2.1 Patient selection for the REAL-TREND dataset

A total of 2778 consecutive patients with newly diagnosed DLBCL were enrolled between January 2010 and December 2015 at 8 centers in China, including Shanghai Rui Jin Hospital (Shanghai, Shanghai, China), Fujian Medical University Union Hospital (Fuzhou, Fujian, China), the First Affiliated Hospital of Medical School of Zhejiang University (Hangzhou, Zhejiang, China), Xinqiao Hospital (Chongqing, Chongqing, China), Wuhan Union Hospital (Wuhan, Hubei, China), Shanghai Ninth People's Hospital (Shanghai, Shanghai, China), Huashan Hospital Affiliated to Fudan University (Shanghai, Shanghai, China), and the First Affiliated Hospital of Soochow University (Suzhou, Jiangsu, China). Pathological diagnosis was established according to the World Health Organization (WHO) classification [13]. The data, including patient characteristics at diagnosis and determination of refractory status, treatment modality and response to each line of therapy, date of diagnosis, date of relapse or progression, and date of death or last follow-up, were collected from the electronic medical records by a team of research coordinators who had received the same training regarding how to perform a medical chart review following a standardized data collection protocol. A central electronic database was built for collecting the study data online. A data surveillance team performed monthly data quality control and gave feedback to all 8 centers.

When data collection was complete, a patient selection committee systematically identified patients treated with curative intent, which was defined as receiving rituximab-based immunochemotherapy in at least one adequate line of therapy. Standard R-CHOP-based regimens were used by all centers as first-line treatment, while regimens for second- or later-line therapy could be varied from centers. Patients with primary central nervous system (CNS) lymphoma, patients with primary mediastinal large B-cell lymphoma, and patients receiving CAR-T therapy or palliative care were excluded.

The covariates for the clinical characteristics and subgroup analyses within the refractory DLBCL dataset were age, gender, Eastern Cooperative Oncology Group performance status (ECOG PS), Ann Arbor stage, international prognostic index (IPI) score, treatment modalities (immunochemotherapy plus SCT, immunochemotherapy, and palliative treatment) after the determination of refractory disease and CNS relapse defined as a secondary CNS invasion that occurred either before or after the determination of refractory disease.

All study activities were approved by the Institutional Review Boards of all participating centers, and informed consent was obtained in accordance with the Declaration of Helsinki.

2.2 Validation of previous definitions of refractory DLBCL

Within the REAL-TREND dataset, 4 definitions of refractory DLBCL were applied: Definition 1 reported by Coiffier et al. [9], first-line SD/PD or relapse within 12 months after CR; Definition 2 reported by Cheson et al. [10], second-line/third-line SD/PD or relapse within 6 months after second-line/third-line CR/PR; Definition 3 reported by Hitz et al. [11], first-line SD/PD or relapse within 3 months after CR/PR; Definition 4 reported by the SCHOLAR-1 study [12], SD/PD to any line of therapy or relapse within 12 months after SCT. The definition for identifying patients with homogenous overall survival (OS) within subgroups will be used for further analysis of incidence, response rate, and prognostic factors of refractory DLBCL.

2.3 Incidence and outcomes of refractory DLBCL

The incidence of refractory DLBCL was defined as the number of patients with refractory DLBCL that occurred during a specified period of time. The outcomes were OS, overall response rate (ORR), and complete remission rate (CRR) after the determination of refractory DLBCL. The treatment response was evaluated according to the Lugano criteria [14]. Assessment of treatment response was evaluated by follow-up clinical, radiological, or laboratory studies, as determined by the clinicians [15]. Ambiguous cases were clarified by consulting institutional investigators and disagreements were resolved by consensus. ORR and CRR were calculated according to the best treatment response after the determination of refractory disease.

2.4 Statistical analysis

Categorical variables were summarized using frequencies and percentages. Continuous variables were summarized using median and ranges. Missing data in any variable analyzed were summarized using frequencies. OS was defined as the time interval from the date of event to the date of death or last follow-up. Patients were followed up for survival per institution standard procedures. Patients who were alive at the time of data collection were censored at the date of the last follow-up. In the REAL-TREND dataset, OS was estimated by the Kaplan-Meier method, and comparison of survival curves was performed using the log-rank test. Five-year cumulative incidence of refractory DLBCL was estimated from diagnosis to the determination of refractory status or last follow-up, as reported previously [16]. In the refractory DLBCL dataset, the heterogeneity of the ORR was first tested across the participant centers by Cochran's Q and I² statistics. If the variation in response rates across centers was significant, the ORR and CRR would be pooled using random-effect models. Otherwise, fix-effect models would be used. Survival data would be pooled directly if the log-rank test showed no significant difference across center, otherwise stratified by center. Univariable analysis of the covariates was performed using the Cox proportional hazards regression analysis. Statistically significant covariates in univariable analysis were included in the multivariable analysis. All analyses were performed using R version 3.4.3 (R Foundation, Vienna, Austria) and two-sided P values and 95% confidence intervals were reported. Significance was indicated at P <0.05.

