Mwangi, Raphael

Background: Perinatal and early-life factors may influence non-Hodgkin lymphoma (NHL) risk, but study results have been mixed and exposure data for childhood ages are limited. Herein, we investigated associations of these exposures with risk of NHL and common subtypes. Methods: This case–control study included 2280 NHL cases and 2253 controls, enrolled from 2002 to 2014 at the Mayo Clinic Rochester. Self-reported perinatal and early-life exposures included maternal age at birth, birth order, birthweight, time breastfed, height and weight at ages 7, 12, and 18 years relative to peers, and age and weight when growth ceased. We used logistic regression to estimate odds ratios (ORs) and 95% confidence intervals (CIs), adjusting for design variables and potential confounders. Linear trend tests were performed for ordinal variables, and heterogeneity tests were performed to evaluate whether associations varied across four NHL subtypes. Results: After multivariable adjustment, greater birth weight (OR = 1.15 for quartile 4 vs. 1; p-trend = 0.04), weight at age 7 relative to peers (compared to average, OR = 1.05 for heavy and OR = 0.87 for thin; p-trend = 0.002), and weight when growth ceased (OR = 1.30 for quartile 4 vs. 1; p-trend < 0.001) were associated with increased NHL risk. An inverse association was observed for breastfeeding duration with NHL risk (OR = 0.77 for >6 months vs. never; 95% CI 0.61–0.97). Associations did not significantly vary by major NHL subtype (p-heterogeneity > 0.05). Other perinatal and early-life exposures were not associated with NHL risk. Conclusions: Greater body weight at birth and through childhood may be associated with increased NHL risk, potentially extending the life-course perspective on adiposity and lymphomagenesis, with implications for prevention.

Abstract Background: Although some vaccines seem to be associated with a lower risk of lymphoma, their impact on outcomes is largely unknown. Methods: Data from a large prospective lymphoma cohort study were utilized to estimate associations of self-reported history of vaccination against hepatitis A, hepatitis B, influenza, and yellow fever with event-free survival (EFS) and overall survival (OS) using Cox proportional hazard models. Results: None of the vaccines were associated with EFS [hazard ratio (HR) 0.77–1] or OS (HRs 0.96–1.04). Similarly, there were no significant associations between the vaccines and outcomes for each lymphoma subtype. Conclusions: Prior vaccination against selected viruses was not associated with EFS or OS among patients with lymphoma. Impact: We did not find evidence for an impact of prior vaccination on lymphoma prognosis.

Abstract Prognostic indices such as the CLL International Prognostic Index (CLL-IPI) and the International Prognostic Score for Early-stage CLL (IPS-E) can predict time to first treatment (TTFT) in patients with early-stage CLL. Their ability to predict risk of infection - a leading contributor to morbidity and mortality in CLL - remains uncertain. The CLL-TIM is a machine-learning model developed in Europe that integrates clinical, laboratory, and infection-related data to predict TTFT or infection risk within 2-years of diagnosis, Receiver Operating Characteristic - Area Under the Curve (ROC-AUC) 0.74; Agius et al., Nat Comm 11, 363 2020. We conducted the first validation of CLL-TIM in a US cohort of newly diagnosed CLL patients and compared its performance with the CLL-IPI and IPS-E indices. Adults with newly diagnosed CLL (2000-2020) were identified through the Rochester Epidemiology Project using International Classification of Disease (ICD) codes; all diagnoses were confirmed. We replicated the variable selection from the original CLL-TIM model and applied it to our cohort. The primary endpoint was a 2-year composite of either TTFT or incident