We conducted a retrospective analysis on patients diagnosed with early-stage FL, grades 1–3A, at the National Cancer Center, Cancer Hospital, Chinese Academy of Medical Sciences, Peking Union Medical College between 2000 and 2020. The inclusion criteria were: (1) histologically confirmed grade 1–3A FL (Alaggio et al., 2022); (2) Ann Arbor stage I–II disease; and (3) age ≥18 years at the time of diagnosis. The exclusion criteria were: (1) Ann Arbor stage III–IV disease; (2) incomplete data on primary treatment; and (3) loss to follow-up within 3 months after diagnosis. The study included 107 patients with early-stage FL. Bulky disease was defined as a nodal or extranodal mass with the largest diameter exceeding 5 cm. Because the study period spanned 2 decades, staging procedures evolved over time. PET/CT was performed in 31.8% of patients overall, with lower utilization in the earlier years and increasing adoption after 2010. All patients underwent bone marrow biopsy as part of baseline staging. Marrow flow cytometry was not routinely performed during the study period. For patients without PET/CT, staging was determined using contrast-enhanced CT, clinical evaluation, and bone marrow biopsy findings. Ann Arbor stage I–II was assigned based on the best available imaging and pathological information at diagnosis. This study was conducted in accordance with the Declaration of Helsinki and approved by the PUMC/CAMS independent ethics committee (no. 25/164-5110). Given the retrospective nature of the study and the use of de-identified data, the requirement for informed consent was waived by the ethics committee.

The first-line treatment strategy included RT with or without chemoimmunotherapy or chemotherapy (RT ± CIT, n = 57), chemoimmunotherapy or chemotherapy (CIT, n = 34), and observation (n = 16). The RT ± CIT category included patients treated with radiotherapy alone or with radiotherapy combined with chemotherapy or chemoimmunotherapy, whereas the CIT group consisted of patients receiving chemotherapy or chemoimmunotherapy without radiotherapy. Observation was defined as a documented decision in the medical record to defer treatment for at least 3 months from the date of diagnosis, with no treatment during this period. The treatment decision was at the discretion of the physician or the patient’s choice. Treatment response was primarily assessed using the Lugano criteria (Cheson et al., 2014).

The RT ± CIT group received involved-field or involved-site RT. The median dose was 30 Gy (interquartile range [IQR], 30–40 Gy), at 1.8–2.0 Gy per fraction. Most patients received intensity-modulated or three-dimensional conformal RT. The specific dose fractionation schedules and RT techniques are shown in Supplementary Table S1.

Systemic therapy regimens reflected real-world management of follicular lymphoma across the two-decade study period. For analytical purposes, systemic therapy was defined as the administration of chemotherapy and/or immunotherapy without radiotherapy, or as part of combined-modality treatment combined with radiotherapy. Regimens were classified according to their dominant components and included anthracycline-based chemoimmunotherapy (e.g., CHOP or R-CHOP–based), chemotherapy alone, rituximab-based immunotherapy without cytotoxic agents, and other less commonly used regimens such as lenalidomide-containing or alkylator-based therapies. Treatment selection was individualized based on patient age, comorbidities, performance status, and disease characteristics, consistent with real-world practice during the study period.

To more comprehensively evaluate differences across therapeutic strategies, we additionally applied a four-category grouping—radiotherapy alone (RT), systemic therapy alone (CIT), combined systemic therapy plus radiotherapy (CIT + RT), and observation—in exploratory analyses. This treatment-modality–based grouping provided clinical granularity for comparing specific approaches, while the primary three-group framework supported coherent statistical modeling and interpretation. The exploratory categorization was not used for primary endpoints but served to complement and contextualize treatment-related findings.

