Persistent B-cell aplasia predicts late severe infections after anti-CD19 CAR T-cell treatment in B-cell lymphoma : A methodological approach to long-term safety reporting

Abstract

Abstract Introduction B-cell aplasia (BCA) is an on-target, off-tumor effect of CD19 CAR T-cells that correlates with their persistence but increases infection risk. Data on B-cell recovery (BCR) incidence, determinants, and impact on infections following CAR T-cell treatment are poorly reported. Many long-term studies of CAR T-cells have relied on Kaplan–Meier estimates and treated time-dependent exposures as static covariates, overlooking competing risks and immortal-time bias. Methods ​​We retrospectively analysed 208 adults who received CD19 CAR-T cells between Sep 2018 and Dec 2023 at a single center. BCR was defined as ≥0.01G/L B-cells on flow cytometry assay. From a day-100 landmark, cumulative incidence of BCR, non-relapse mortality (NRM) and severe infection (grade ≥3 or requiring hospitalization) were estimated, treating death or relapse as competing risks. To assess the time-dependent impact of BCR on severe infection risk we performed competing-risk analyses with annual landmarks. Predictors of BCR were identified by the Fine-Gray model.​ To illustrate methodological pitfalls, we also performed standard Kaplan–Meier analyses, non-landmarked competive-risk models and standard Cox models. Results Median follow-up was 36.1 months, median age was 64 years (range 19–78) and 61.2% were male. Indications were mostly diffuse/transformed large B-cell lymphomas (81.7%). CAR T products included axi-cel (57.2%), tisa-cel (23.1%), liso-cel (13.9%) and brexu-cel (5.3%). Prior to lymphodepletion (preLD), median LDH was 272 U/L and median MCV 93.0 fL. Median CAR-HEMATOTOX score was 2, with 41.8% high risk (≥2). Patients received a median of two prior therapy lines; 9.1% had prior ASCT. CRS occurred in 86.6% (11.1% ≥grade 3), ICANS in 28.4% (7.7% ≥grade 3). At 1, 3 and 5 years post-infusion, BCR among patients in complete remission at day 100 was 22.2% (95%CI 15.6–29.5), 39.9% (31.3–48.3) and 45.7% (35.9–54.9) respectively. Comparatively, Kaplan–Meier overestimated 5-year BCR at 58.0% (43.9–68.6) by treating competing events as simple censoring, while competing-risk models without landmarks underestimated it at 30.8% (23.8–38.0), due to high early mortality and relapse. Similarly, the day-100 landmarked competing-risk model estimated NRM at 1.48% (0.3–4.8) at 1 year and 5.56% (2.2–11.2) at 3 and 5 years, compared with a 5-year estimate of 7.5% (3.3–16.4) by Kaplan–Meier and 3.75% (1.5–7.6) by the non-landmarked model. Severe infections were common: day-100 landmarked cumulative incidences were 14.9% (9.5–21.5), 41.7% (32.1–51.0), and 50.5% (38.6–61.2) at 1, 3, and 5 years post-infusion, respectively. Annual landmark analyses, allowing multiple events per patient, showed similar 5-year estimates: 54.9% (37.3–69.4) from day 365, 47.2% (27.0–65.0) from day 730, and 34.8% (15.2–55.3) from day 1095. Annual severe infection rates were consistently higher in patients with persistent BCA compared to those with BCR: 23.0% (16.0–31.0) vs 0% in year 1, 23.0% (13.0–34.0) vs 5.9% (0.35–24.0) in year 2, 36.0% (21.0–52.0) vs 5.6% (0.34–23.0) in year 3, 29.0% (11.0–50.0) vs 0% in year 4, and 19.0% (2.4–47.0) vs 13.0% (0.48–44.0) in year 5. Over the five-year period, severe infections were 7.4 times more frequent in the BCA group (incidence rate: 35.7 vs. 4.8 per 100 patient-years; 95%CI 2.4–23.2, two-sample Poisson test: p<0.001). In univariate analyses, using Fine-Gray model, older age (HR 0.98 per year; 95%CI 0.96–1.00; p = 0.044), ≥3 prior therapy lines (0.58; 0.34–0.98; p = 0.041), high CAR-HEMATOTOX score (HR 0.40; 0.22–0.73; p = 0.003), preLD MCV (0.96; 0.92–1.00; p = 0.030), preLD LDH (1.00; 0.99–1.00; p = 0.029), CRS duration (0.93 per day; 0.86–1.00; p = 0.036), use of tocilizumab (0.59; 0.35–0.99; p = 0.047) or corticosteroids (0.42; 0.22–0.79; p = 0.007) were significantly associated with lower BCR. In contrast, a conventional Cox model that did not account for competing risks identified sex—but not age, LDH, or tocilizumab use—as significant predictors, illustrating how inappropriate methods can yield misleading conclusions. Conclusion In this real-world cohort, sustained BCA was observed in over half of complete responders five years after CD19 CAR T-cell infusion, resulting in a 7.4-fold higher risk of severe infections. Using BCA as a case-study, we demonstrate that rigorous statistical methods and prolonged follow-up are essential to deliver reliable, unbiased assessments of the long-term safety of CAR-T cell therapy.