3.1 Clinical characteristics and remission status of the REAL-TREND dataset

A total of 2778 DLBCL patients were enrolled in this study. Among them, 2342 patients were included in the REAL-TREND dataset. The clinical characteristics of the REAL-TREND cohort were summarized in Supplementary Table S1. The mean age at diagnosis was 54 years old (range = 14-93 years), with a 56.4% male composition. Overall, 412 (17.6%) patients had ECOG PS 2-4, 1276 (54.5%) had advanced disease (Ann Arbor stage III-IV), and 689 (29.4%) had IPI scores ≥3.

Within the REAL-TREND cohort, 2138 (91.3%) patients were evaluated by positron emission tomography/computed tomography (PET-CT) using the Lugano PET-CT criteria, and the remaining 204 (8.7%) patients by CT using Lugano CT criteria. No statistically significant difference in ORR and CRR was observed between patients evaluated by PET-CT and by CT (Supplementary Table S2).

As illustrated in Figure 1, 181 (7.7%) patients failed to respond to first-line treatment, including 33 (18.2%) with SD and 148 (81.8%) with PD. Additionally, 2161 (92.3%) patients responded to first-line treatment, including 1714 (79.3%) patients achieved CR and 447 (20.7%) patients achieved PR (CRR = 73%; ORR = 92%). Among them, 89 (4.1%) patients underwent consolidation SCT. Afterwards, 438 (20.3%) patients relapsed or progressed, including 16 (3.7%) who relapsed within 12 months after SCT. Among them, 119 (27.2%) patients who underwent palliative care were excluded, while the other 319 (72.8%) patients received second-line treatment. In the group that received second-line treatment, 66 (20.7%) of them achieved CR and 138 (43.3%) achieved PR (ORR = 64%; CRR = 21%; 42 [20.6%] of them underwent second-line salvage SCT). In comparison, 115 (36.1%) patients failed to respond, including 40 (34.8%) with SD and 75 (65.2%) with PD. Finally, excluding 17 (21.3%) patients who later underwent palliative care, the remaining 63 (78.8%) patients received third-line or later-line treatment. Among them, 9 (14.3%) patients achieved CR and 21 (33.3%) achieved PR (ORR = 48%; CRR = 14%), while 33 (52.4%) failed to respond, including 9 (27.3%) with SD and 24 (72.7%) with PD.

3.2 Validation of previous definitions of refractory DLBCL in the REAL-TREND dataset

As defined by Definition 1 (reported by Coiffier et al. [9]), internal heterogeneity was observed within subgroups, where patients with first-line SD/PD had significantly worse OS than those who relapsed within 12 months after CR (median OS: 7.1 months [95% confidence Interval {CI} = 5.9-9.0 months] vs. 22.8 months [95% CI = 19.6-36.6 months]; P < 0.001; Figure 2A). Similar results were shown according to Definition 2 (reported by Cheson et al. [10]; median OS: second-line/third-line SD/PD, 5.8 months [95% CI = 5.1-7.1 months]; relapsed within 6 months after second-line/third-line CR/PR, 14.1 months [95% CI = 10.9-72.8 months]; P < 0.001; Figure 2B) and Definition 3 (reported by Hitz et al. [11]; median OS: first-line SD/PD, 7.1 months [95% CI = 5.7-9.0 months]; relapsed within 3 months after CR/PR, 14.9 months [95% CI = 9.9-52.2 months], P < 0.001; Figure 2C). Only Definition 4 (reported in SCHOLAR-1 [12]) identified patients with homogenously inferior survival, and there was no significant difference of OS between patients with SD/PD to any line of therapy and patients who relapsed within 12 months after SCT (median OS = 5.9 months [95% CI = 5.5-7.1 months] vs. 5.9 months [95% CI = 3.2 months to not reached]; P = 0.564; Figure 2D). Thus, Definition 4 was applied to select refractory DLBCL patients for the following analysis.

