Hubei, China
8 Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, Hubei, China
9 Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China
10 Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, Hubei, China
ABSTRACT
Background
Severe early immune effector cell-associated haematotoxicity (eICAHT) is a frequent and clinically important complication after chimeric antigen receptor T-cell (CAR-T) therapy, contributing to infection, prolonged hospitalisation, transfusion dependence, and non-relapse mortality. Existing prediction scores show limited performance for severe eICAHT defined by the current European Hematology Association/European Society for Blood and Marrow Transplantation (EHA/EBMT) consensus criteria and may not fully capture complex pretreatment risk patterns. We therefore developed and temporally validated an interpretable machine-learning model for pretreatment prediction of severe eICAHT.
Methods
We retrospectively analysed 430 consecutive patients with relapsed/refractory B-cell non-Hodgkin lymphoma (R/R B-NHL) who received CAR-T therapy between January 2015 and December 2022. Patients were randomly divided into training (n=301) and internal validation (n=129) cohorts. A non-overlapping cohort of 78 patients treated between November 2023 and April 2026 served as temporal external validation. Severe eICAHT defined by EHA/EBMT criteria was the primary endpoint. Ninety-six routinely available pretreatment variables were screened, twelve machine-learning algorithms were compared, and SHapley Additive exPlanations (SHAP) were applied to interpret the final model.
Results
Severe eICAHT occurred in 254/430 patients (59.1%) in the development cohort. The final XGBoost model incorporated only ten routinely available pretreatment laboratory variables, including white blood cell count, neutrophil count, monocyte count, haemoglobin, platelet count, sodium, lactate dehydrogenase, uric acid, bicarbonate, and estimated glomerular filtration rate. In the internal validation cohort, the model achieved an AUROC of 0.771 (95% CI, 0.679–0.856), an AUPRC of 0.756, and a Brier score of 0.193. In a head-to-head comparison within our internal validation cohort using the same EHA/EBMT severe eICAHT endpoint, the XGBoost model demonstrated higher discrimination than the CAR-HEMATOTOX score (AUROC 0.771 vs 0.597), with improved specificity (0.660 vs 0.477). Similar improvements were also observed for sensitivity, overall accuracy, and F1 score. Performance remained stable in the temporally distinct validation cohort (AUROC 0.701, 95% CI 0.572–0.811). SHAP analysis identified impaired baseline haematopoietic reserve together with metabolic stress and disease-burden markers as the principal contributors to predicted severe eICAHT risk.
Conclusions
Using only ten routinely available pretreatment laboratory variables obtained before lymphodepleting chemotherapy, we developed and temporally validated an interpretable machine-learning model for severe eICAHT that demonstrated favourable performance and retained discrimination in a temporally independent cohort. Compared with an established clinical risk score, the model provided improved prediction of severe eICAHT while maintaining transparent patient-level interpretation. Because all predictors are routinely available before lymphodepleting chemotherapy, this model enables individualized pretreatment risk assessment and may support risk-adapted clinical decision-making, including optimization of treatment timing, supportive-care planning, monitoring intensity, and prospective evaluation of preventive interventions for patients at high risk of severe eICAHT. Prospective multicentre validation is warranted to determine whether pretreatment risk stratification can safely guide risk-adapted supportive strategies and improve patient outcomes.
DISCLOSURES
Zekai Mao: No relevant conflicts of interest to declare.