infection (defined as having blood cultures drawn). Model performance was evaluated using ROC-AUC consistent with the original CLL-TIM methods. CLL-IPI and IPS-E performance was assessed using time-dependent ROC-AUC; pairwise differences in discrimination were tested using DeLong. We identified 454 CLL patients with a median age of 72 years [range, 30-97], 166 (37%) were female, Rai Stage was 0-II in 399 (92%) patients. IGHV genes were unmutated in 130/305 (42%) patients and TP53 disruption (either del17p by FISH or TP53 mutation) was present in 24/329 (8%) patients. At 2-years, 55 (12%) patients received CLL therapy, and 56 (12%) patients had an infection. The 2-year composite endpoint was observed in 94 (21%) patients. The ROC-AUC was 0.74 (95% CI 0.68-0.80) for CLL-TIM, 0.69 for CLL-IPI (95% CI 0.63-0.75, p=0.11 compared to CLL-TIM) , and 0.68 for IPS-E (95% CI 0.62-0.74, p=0.02 compared to CLL-TIM). We also evaluated the ROC-AUC using all the prognostic models for each individual endpoints. The corresponding ROC-AUC for TTFT at 2 years were 0.79, 0.74, and 0.76 for CLL-TIM, CLL-IPI, and IPS-E, respectively; and for infection at 2 years were 0.68, 0.57, and 0.52, respectively. This is the first validation of the CLL-TIM in a US-based cohort of newly diagnosed CLL patients. Although the CLL-TIM model exceeded the minimally meaningful discrimination threshold (AUC &gt;0.7) for individual patient-level risk prediction, there was no statistical difference in AUC between CLL-TIM and CLL-IPI, suggesting either model can predict the composite endpoint at 2 years. Most of the observed discrimination was driven by prediction of TTFT rather than infection, underscoring the need to develop more infection-specific risk models for newly diagnosed CLL. Citation Format: Raphael Mwangi, Tait D. Shanafelt, Soren Basnet, Emily L. West, Owen Keegan, Timothy G. Call, Yao Yuan, Bryan Alexis Vallejo, Paul J. Hampel, Lindsey E. Roeker, Yucai Wang, Saad J. Kenderian, Sara J. Achenbach, Aaron D. Norman, Kari G. Rabe, Neil E. Kay, James R. Cerhan, Curtis A. Hanson, Susan L. Slager, Sameer A. Parikh. Validation of the CLL treatment infection model (CLL-TIM) in patients with newly diagnosed chronic lymphocytic leukemia (CLL) [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 4228.

Summary Family history of haematological malignancy (FHHM) reflects inherited susceptibility to non‐Hodgkin lymphoma (NHL), but its prognostic significance remains poorly understood. We used a cohort study of 1994 NHL patients to evaluate associations between FHHM and clinical outcomes across lymphoma subtypes. We estimated associations, adjusted for sex and lymphoma‐specific prognostic index, of self‐reported FHHM in a first‐degree relative with overall survival (OS), event‐free survival (EFS) and lymphoma‐specific survival (LSS) using multivariable Cox regression models and with failure to achieve EFS at 24 months (EFS24) using logistic regression. FHHM prevalence was 12.8%, and the median follow‐up among survivors was 12.0 years. FHHM was not associated with OS. In subtype‐specific analyses, FHHM was associated with inferior LSS (hazard ratio [HR] = 1.98, 95% confidence interval [CI] 1.11–3.53) in follicular lymphoma (FL); this association and the association with EFS24 failure strengthened among those treated with front‐line immunochemotherapy (EFS24 failure odds ratio = 2.57, 95% CI 1.08–5.99; LSS HR = 2.68, 95% CI 1.09–6.58). In mantle cell lymphoma, FHHM was associated with inferior EFS (HR = 1.70, 95% CI 1.02–2.85). Interaction testing did not demonstrate significant heterogeneity of FHHM effects by lymphoma subtype. FHHM may be associated with adverse outcomes in select NHL groups, particularly FL. These findings warrant further investigation of germline and environmental factors potentially contributing to more aggressive FLs.