3.3 Incidence and clinical features of refractory DLBCL

The incidence rate of refractory DLBCL in the REAL-TREND dataset was 14.9% (350/2342), including 181 (51.7%) primary refractory patients, 148 (42.3%) refractory patients to second or later-line therapy, and 21 (6.0%) relapsed patients within 12 months after SCT. The estimated 5-year cumulative incidence of refractory DLBCL was 20% (95% CI = 18%-22%). The clinical characteristics of the patients with refractory DLBCL at the determination of refractory disease are summarized in Table 1. Overall, 104 (29.7%) patients had ECOG PS 2-4, 197 (56.3%) had advanced disease (Ann Arbor stage III-IV), and 138 (39.4%) had IPI scores ≥3. After the determination of refractory disease, 245 (70.0%) of patients received salvage chemotherapy and 30 (8.6%) further underwent SCT.

3.4 Response rate and prognostic factors after the determination of refractory DLBCL

The variation in response rates across centers was significant, indicating that it was due to heterogeneity rather than chance (I² = 61.6%; Q = 19.1; P = 0.008). Then, the pooled ORR and CRR were estimated with random-effects models. Survival data were pooled directly using observed values since the log-rank test showed no significant difference across centers (P = 0.562). The pooled ORR after determination of refractory disease was 30% (95% CI = 22%-38%) and the CRR was 9% (95% CI = 4%-15%). Primary refractory patients had an ORR of 39% (95% CI = 26%-52%) and CRR of 16% (95% CI = 3%-28%). Patients’ refractory to second- or later-line therapy had an ORR of 18% (95% CI = 13%-24%) and CRR of 5% (95% CI = 2%-8%). Patients who relapsed within 12 months after SCT had an ORR of 24% (95% CI = 9%-39%) and CRR of 15% (95% CI = 2%-28%).

Refractory DLBCL patients had a median OS of 5.9 months (95% CI = 5.5-7.1 months) and 2-year OS rate of 16% (95% CI = 12%-20%). As shown in Table 2, univariable analysis identified the clinical characteristics of refractory diseases, including age (median OS: age >60 years old, 4.0 months [95% CI = 3.5-5.6 months]; age ≤60 years old, 7.1 months [95% CI = 6.1-9.0 months]), ECOG PS (median OS: ECOG PS 0-1, 7.3 months [95% CI = 6.2-9.3 months]; ECOG PS 2-4, 3.5 months [95% CI = 2.5-4.5 months]), Ann Arbor stage (median OS: stage I-II, 7.5 months [95% CI = 6.3-11.0 months]; stage III-IV, 4.6 months [95% CI = 3.8-6.1 months]), IPI score (median OS: IPI 0-1, 8.1 months [95% CI = 6.3-11.8 months]; IPI 2, 7.0 months [95% CI = 6.0-11.8 months]; IPI 3, 4.2 months [95% CI = 3.6-7.4 months]; IPI 4-5, 2.5 months [95% CI = 1.3-3.7 months]; Figure 3A) and CNS relapse (median OS: with CNS relapse, 4.5 months [95% CI = 3.6-7.1 months]; without CNS relapse, 6.1 months [95% CI = 5.5-7.4 months]; Figure 3B) had significant impact on OS. After the determination of refractory disease, 215 (61.4%) patients received immunochemotherapy, 30 (8.6%) patients received immunochemotherapy plus salvage SCT, and 105 (30.0%) patients received palliative care. According to treatment modalities, patients treated with immunochemotherapy plus salvage SCT had longer survival time (median OS: 14.1 months [95% CI = 11.1 months to not reached]) than those who received immunochemotherapy (median OS: 6.7 months [95% CI = 5.8-7.8 months]) or palliative treatment (median OS: 2.8 months [95% CI = 1.8-3.7 months]; P < 0.001; Figure 3C). The best response achieved after the determination of refractory disease also had significant impact on OS (P < 0.001; Figure 3D). Patients who achieved CR had a median OS of 50.4 (95% CI = 39.2 months to not reached), remarkably superior to patients who achieved PR (median OS: 12.2 months [95% CI = 11.3-18.5 months]) or patients with SD/PD (median OS: 4.0 months [95% CI = 3.6-5.0 months]).

Since age, ECOG PS and Ann Arbor stage were factors integrated into IPI scores, they were not included in multivariable analysis. Multivariate analysis showed that IPI score 4-5 (Hazard ratio [HR] = 2.22; 95% CI = 1.47-3.35), CNS relapse (HR = 1.43; 95% CI = 1.04-1.97) and best response after the determination of refractory disease (PR: HR = 2.68; 95% CI = 1.42-5.03; SD/PD: HR = 5.97, 95% CI = 3.21-11.11) were independent unfavorable prognostic factors for OS (Table 2).