A total of 613 DLBCL patients treated with rituximab-containing regimens between January 2002 and March 2012 (a dataset identical to that presented in a previous study) [11] were enrolled. In contrast to the previous study, cases with DLBCL variants or those showing DLBCL transformed from low-grade B-cell lymphomas were not excluded. However, of these, five cases with insufficient clinical information, particularly factors related to the IPI or dose of chemotherapeutic agents, were excluded. Finally, 608 cases were enrolled in the study. The medical records of all enrolled patients were reviewed to obtain data related to birth date, sex, Eastern Cooperative Oncology Group performance score, presence B symptoms, involved anatomic site, initial serum lactate dehydrogenase (LDH) titer, and Hans cell-of-origin subgroup [12] evaluated at the time of initial diagnosis. Cases were staged according to Ann Arbor stage [13] and classified according to age-adjusted IPI risk group [3], which could serve as a prognostic group free from the patient age. Patient age was re-calculated based on birth date. The relative dose intensity (RDI) of each chemotherapeutic agent (cyclophosphamide, doxorubicin, or vincristine) was calculated by dividing the actual dose over the actual treatment duration by the standard dose over the standard treatment duration [14]. The average RDI (ARDI) of each patient was then calculated; patients with an ARDI < 85% were assigned as the low-dose intensity group [15].

The impact of age at the time of diagnosis on survival outcome was estimated using scatterplot smoothing curves between the Martingale residual of the null Cox proportional hazard (PH) model and age (as a continuous variable) [16]. The Cox PH model fitted with the age-adjusted IPI and ARDI < 85% variables was used as a null model for the multivariate setting. These estimates were confirmed using a Cox PH model fitted with the transformed age variable using the restricted cubic spline function [17]. The prognostic significance of age variables categorized according to known cutoff values (such as the conventional IPI [3], NCCN-IPI [7], and GELTAMO-IPI [8]) (Table 1) were evaluated using both univariate Kaplan-Meier survival analysis and multivariate Cox PH models. The statistical procedures described above were also carried out for the DLBCL-NOS (not otherwise indicated) subgroup. All statistical analysis was performed using R ver. 3.6.1 (R foundation for statistical computing, Vienna, Austria). Two-sided p-values < 0.05 were considered statistically significant.

The clinical characteristics of the enrolled patients are summarized in Table 2. The mean age was 53.3 years. About one third (32.7%) of patients were aged > 60 years, and 15 (2.5%) were elderly (> 75 years). Only six patients (1.0%) were aged > 80 years. More than 90% (91.9%) of patients were diagnosed with DLBCL-NOS histologically; 26 cases (4.1%) had variants of large B-cell lymphoma, including Epstein-Barr virus–positive large B-cell lymphoma (12 cases), T-cell/histiocyte-rich B-cell lymphoma (seven cases), primary mediastinal B-cell lymphoma (four cases), primary cutaneous large B-cell lymphoma (two cases), and lymphomatoid granulomatosis (one case). Another 23 cases were DLBCL with concurrent low-grade B-cell lymphomas, including extranodal marginal zone B-cell lymphoma (14 cases), high-grade (four cases) or low-grade (three cases) follicular lymphoma, and marginal zone B-cell lymphoma (two cases). About half (49.3%) of cases had advanced disease (Ann Arbor stage III–IV) at the time of the diagnosis, and 47.2% of cases had elevated serum LDH titers. Multiple extranodal site involvement was identified in about one third of patients (31.4%). More than half of cases (50.7%) were classified as low IPI risk. About two thirds of the tumors were classified as non–germinal center B-cell-like subgroup by the Hans immunohistochemical algorithm in available patients (n=402). Almost all patients were treated with an R-CHOP (rituximab with cyclophosphamide, doxorubicin, vincristine and prednisone) regimen; only three patients (0.5%) were treated with an R-CVP (rituximab with cyclophosphamide, vincristine and prednisone) regimen. Due to dose reduction during treatment, about 30% of patients received chemotherapeutic treatment with low ARDI levels (< 85%). Nevertheless, 88.8% of patients achieved complete remission after cessation of initial treatment, and 83.6% attained 3-year survival. On univariate survival analyses, all of the components of the conventional IPI system, presence of B symptoms, and low ARDI level showed statistically significance. However, the Hans cell-of-origin subgroup did not significantly correlate with survival outcome (p > 0.05) (Table 1).

First, we examined the relationship between the age and survival outcome using the Martingale residuals from both null (no predictor included) and multivariate Cox PH models (Fig. 1A and B). Age-adjusted IPI risk, presence of B symptom, and categorized ARDI variables (less or more than 85%) were included as predictor variables in the multivariate Cox PH models. Scatterplot smoothing lines between the Martingale residuals of the null Cox models and age variables were not significantly different between patients aged up to 62 years. However, when age exceeded 62 years, the Martingale residuals started to increase rapidly and exponentially; which is apparent in the models for overall survival (OS) and progression-free survival (PFS). These relationships were maintained when data was examined using a multivariate Cox model (Fig. 1C and D) and the DLBCL-NOS group (S1A-D Fig.). These findings suggest that the relative hazard increases only for patients aged > 62 years.

The non-linear relationship between the age variable and relative hazard was also evaluated using variable transformation models. The shape of the predicted relative hazard curves (Fig. 1E and F) calculated from the selected models was similar to that of the smoothened curve of the Martingale residual plot (Fig. 1A and B). Also, these relationships were reproduced identically in subgroup analysis of the DLB-CL-NOS group (S1E and S1F Fig.). Therefore, it can be concluded that the age variable and survival outcome are not linearly correlated; the relative hazard was not significantly different for those aged younger than 62 years, but increased rapidly when the age variable exceeded 62 years.

Finally, we examined the prognostic significance of three known cutoff criteria for age variables: the conventional IPI, the NCCN-IPI, and the GELTAMO-IPI (Table 1). Although all univariate survival analyses of age variables categorized according to the conventional IPI, the NCCN-IPI, and the GELTAMO-IPI schemes were statistically significant both for OS and PFS, only the elderly patient groups (age > 75 or ≥ 80 years) showed markedly inferior survival outcomes (Table 1, Fig. 2A–F), which was also observed in the DLBCL-NOS subgroup (data not shown). In line with this trend, only the patients aged > 75, classified according to the NCCN-IPI criterion, displayed an independently poor survival outcome in multivariate analysis adjusted for age-adjusted IPI and ARDI variables (Table 3). Age variables divided by the cutoffs of conventional IPI and 62 years were non-significant in multivariate analyses; nevertheless, some of the significance levels were borderline (Table 4, S2 Table). In addition, multivariate analysis revealed that age > 80 years, classified according to the GELTAMO-IPI criterion, was statistically significant only for OS but significantly borderline for PFS, which could be attributable to the small number of the patients with > 80 years old (n=6). Interestingly, age 65 to < 80 was statistically significant for PFS, but was significantly borderline for OS (Table 5). These results of multivariate analyses were also recapitulated when the DLBCL-NOS subgroup was subjected to the same analysis (S3–S5 Tables), except that patients > 80 years old were borderline significant for PFS (p=0.086) (S4 Table). Further analyses of the Hans cell-of-origin-available subgroup revealed a significantly poor OS outcome only for patients > 75 years old (p=0.045) (data not shown).