AI Predicts Early Relapse Post-Axicabtagene Ciloleucel in Diffuse Large B-Cell Lymphoma Patients in a Multi-Center Real-World Study

Importance: Accumulating real-world (RW) evidence of axicabtagene ciloleucel (axi-cel) has demonstrated comparable safety and effectiveness to the pivotal trials in patients with relapsed or refractory diffuse large B-cell lymphoma (DLBCL). However, nearly 60% of patients receiving axi-cel eventually experience relapse, with most requiring additional therapies. Being able to identify or forecast patients developing early relapse enables clinicians to consider additional interventions to improve survival outcomes.

Objective: This study aimed to develop a clinically interpretable decision tree machine learning (ML) model to identify patients with risks of early relapse of their disease within six months after receiving axi-cel. We also comprehensively evaluated the real-world safety, including cytokine release syndrome (CRS), immune effector cell associated neurotoxicity syndrome (ICANS) and secondary T cell lymphoma incidence, and the effectiveness of axi-cel in a multi-center population across the University of California (UC) Health Systems.

Methods: This retrospective study included adult patients with DLBCL receiving axi-cel between October 2017 and December 2023 at five large UC tertiary academic medical centers. Computational pipelines were developed to perform data extractions, statistical analyses, and ML model development and evaluation. The primary endpoints included the survival and safety outcomes of axi-cel and the f1 score, a ML evaluation metric that combines precision and recall measuring the accuracy of the decision tree ML model in predicting early relapse within six months post-CAR-T infusion. The outcome endpoints included the medians and rates of progression-free (PFS) and overall (OS) survival at 18 months, as well as the rates of severe (≥ grade 3) CRS, and severe (≥ grade 3) ICANS. The secondary endpoints included the rates and median durations of other severe adverse events related to hematological profile and organ functions, hazard ratios (HR) of factors associated with PFS using the Cox proportional hazard regression model, and biomarkers associated with severe CRS and ICANS identified using Non-parametric Mann-Whitney tests. Finally, given recent reports, we assessed the frequency of secondary T cell malignancies within the treated population.

Results: 335 patients who received axi-cel were included in the study. 203 (60.6%) had aggressive DLBCL and comorbidities that would have excluded them from enrollment in the pivotal ZUMA-1 study of axi-cel. The median follow-up, PFS, and OS were 19.2, 10.1, and 38.4 months, respectively. The 18-month PFS and OS rates were 44.3% (95% CI, 38.2-50.2%) and 61.3% (95% CI, 55.0-67.1%), respectively. Severe CRS and ICANS were observed in 16.1% and 33.4% of patients, respectively. We developed a decision tree ML model using age and four routinely measured laboratory measures (LDH, ferritin, hematocrit, and platelet count). This straightforward ML model achieved a high f1 score of 0.72 at predicting patients experiencing early relapse within six months after receiving axi-cel. Our Cox risk factor analysis found that age ≥ 65 years, lower levels of lactate dehydrogenase (LDH), ferritin, and total bilirubin, and higher circulating platelets, hemoglobin, and eosinophils were associated with longer PFS. Our biomarker analysis found that severe CRS or ICANS were associated with significantly lower post-infusion albumin, prothrombin time, and circulating blood cells including, neutrophils, eosinophils, and leukocytes, but higher levels of LDH, c-reactive protein (CRP), and ferritin. One of 335 patients (0.3%) developed secondary T-cell lymphoma 28 days after CAR-T infusion.

Conclusion: This study computationally evaluated the RW performance of axi-cel using a diverse patient population across the UC Health System and further confirmed the safety and benefit of axi-cel in real-world patients. Our interpretable decision tree ML model demonstrated a novel approach to identify patients with risks of developing early relapse, using only age and 4 laboratory tests measured within 24h of CAR-T infusion, which enables clinicians to consider potential interventions in time to further extend survival outcomes of CAR-T therapy. Following confirmation, study findings suggest that our model and decision tree approach has strong potential to guide clinical decision making in patients with high-risk DLBCL.