The primary endpoints of this study were overall survival (OS), lymphoma-specific survival (LSS), and PFS. OS was defined as the time from diagnosis to death from any cause. LSS was defined as the time from diagnosis to death due to FL. PFS was defined as the time from diagnosis to disease progression, relapse, or death from any cause. Non-lymphoma-related death (non-LRD) was defined as death due to causes other than FL, including cardiovascular disease, secondary malignancies, and other non-lymphoma conditions. Disease failure was defined as the recurrence or progression of FL after initial treatment. Progression of disease within 24 months (POD24) was defined as disease progression, relapse, transformation, or death from any cause occurring within 24 months of the initiation of first active anti-lymphoma therapy. For patients who received immediate treatment after diagnosis, the 24-month interval was measured from the start of first-line therapy. For patients initially managed with observation, the interval was measured from the initiation of their first systemic therapy or radiotherapy. Patients who remained untreated and alive beyond 24 months from diagnosis were classified as non-POD24, consistent with definitions used in prior early-stage and real-world FL studies. Locoregional failure (LRF) was defined as the recurrence or progression of FL within the original tumor site or regional lymph nodes. Systemic recurrence of FL was defined as systemic disease progression or recurrence at sites distinct from the originally involved sites, including distant lymph nodes or extranodal locations.

Categorical variables were compared using the chi-squared or Fisher’s exact test. Survival outcomes were estimated using the Kaplan-Meier method, and survival differences were compared by the log-rank test. The Fine-Gray competing risk model was applied to evaluate the cumulative incidence of lymphoma-specific mortality and non-LRD, accounting for competing risks. Multivariable analysis was performed using Cox proportional hazards models to identify independent prognostic factors for OS and PFS, while competing risk regression was used for LSS and non-LRD. LASSO regression with ten-fold cross-validation was performed using the glmnet package in R to identify prognostic variables associated with overall survival from 24 clinical and hematologic factors (Sex, Age, Year, Stage, Primary site, Grade, Ki-67, B symptoms, FLIPI1, FLIPI2, Bulky disease, Lymph involvement, ECOG, LDH, Leukopenia, Anemia, PLT, Lymphopenia, NLR, RT, PFS24, CT, Surgery, and CD20). Variables with nonzero coefficients at the optimal penalty parameter (λ) were selected for model construction. Three survival prediction models were subsequently developed, including a Cox proportional hazards model, a Gradient Boosting Machine (GBM) model, and a Random Survival Forest (RSF) model. Model performance was evaluated using the concordance index (C-index) and time-dependent receiver operating characteristic (ROC) curves at 3-, 5-, and 10-year time horizons. The RSF model demonstrated the best predictive discrimination and was further interpreted using SHAP (SHapley Additive exPlanations) analysis to quantify the global contribution and relative importance of each feature across different prediction horizons. Statistical analysis was conducted using R software (version 4.3.3), and a P-value <0.05 was considered statistically significant.

The baseline patient characteristics are shown in Table 1. The median age at diagnosis was 53 years (IQR, 44–62), with 30.8% of patients under the age of 60. Most patients (98.1%) had an Eastern Cooperative Oncology Group (ECOG) performance status of 0–1, and 57.9% presented with a stage II disease. Histologically, 61.7% of patients had grade 1–2 FL, 29.0% had grade 3A disease, while grade information was not specified for the remaining 9.3% of patients. Bulky disease was observed in 11.2% of patients. Hematologic abnormalities were observed at diagnosis in some patients, including leukopenia (13.1%), anemia (2.8%), lymphopenia (4.7%), and thrombocytopenia (1.9%). Elevated lactate dehydrogenase (LDH ≥245 U/L) was found in 8.4% of patients, and 36.4% had elevated β2-microglobulin (≥3 mg/L). The neutrophil-to-lymphocyte ratio (NLR) was <1.8 in 30.8% of patients. Furthermore, 63.5% of patients were diagnosed between 2015 and 2020, which suggests a recent increase in early-stage FL diagnoses.

Among the 72 patients who received systemic therapy (either alone or in combination with radiotherapy), treatment patterns reflected contemporary real-world practice. R-CHOP–based chemoimmunotherapy was the most common regimen (49 patients, 68.1%), followed by CHOP alone (14 patients, 19.4%). Rituximab monotherapy was administered in 7 patients (9.7%), and lenalidomide-based regimens, including R²-like combinations, were used in 4 patients (5.6%).