Abstract Introduction The International Prognostic Index (IPI) is widely used to stratify risk in large B-cell lymphoma (LBCL) across all ages, with age-adjusted (aa)IPI applied to younger patients. While there have been efforts to develop prognostic indices (PIs) tailored to older patients, few are adopted in clinical practice due to their complexity and reliance on functional assessments, which are not routinely collected in real-world registry datasets. Recently proposed SENIOR-IPI (Dubois, ASH 2024) was developed for older patients with LBCL receiving anthracycline (A)-based therapy. We assessed the performance of standard and novel PIs in patients aged 80+ years and externally validated the SENIOR-IPI. Methods We harmonized and pooled data of patients aged 80+ years with newly diagnosed LBCL from a national registry and two prospective cohorts: NiHiL (Czech Republic, enrolled 2010–2023), LEO (USA; n=222, 2015–2020), and MER (USA; n=77, 2010–2015), forming a cohort of 819 individuals. We included 536 individuals with complete data on IPI components and albumin, including 485 treated with chemotherapy and 397 with A-based regimens (primary cohort). Following PIs were evaluated: IPI, NCCN-IPI, (aa)IPI, and SENIOR-IPI (age per year, aaIPI components, LDH &gt;3x ULN as extra point, and albumin, all weighted six-fold; extranodal involvement excluded due to lack of prognostic significance). Multivariable hazard ratios (HR) were estimated using Cox proportional hazards models; discriminative ability was assessed using C-index and calibration plots. The primary endpoint was overall survival (OS); event-free survival results were consistent (230/357 events were deaths) and are not reported here. Results Among the 397 A-treated patients (median age 82 years, range 80–99), ECOG performance status (PS) 2–4 was observed in 35% of patients, clinical stage III–IV in 67%, elevated LDH in 65% (8% with &gt;3x ULN), &gt;1 extranodal (EN) site in 32%, and albumin levels &lt; 3.5 g/dL in 27%. In multivariable modeling, only age per year (HR=1.07, P&lt;0.01), and clinical stage (HR=1.36, P=0.048) were independent predictors of OS, whereas ECOG PS (HR=1.27, P=0.12), trichotomized LDH by ULN (HR=1.23, P=0.08) and albumin (HR=1.17, P=0.35) were not significant. SENIOR-IPI had the highest discrimination (C-index 0.631) in A-treated patients, outperforming IPI (0.617), NCCN-IPI (0.591), and aaIPI (0.620), with consistent results across cohorts: NiHiL (0.633), LEO (0.619), and MER (0.621). SENIOR-IPI stratified patients into low- (L-Risk, 0–9 points, 28%), low-intermediate (L-Int, 10–19 points, 42%), high-intermediate (H-Int, 20–29 points, 23%), and high-risk (H-Risk, 30–40 points, 7%) categories, with OS medians of 6.3, 3.8, 2.4, and 1.4 years, respectively (P&lt;0.01), and significant separation between L-Int and H-Int (HR=1.47, P=0.02). By comparison, IPI categorized patients as L-Risk (12%), L-Int (24%), H-Int (27%), and H-Risk (37%), with median OS of 8.7, 4.6, 4.8, and 1.9 years, resp. (P&lt;0.01). aaIPI yielded following distributions: 13%, 29%, 37%, and 21%; and median OS: 8.7, 5.3, 3.4, and 1.5 years; P&lt;0.01. NCCN-IPI stratified patients into three groups: L-Int (13%), H-Int (49%), and H-Risk (39%), with respective OS medians of 8.7, 4.1, and 2.5 years (P&lt;0.01). All PIs showed modest overestimation of mortality (calibration slopes: SENIOR-IPI 1.36; IPI 1.24; aaIPI 1.30; NCCN-IPI 1.26; intercept: -0.97, -0.87, -0.92, -0.87, resp.). In the entire cohort (n=536), SENIOR-IPI showed highest C-index (0.651 vs IPI 0.622, aaIPI 0.631, NCCN-IPI 0.606), and stratified patients into L-Risk (26%, 2-year OS: 6.4 years), L-Int (40%, 3.2 years), H-Int (24%, 1.5 years), and H-Risk (10%, 0.6 years). Calibration slope was 1.18 (intercept -0.33). Among patients not receiving chemotherapy, SENIOR-IPI also performed best (C-index 0.759 vs IPI 0.738, aaIPI 0.757, NCCN-IPI 0.728). Conclusion SENIOR-IPI enables effective risk stratification of patients aged 80+ years with LBCL into four prognostically distinct groups, outperforming traditional indices (IPI, NCCN-IPI, aaIPI) in patients receiving A-based therapy, as well as in the entire cohort and in patients not receiving chemotherapy. While its discrimination is superior, modest miscalibration suggests that future recalibration may improve precision. These findings support SENIOR-IPI as a clinically relevant prognostic tool for the elderly LBCL patient population. Funding: NU21-03-00411, P50 CA97274, U01 CA195568.

