An assessment model for efficacy of autologous CD19 chimeric antigen receptor T-cell therapy and relapse or refractory diffuse large B-cell lymphoma risk

To the Editor: Autologous CD19 chimeric antigen receptor (CAR) T-cell therapy has emerged as a crucial treatment option for patients with relapsed or refractory diffuse large B-cell lymphoma (r/r DLBCL).[1] While CAR-T-cell therapy has the potential for curing a subset of patients despite its high cost, over 60% of recipients eventually experience relapse.[2] This troubling statistic has led several clinical centers to investigate the underlying reasons for CAR-T-cell therapy failure or suboptimal outcomes.[3,4] Currently, a reliable method for predicting the curative efficacy of CAR-T cells is lacking.[5] In this study, we conducted a comprehensive analysis to identify various predictors influencing the success of CAR-T-cell therapy. Our efforts have culminated in the development of a predictive model designed to assess the therapeutic efficacy of CAR-T-cell therapy. A total of 80 patients who were diagnosed with diffuse large B-cell lymphoma (DLBCL) and who were receiving CD19 CAR-T-cell therapy at Shanghai Tongji Hospital and Ruijin Hospital Affiliated to Shanghai Jiao Tong University School of Medicine were enrolled in this study. All patients voluntarily participated in the study and provided signed informed consent. The study protocols were approved by the Ethics Committee of Shanghai Tongji Hospital and Ruijin Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, and were conducted following the principles of the Declaration of Helsinki. We utilized data from CAR-T-cell clinical trials to establish the derivation cohort and real-world data to form the validation cohort [Supplementary Figure 1, https://links.lww.com/CM9/C191]. Details of the model development, validation process, and assessment of clinical utility are outlined in the Supplementary Methods, https://links.lww.com/CM9/C191. The derivation cohort had a median age of 59.5 years, ranging from 32.0 years to 79.0 years, and a median follow-up duration of 18.5 months, ranging from 1.2 months to 50.2 months. Notably, one participant experienced grade 3–4 cytokine relase syndrome (CRS). The validation cohort had a median age of 57.0 years, ranging from 29.0 years to 80.0 years, with a median follow-up period of 13.9 months, ranging from 1.5 months to 36.0 months. In this cohort, three individuals experienced grade 3–4 CRS, and one suffered from CAR-T-cell-related encephalopathy syndrome (CRES). The follow-up period ended on August 31, 2023. A detailed overview of the study population’s baseline demographic and disease characteristics can be found in Supplementary Table 1, https://links.lww.com/CM9/C191. The derivation cohort was categorized into two groups based on treatment efficacy: good and poor. The results of the intergroup analysis are presented in Supplementary Table 2, https://links.lww.com/CM9/C191. To ensure accuracy, we selected predictive factors with P values <0.1 that demonstrated a statistically significant trend in univariable logistic regression analysis. These factors included double expression (P = 0.037, 95% confidence interval [CI]: 1.112, 30.952), Eastern Cooperative Oncology Group (ECOG) performance status at enrollment (P = 0.028, 95% CI: 1.300, 97.372), the presence of bulky disease (P = 0.088, 95% CI: 0.755, 57.724), prior lines of therapy (P = 0.004, 95% CI: 2.095, 43.837), the TP53 status (P = 0.091, 95% CI: 0.818, 15.464), and LDH levels (P = 0.146, 95% CI: 0.999, 1.010). Treatment efficacy was defined as the dependent variable, as detailed in Supplementary Table 3, https://links.lww.com/CM9/C191. LDH levels greater than 250 U/L were binarized according to clinical thresholds and median values and subsequently analyzed via univariable logistic regression (P = 0.106, 95% CI: 0.786, 12.346). On the basis of the results of this single-variable logistic regression, we identified specific factors as potential predictors, applying a significance threshold of P <0.1. We confirmed that there was no collinearity among