Twenty-three patients died during a median follow-up of 86 months (IQR, 59–144), including LRD in 10 patients, and non-LRD in 13. Using non-LRD as a competing risk, the cumulative incidence of LRD slightly increased from 6.5% at 5 years (95% confidence interval [CI]: 2.9%–12.3%) to 10.7% at 10 years (95% CI: 5.4%–18.2%). Conversely, the cumulative incidence of non-LRD markedly increased from 4.8% at 5 years (95% CI: 1.8%–10.1%) to 20.5% at 10 years (95% CI: 10.3%–33.3%; Figure 1A). Among the 13 non-LRD events, cardiovascular disease was the most common cause (6/13, 46.2%), followed by other non–cancer-related medical conditions (4/13, 30.8%) and secondary malignant neoplasms (3/13, 23.1%; rectal, breast, and laryngeal cancers), as shown in Figure 1B.

Twenty-one patients (19.6%) developed disease relapse by the last follow-up. Fourteen patients developed LRF only (five with local failure only, six with regional lymph node failure only, and three with both), six had systemic failure only, and one had both LRF and systemic failures. The 5-year and 10-year cumulative incidences were 14.5% and 16.2% for LRF, 3.1% and 4.8% for systemic failure, and 16.3% and 19.5% for overall failures (Figure 2A).

In most patients, failure occurred early in the course of the disease. The disease had relapsed within 5 years in 16 patients and after more than 5 years in five patients. No significant differences in failure rates were observed among the treatment modality groups. The 5-year cumulative incidences of overall failure were 11.1% for RT ± CIT, 27.5% for CIT (hazard ratio [HR], 1.90; 95% CI: 0.96–3.77; P = 0.064; Figure 2B), and 12.5% for observation (HR, 1.05; 95% CI: 0.35–3.15; P = 0.933; Figure 2B). The corresponding 10-year cumulative incidences were 18.5%, 27.5%, and unreached.

For LRF, the 5-year cumulative incidences were 9.3% for RT ± CIT, 24.5% for CIT (HR, 2.60, 95% CI: 0.82–8.20; P = 0.103; Figure 2C), and 12.5% for observation (HR, 2.18; 95% CI: 0.52–9.13; P = 0.286; Figure 2C). The respective 10-year rates were 9.3%, 24.5%, and unreached. Systemic failure was rare across all groups, with 5-year cumulative incidences of 2.0% for RT ± CIT, 4.0% for CIT (HR, 0.58; 95% CI: 0.06–5.36; P = 0.634; Figure 2D), and 6.3% for observation (HR, 1.63; 95% CI: 0.16–16.75; P = 0.681; Figure 2D). The corresponding 10-year incidences were 10.1%, 4.0%, and unreached.

Among the patients in the RT ± CIT group, ten experienced failures, including five in-field failures (within the composite 95% isodose volume) and five out-of-field failures (outside the irradiated volume). The time to in-field recurrences ranged between 9.9 and 50.9 months. Out-of-field failures primarily involved nodal relapses outside the irradiated field. A detailed breakdown of in-field recurrence patterns, including irradiated volumes and doses, is presented in Supplementary Table S2. To further evaluate whether treatment granularity affected failure patterns, we conducted an exploratory four-group analysis separating RT alone, CIT, CIT + RT, and observation (Supplementary Figure S1). Although not powered for formal statistical comparison due to limited event numbers within individual strata, the cumulative incidence curves for overall failure, locoregional failure, and systemic failure showed broadly similar trajectories across all four categories. No clinically meaningful divergence was observed.

The 5- and 10-year rates for the entire cohort were 88.7% and 68.8% for OS (95% CI: 82.8%–94.9%), 93.3% and 89.9% for LSS (95% CI: 88.2%–94.7%; Figure 3A), and 76.0% and 57.2% for PFS (95% CI: 68.2%–84.7%). POD24 occurred in 6.5% of patients.