Abstract Over the past decade, classification systems using high-throughput genetic data have been developed to identify molecular subtypes of DLBCL. Although not a primary goal, these classifiers were also proposed as prognostic factors in their respective publications. To date, however, no study has directly compared these classifiers' ability to predict overall (OS) and event-free survival (EFS) in the same cohort. Herein, we sought to systematically compare predictive performance, benchmarked against the IPI, of several current established molecular classifiers and evaluate if any add significant independent prognostic information beyond the IPI. WES (including paired germline filter) and RNAseq data from newly diagnosed DLBCL patients (N=432; 2018 WHO criteria) in the Mayo Clinic/University of Iowa Lymphoma Molecular Epidemiology Resource (MER) were available for analysis (WES n=112, RNAseq n=40, or both n=280). The number of cases able to be classified (including as unclassified) for each classifier was as follows: IPI (432, 100%), cell-of-origin (COO, by Hans algorithm; 346, 80.1%), refined COO (by gene expression with Dark Zone Signature; 320, 100%), Haematological Malignancy Research Network (HMRN; 349, 89.0%), B-cell state (320, 100%), Lymphoma Ecotype (320, 100%), Lymphoma Microenvironment (LME; 320, 100%), Tumor microenvironment (TME26; 320, 100%), SubLymE (312, 97.5%), LymphGen (358, 91.3%), and DLBclass (223, 58.2%). OS was defined as time from diagnosis to all-cause death while EFS was defined as time from diagnosis to disease progression, initiation of second-line treatment, or all-cause death. We calculated Harrell's c-statistics for each Cox proportional hazards model evaluating the discriminative performance of each classifier in predicting OS/EFS. Using bootstrap validation, we obtained optimism-corrected c-statistics, which we report in the results. We performed likelihood ratio tests (LRT) comparing a univariable Cox model with IPI alone and models including both IPI and a given classifier to assess if a classifier provided independent prognostic information beyond the IPI. All patients were treated with immunochemotherapy (IC) with curative intent, 247 (57%) were male, median age at diagnosis was 65 years (IQR 55-72 years), 390 (96%) were White, 209 (60%) had GCB subtype by Hans, 22 (5.1%) had MYC double-hit by FISH (MYC and BCL2 or BCL6), and 162 (37.9%) had an IPI ≥ 3. The median follow-up was 7.0 years, 167 (39%) experienced an event, and 137 (32%) died. In a univariable Cox model, the c-statistic for IPI predicting OS was 0.66 (95% confidence interval [CI] 0.61-0.70). All other classifiers had c-statistics of ≤0.56. The c-statistic for IPI predicting EFS was 0.64 (95% CI 0.60-0.69), and all other classifiers had c-statistics of ≤0.58. The univariate DLBclass results for OS and EFS were not significantly impacted when restricted to cases with confidence &gt;0.7 (n=136). In Cox models of IPI alone versus IPI plus an individual classifier, the only classifier that provided statistically significant, independent information for OS was TME26 (LRT p=0.02). Adjusting for IPI, patients who were TME26-negative had significantly inferior OS compared to TME26-positive cases (HR 1.63, 95% CI 1.07-2.47; p=0.02). For EFS, DLBclass (restricted to confidence &gt;0.7) added significant prognostic information beyond IPI (LRT p=0.02). This finding was largely driven by C5 subtype; relative to C1 subtype and adjusting for IPI, C5 subtype had significantly inferior EFS (HR 5.53, 95% CI 1.81-17.0; p=0.003). In summary, we compared prognostic performance of multiple DLBCL molecular classifiers with respect to OS and EFS in a single real-world, prospective cohort of newly diagnosed, IC-treated DLBCL. Our data supports two key findings: all classifiers had modest performance for prognosis, and none individually outperformed the IPI in predicting OS nor EFS. Second, only TME26 and DLBclass (confidence &gt;0.7) provided additional prognostic information beyond that captured by the IPI for OS and EFS, respectively. This finding suggests that the tumor immune microenvironment and characteristics of C5 DLBCLs (ABC-subtype, CD79B mutations) can potentially offer significant independent information for predicting survival and response to front-line immunochemotherapy. If replicated, these findings could inform risk-adapted treatment approaches and aid in identifying high-risk patients for enrollment in genomically-informed clinical trials.