these predictors (P >0.05). These selected predictors were then used to develop a multivariable logistic regression model [Supplementary Table 4, https://links.lww.com/CM9/C191], which included double expression (P = 0.069, 95% CI: 1.004, 74.22), ECOG score at enrollment (P = 0.09, 95% CI: 0.96, 196.324), bulky disease (P = 0.219, 95% CI: 0.481, 136.605), prior lines of therapy (P = 0.071, 95% CI: 0.926, 60.626), and TP53 status (P = 0.499, 95% CI: 0.216, 19.04). For each predictive factor, we calculated regression coefficients, odds ratios (ORs), and 95% CIs. A score nomogram (Model 1) was developed using the regression coefficients of each predictive factor as weights, as shown in Figure 1A. Model 1 estimated the probabilities of poor CAR-T-cell therapy efficacy, with total scores ranging from 20 to 220, and demonstrated a C-index of 0.907 [Figure 1B]. The validity of the model was confirmed via the omnibus test (P <0.001), although no single predictive factor independently predicted the outcome (P >0.05). Given the complex nature of CAR-T-cell therapy efficacy, we sought to improve the identification of patients who would benefit most from CAR-T-cell therapy and might need additional treatments. To enhance the model’s specificity, we assigned equal scores to each factor. Consequently, Model 2 [Figure 1C–D] and Model 3 [Figure 1E–F] achieved 100% specificity in predicting efficacy and were able to independently predict CAR-T-cell therapy efficacy, with a P-value of 0.001 [Supplementary Table 5, https://links.lww.com/CM9/C191]. Although the sensitivity was slightly lower than that of Model 1, the C–indices for Model 2 (0.891) and Model 3 (0.891) were comparable to that of Model 1 (0.907). The calibration curves for the three models revealed only minor differences in Brier scores, with values of 11.5, 12.2, and 12.3 [Figure 1G]. The Hosmer–Lemeshow goodness-of-fit test indicated good calibration for all the models in the derivation cohort [Supplementary Table 6, https://links.lww.com/CM9/C191], with P-values of 0.921, 0.552, and 0.112, respectively.Figure 1: Development and calibration of prediction models. (A) The Model 1 nomogram predicts CAR-T-cell treatment efficacy by assigning weights on the basis of the regression coefficients of each predictive factor. In this nomogram, Row 1 represents the point scale, Row 7 represents the total points scale, and Row 8 represents the predicted probability of poor CAR-T-cell treatment efficacy. The probability of poor CAR-T-cell efficacy ranges from 0.20 to 0.95, corresponding to total scores between 20 and 220. (B) The ROC curve for Model 1. The AUC was 0.907 (95% CI: 0.821–0.993). The optimal cutoff value was 1.733 (sensitivity: 0.923, specificity: 0.774). (C) The Model 2 nomogram predicts CAR-T-cell therapy efficacy by identifying predictive factors on the basis of clinical expertise. It estimates the probability of poor CAR-T-cell efficacy within a range of 0.20–0.95. (D) The ROC curve for Model 2. The AUC was 0.891 (95% CI: 0.801–0.910). The optimal cutoff value is 3 risk factors (sensitivity: 1.000, specificity: 0.613). (E) The Model 3 nomogram predicts CAR-T-cell therapy efficacy by identifying predictive factors through a grading system informed by clinical experience. The probability of poor CAR-T-cell efficacy is estimated to be between 0.20 and 0.95. (F) The ROC curve for Model 3. The AUC was 0.891 (95% CI: 0.801–0.910). The optimal cutoff value is very high risk (sensitivity: 1.000, specificity: 0.613). (G) Calibration curves show the alignment between the predicted and observed efficacy of CAR-T-cell therapy. The x-axis denotes the predicted efficacy, whereas the y-axis represents the observed efficacy. The diagonal line represents the ideal model with perfect prediction. The curves for the three models illustrate their performance, with curves closer to the diagonal line indicating better predictive accuracy. Brier scores quantify the disparity between the predicted probabilities and actual outcomes, with lower values (closer to 0) reflecting greater predictive accuracy. The Brier scores are as follows: Model 1: 0.115 (95% CI: 0.055–0.175), Model 2: 0.122 (95% CI: 0.062–0.182), and Model 3: 0.123 (95% CI: 0.061–0.184). CAR: Chimeric antigen receptor; CI: Confidence interval.Validation in the validation cohort confirmed the high accuracy of the models, with C-indices of 0.805 for Model 2 and 0.800 for Model 3 [Supplementary Figure 2, https://links.lww.com/CM9/C191]. The Brier scores for Models 2 and 3 were close, at 17.5 and 17.8, respectively. The Hosmer–Lemeshow goodness-of-fit test indicated good calibration for both models [Supplementary Table 6, https://links.lww.com/CM9/C191], with P-values of 0.611 and 0.260. These results highlight the importance of considering clinical experience when evaluating the significance of predictive factors in the development of prediction models. Decision curve analysis (DCA) and clinical impact curve (CIC) analysis were employed to assess the clinical utility and effectiveness of the prediction models [Supplementary Figure 3, https://links.lww.com/CM9/C191]. In the derivation cohort, the clinical decision curve revealed that Model 3 demonstrated clinical utility across a risk threshold range of 0.15–1.00. Similarly, in the validation cohort, Model 3 was found to be clinically useful within a risk threshold range of 0.20–0.92. The CIC results from the validation cohort further support that Model 3 is the most applicable model in clinical practice. On the basis of the identified factors, the final index model incorporates the following risk factors: double expression, TP53 alteration, ECOG score ≥2, bulky disease ≥5 cm, and prior lines of therapy ≥4 [Supplementary Table 7, https://links.lww.com/CM9/C191]. Depending on the number of these risk factors present in DLBCL patients before lymphocyte clearance, the corresponding risk levels for CD19 CAR-T-cell therapy efficacy are categorized as follows: patients with 0 risk factors are classified as low risk (risk score <0.2), patients with 1 risk factor are categorized as medium risk (risk score ranging from 0.4 to 0.5), patients with 2 risk factors are designated as high risk (risk score ranging from 0.8 to 0.9), and patients with 3 or more risk factors are considered to have a very high risk (risk score ≥0.95). Using the index model, we conducted a thorough analysis of the effectiveness and prognosis across the entire patient cohort [Supplementary Figure 4, https://links.lww.com/CM9/C191]. The cohort was divided into four distinct risk groups, which highlighted significant differences in CAR-T-cell therapy efficacy, progression-free survival (PFS), and overall survival (OS). These differences were statistically significant (P <0.001), providing strong evidence of the impact of risk stratification on CAR-T-cell therapy outcomes and demonstrating the index model’s effectiveness in assessing prognosis. However, this study has several limitations. First, the sample size was relatively small. Second, while the choice of costimulatory domain might affect CAR-T-cell therapy efficacy, our study did not address this potential influence. Although we considered most factors that could impact prognosis in our baseline assessments, there may still be unexamined underlying factors that could affect outcomes. In conclusion, our research successfully developed and validated an index system for assessing the efficacy of CD19 CAR-T-cell therapy in patients with r/r DLBCL. This system is notable for its precision and ease of use, making it a practical tool for clinical applications. This study provides clinicians with valuable insights for decision-making, trial design, and patient prognosis discussions, thereby enhancing the real-world utility of CAR-T-cell therapy. Funding This work was supported by grants from the National Key R&D Program of China (No. 2021YFA1100800), National Natural Science Foundation of China (Nos. 81830004 and 82070168), Translational Research Grant of NCRCH (No. 2020ZKZC04), Shanghai Municipal Health Commission (No. 2020CXJQ02), and Cultivation project of National Natural Science Foundation of Shanghai Tongji Hospital (No. GJPY2332). Conflicts of interest None.