There were no statistically significant differences in survival between treatment groups. The 5-year OS, PFS, and LSS rates were 87.5%, 80.4%, and 92.9% in the RT ± CIT group; 85.3% (HR, 1.59; 95% CI: 0.66–3.81; P = 0.295; Figure 3B), 63.3% (HR, 1.90; 95% CI: 0.96–3.77; P = 0.064; Figure 3C), and 90.7% (HR, 1.21; 95% CI: 0.34–4.28; P = 0.771; Figure 3D) in the CIT group; and 100% (HR, 0.41; 95% CI: 0.05–3.20; P = 0.395; Figure 3B), 87.5% (HR, 1.05; 95% CI: 0.35–3.15; P = 0.933; Figure 3C), and 100% (HR, not available; P = 0.988; Figure 3D) in the observation group.

To complement the primary treatment-group comparison, an exploratory four-category survival analysis was also performed (RT alone, systemic therapy alone, systemic therapy plus RT, and observation; Supplementary Figure S2). As expected with small sample sizes and few survival events, the curves did not demonstrate statistically or clinically significant separation. The overall patterns remained consistent with the main analysis, indicating that the absence of major survival differences is not dependent on the specific grouping scheme.

We analyzed failure within a competing risk framework, considering non-LRD as a competing event. As shown in (Supplementary Figure S3A), the cumulative incidence of failure increased steadily over time, surpassing that of non-LRD after 24 months. Stratification by treatment modality revealed no statistically significant differences in failure risk (Supplementary Figure S3B) or the risk of non-LRD (Supplementary Figure S3C) among groups.

Assessment of OS by failure incidence showed 5-year OS rates of 89.5% without failure, 83.9% with LRF and 87.5% with systemic failure, with no significant difference between groups (Supplementary Figure S4A). Patients who experienced POD24 exhibited significantly poorer OS than those those who did not experience POD24 (P = 0.011; Supplementary Figure S4B). Cumulative incidence analysis further confirmed that patients who experienced POD24 had a significantly higher incidence of treatment failure and death than those who did not experience POD24, underscoring the critical prognostic impact of early disease progression (Table 2).

Univariate Cox regression analysis identified several significant prognostic factors for mortality in early-stage FL (Table 2). These included age over 60 years (P < 0.001), male sex (P = 0.007), ECOG score ≥2 (P = 0.007), NLR ≥1.8 (P = 0.017), anemia (P = 0.023), leukopenia (P = 0.021), thrombocytopenia (P = 0.005), and lymphopenia (P = 0.007). Elevated LDH (P = 0.068) and β2-microglobulin (P = 0.235) lacked statistical significance. Male sex (P = 0.026) and NLR <1.8 (P = 0.025) were significantly associated with an increased risk of failure.

Kaplan-Meier survival analysis showed that, compared to patients with NLR ≥1.8, those with NLR <1.8 exhibited significantly worse 5-year OS (81.0% vs. 91.9%; HR, 2.65; 95% CI: 1.16–6.06; P = 0.021; Figure 4A) and PFS (62.5% vs. 82.1%; HR, 2.76; 95% CI: 1.44–5.31; P = 0.002; Figure 4B).

To explore prognostic factors, LASSO regression was used to select survival-associated clinical and laboratory variables (Supplementary Figure S5A,B). These variables were incorporated into three prediction models (Cox, GBM, RSF). The RSF model showed the best discrimination, with a C-index of 0.943 compared with 0.888 for the Cox model and 0.734 for the GBM model. Time-dependent ROC curves consistently demonstrated superior AUCs for the RSF model at 3, 5, and 10 years (Supplementary Figure S5C–E). Model interpretability was assessed using SHAP analysis. Age was the most influential predictor, followed by lymphopenia and FLIPI1, indicating the importance of both demographic and immune-related factors. SHAP heatmaps demonstrated stable feature importance across time horizons (Supplementary Figure S5F,G).