Abstract Introduction Although chemo-free regimens are increasingly used in elderly patients with large B-cell lymphoma (LBCL), most still receive conventional chemoimmunotherapy. Patients up to 75–80 years old are usually treated with full-dose R-CHOP, while 80+ patients receive the attenuated R-miniCHOP regimen as standard of care. Prior studies have shown that in patients aged 65+ years, R-CHOP was associated with improved survival based on propensity score matching (PSM) analysis (Al-Sarayfi, Am J Hematol, 2023), though this did not include comorbidity data. Furthermore, no formal studies have focused specifically on 80+ patients. We aimed to perform a PSM comparison of R-CHOP versus R-miniCHOP to better inform treatment decisions in this population. Methods We harmonized and pooled data from 237 LBCL patients receiving chemotherapy with rituximab (out of all 263 80+ patients) from the prospective LEO cohort (USA, n=147; 2015–2020) and NiHiL registry (Czech Republic, single center, n=116; 2010–2023). Of them, 205 (86%) receiving anthracycline(A)-based regimens were included. Baseline characteristics, comorbidities, treatments, and outcomes were collected. PSM (1:1, caliper 0.2) was performed to compare R-CHOP and R-miniCHOP using age, ECOG performance status (PS), clinical stage, lactate dehydrogenase (LDH), and major comorbidities (heart disease, history of stroke, and other recent cancer) as matching variables. The endpoints were overall survival (OS), event-free survival (EFS), and lymphoma-specific survival (LSS, defined as death attributed to lymphoma progression). Results Among 205 A-treated patients, median age was 83 years (range 80–99); 30% had ECOG PS 2–4, 67% clinical stage III–IV, 34% ≥2 involved extranodal sites, 59% elevated LDH, and 58% had International Prognostic Index (IPI) scores 3–5. Median follow-up was 5.5 years; median OS and EFS were 4.1 and 3.4 years, resp. A-based regimens comprised full-dose R-CHOP-like (n=79; 39%; R-CHOP n=74, DA-EPOCH-R n=5) and R-miniCHOP (n=126; 61%). Patients receiving R-CHOP-like regimens had more favorable baseline characteristics (median age 82 vs 83 years, P=0.04; ECOG PS 2–4: 16% vs 39%, P&lt;0.01; IPI 3–5: 46% vs 66%, P=0.03) and a lower, though not statistically significant, comorbidity burden (mean 1.14 vs 1.39; heart disease 35% vs 45%, P=0.22; history of stroke 4% vs 6%, P=0.64; other cancer 6% vs 5%, P=0.87). R-CHOP-like regimens vs R-miniCHOP had superior OS (median 6.3 vs 3.4 years; HR=0.61, 95% CI 0.42–0.90; P=0.01) and EFS (median 5.3 vs 2.1 years; HR=0.58, 95% CI 0.40–0.83; P&lt;0.01), with no significant difference in LSS (2-year 89% vs 82%, HR=0.77, 95% CI 0.39–1.54, P=0.46). To mitigate selection bias, we performed PSM on patients with available baseline characteristics treated with R-CHOP (n=69) and R-miniCHOP (n=121), resulting in 63 matched pairs. Matched cohorts were well-balanced on age (median 82 years in both cohorts; P=0.68), ECOG PS 2–4 (17% vs 19%; P=0.82), IPI 3–5 (52% vs. 59%; P=0.41), and comorbidity burden (mean 1.10 vs. 1.29; P=0.30). The presence of heart disease (40% in both cohorts), history of stroke (5% vs 6%), and cancer (5% vs 6%) were similar. Survival modestly favored R-CHOP, but differences were not significant: median OS 5.3 vs 3.5 years (HR=0.73; 95% CI 0.46–1.15; P=0.17), median EFS 3.8 vs 2.8 years (HR=0.76; 95% CI 0.49–1.18; P=0.23), and 2-year LSS 88% vs 84% (HR=1.09; 95% CI 0.45–2.63; P=0.85). Treatment-related mortality was similar in both cohorts (16%). Among R-miniCHOP patients, survival outcomes did not differ between matched (n=63) and unmatched (n=63) groups (OS P=0.72; EFS P=0.23; LSS P=0.27). Unmatched (n=16) R-CHOP patients had a trend toward improved survival compared to matched (n=63) individuals (OS P=0.10; EFS P=0.049; LSS P=0.24). Conclusion This is the first PSM-based analysis comparing R-CHOP and R-miniCHOP in LBCL patients 80+ years old. Despite inherent limitations of a non-randomized design and limited number of matched patients, our findings – derived from two prospective cohorts – suggest that R-miniCHOP may offer outcomes comparable to R-CHOP, particularly with respect to LSS. While these results support R-miniCHOP as an appropriate frontline option, treatment decisions remain highly individualized. These data may help inform clinical decision-making and provide a basis for designing future trials targeting this population, where R-miniCHOP could serve as a comparator arm. Funding: NU21-03-00411, U01 CA195568.

Abstract Introduction There is currently no standard-of-care for patients (pts) who cannot tolerate full dose anthracycline regimens in frontline large B-cell lymphoma (LBCL). R-miniCHOP is a commonly used regimen that provides curative intent treatment with manageable toxicity. Despite this, lymphoma and treatment-related mortality are the leading causes of death in these pts. Additionally, many pts may not be eligible to receive R-miniCHOP due to comorbidity or frailty. There are limited clinical trial and real-world data focused on this population. Methods Pts with LBCL age 18 years (yrs) and older were identified from the Lymphoma Epidemiology of Outcomes (LEO) cohort, a prospective study of newly diagnosed pts with non-Hodgkin lymphoma enrolled from 2015-2020 at 8 academic medical centers in the US. Pts with LBCL were included in this analysis if they had received R-miniCHOP or a regimen that did not include an anthracycline (nonAC). NonAC regimens included R-CVP, BR, rituximab, radiation and lenalidomide and rituximab. Pts completed the Vulnerable Elders Survey (VES-13) and were considered vulnerable if they had a score &gt;3. Event-free survival (EFS) was defined as the time from start of treatment until progression/relapse, retreatment, or death due to any cause; EFS at 24 months (EFS24) was a secondary endpoint. Overall survival (OS) was defined as the time from start of treatment until death due to any cause. EFS and OS for R-miniCHOP and non-IC groups, as well as vulnerable and not vulnerable by VES-13 were estimated using Cox models. Results A total of 190 pts received R-miniCHOP (N=139) or nonAC (N=51). R-miniCHOP-treated pts were older (median 82 vs 72 yrs) and more commonly female (55% vs 35%). The majority in both groups (86% vs 78%) received their treatment at a LEO center and did not receive therapy on a clinical trial (89% vs 100%). Most pts receiving R-miniCHOP had high-intermediate (32%) or high (30%) international prognostic index (IPI) score, while both high-intermediate (22%) and high (20%) IPI were lower in the nonAC-treated pts. Cell of origin was determined by immunohistochemistry as available clinically, with GCB 42% vs 31%, non-GCB 32% vs 29%, and not done/unknown 26% vs 39% in R-miniCHOP vs nonAC pts. Double-hit status by FISH was known in 115 pts and was positive in 13% of R-miniCHOP pts and 10% of nonAC pts. The diagnosis-to-treatment interval (DTI) was ≤14 days in 30% of R-miniCHOP pts and 24% of nonAC pts. Of pts who had a VES-13 score (N=119), vulnerable status trended higher in the R-miniCHOP (66%) vs the non-IC group (50%; p=0.15). With a median follow-up of 50.5 months, the median EFS was 37.1 months and median OS was 50.0 months. R-miniCHOP, compared with non-IC, was associated with improved EFS (unadjusted HR=0.54, 95% CI 0.36-0.81; p=0.003) and OS (unadjusted HR=0.58, 95% CI 0.38-0.88; p=0.011). When adjusted for age, IPI, LBCL subtype (diffuse large B-cell lymphoma, not otherwise specified vs other), DTI and VES-13, R-miniCHOP was strongly associated with improved EFS (HR=0.26, 95% CI 0.14-0.50; p&lt;0.001) and OS (adjusted HR=0.29, 95% CI 0.15-0.57; p&lt;0.001). EFS24 rate was 56% in all pts and was superior in pts treated with R-miniCHOP vs non-IC (62.7% vs 34.1%; p=0.001). In the R-miniCHOP group, vulnerable status (adjusted for IPI) was associated with inferior OS (HR=2.22, 95% CI 1.09-4.50; p=0.028) and trended towards inferior EFS (HR=1.76, 95% CI 0.93-3.32; p=0.083). In a competing-risk cumulative-incidence analysis, lymphoma was the leading cause of death for pts receiving R-miniCHOP at 2 (0.16, 95% CI 0.11-0.24) and 5 (0.20, 95% CI 0.14-0.28) yrs, as well as for those receiving non-IC at 2 (0.35, 95% CI 0.24-0.53) and 5 (0.38, 95% CI 0.26-0.56) yrs. Treatment-related mortality was similar in both groups. In the R-miniCHOP group, the 2-yr (0.20 vs 0.03) and 5-yr (0.20 vs 0.11) lymphoma-related mortality was higher in pts who were vulnerable vs not by VES-13, while there were no differences for mortality due to therapy or other causes by vulnerable status. Conclusions These data support R-miniCHOP as a curative regimen in older and frail pts with the majority achieving EFS24. R-miniCHOP compared to non-IC use is strongly associated with improved EFS and OS when adjusted for disease and patient related variables. Within the R-miniCHOP group, vulnerable status is associated with inferior OS. Additional research efforts are needed to optimize outcomes for this group of pts.

Abstract Background: Recent advances in computer vision have given rise to image foundation models (FMs), large-scale neural networks pre-trained on millions of unlabeled histopathology patches, which can learn generalizable representations of tissue morphology. In many cancers, these FMs have predicted genetic mutations and prognoses directly from histology, suggesting that morphology encodes “hidden” molecular information. In lymphoma, this field is nascent but growing.General-purpose histopathology FMs have limited representation of lymphoid morphology. We developed and evaluated LymphoVision, a lymphoma-specialized image foundation model pretrained with self-supervised learning, to extract clinically meaningful features from routine hematoxylin and eosin (H&E) histology. Methods: LymphoVision was pre-trained using knowledge distillation and consistency (DINOv2) on over 37 million multi-resolution patches (5x, 10x, 20x, 40x) sampled using a density-aware strategy from 31,211 H&E whole slide images (WSIs) from 9,155 archival lymphoma and reactive lymphoproliferative cases. To ensure diverse, representative samples, our patch sampling method uses an efficient graph-based clustering algorithm and prioritizes morphologically informative regions by clustering patch embeddings and performing density sampling within each cluster. The pretrained encoder was evaluated using a clustering-constrained attention multiple instance learning (MIL) framework on three diagnostic tasks: (1) cell-of-origin (COO) classification in diffuse large B-cell lymphoma (DLBCL), distinguishing germinal center B-cell (GCB) from non-GCB subtypes based on the Hans immunohistochemistry algorithm; (2) follicular lymphoma (FL) grading, differentiating low-grade (grades 1–2) from high-grade (grade 3); and (3) multiclass lymphoma subtype classification across six categories: DLBCL, FL, marginal zone lymphoma (MZL), mantle cell lymphoma (MCL), classical Hodgkin lymphoma (CHL), and benign/reactive lymphoid conditions. WSI-level labels were derived from pathology reports, and none of the evaluation cases were included in pretraining. Models used a ViT-Giant LymphoVision backbone to generate 20x patch embeddings, followed by MIL using 50% training data and 2-fold cross-validation (25% validation, 25% test). Performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, recall (sensitivity) and F1 score. Results: For DLBCL COO classification (total n=358 cases), LymphoVision achieved a mean test AUC of 0.93 (range 0.92–0.94), accuracy of 84.5% and F1 of 0.85 across two independent test folds (n=91 per fold, 50% GCB). These results outperformed general-purpose FMs on the same task and cases, including Virchow-v2 (AUC 0.85) and UNI-v2 (AUC 0.88), despite those models being pretrained on 10–100 times more images. For FL grading (total n=660 cases), the test set in each fold included 166 cases (18% grade 3), yielding a mean test AUC of 0.88 (range 0.87–0.90), accuracy of 89.2% and F1 of 0.80. In multiclass lymphoma subtype classification (total n=1,176 cases), the test set included 296 cases per fold. LymphoVision achieved a mean test AUC of 0.98 (range 0.97-0.99), mean test accuracy of 87.1% and mean F1 of 0.84. Class-wise recall was 92.0% for DLBCL (n=112), 88.2% for FL (n=72), 77.2% for MZL (n=29), 70.8% for MCL (n=35), 87.5% for CHL (n=24), and 95.8% for benign/reactive lymphoid conditions (n=24). Conclusion: LymphoVision is a lymphoma-specialized FM trained using self-supervision that achieves high diagnostic accuracy across multiple clinically meaningful tasks using weak supervision and standard H&E images, suggesting the encoder has learned fine-grained lymphoma morphology. Based on historical benchmarks from the 1997 International Lymphoma Study Group classification project, expert hematopathologists achieved diagnostic agreement rates of 55% to 84% using H&E morphology alone for the non-Hodgkin lymphoma (NHL) entities included in our multiclass task. LymphoVision approached or exceeded these benchmarks. These findings underscore the value of our approach using disease-specialized pretraining and intelligent patch sampling for data-efficient histopathology learning. LymphoVision has the potential to augment diagnostics, enable scalable AI-powered classification in real-world workflows, and serve as a foundation for prognostic and theragnostic applications. External validation and multimodal integration are ongoing.

Abstract Although follicular lymphoma (FL) typically follows an indolent course, patients with FL who experience early events, such as transformation or progression, have increased risk of death related to lymphoma. The FL24Cx is an algorithm based on a 45-target gene expression profiling (GEP) assay, which was developed and trained using 265 formalin-fixed, paraffin-embedded tissue samples on a reliable platform to predict, at the time of diagnosis, whether a patient will experience an event within 24 months. The modeling also confirmed and relied upon previously reported synergy between immune response (IR) gene expression signatures IR1 and IR2. Once locked, the 5-factor logistic regression FL24Cx model was independently validated in a retrospectively assessed cohort of 232 patients from 2 immunochemotherapy-treated arms of SWOG Cancer Research Network S0016 phase 3 clinical trial, in which it assigned 169 patients to the low-risk group with 29 events before 24 months (17.2%) and 63 patients to the high-risk group with 24 events before 24 months (38.1%). The relative risk of an event within 24 months after registration among patients who were classified into the high-risk group relative to patients who were classified into the low-risk group was 2.2 (95% confidence interval, 1.41 to 3.51). An up-front GEP biomarker, such as the FL24Cx, rigorously validated in a clinical laboratory and with a clinically relevant turnaround time, could identify and steer enrollment of patients at high risk for early events in clinical trials, thus enabling timely interpretation of such trials and increasing the pace of innovation.