The research, led by Jan-Michel Heger of University Hospital Cologne together with colleagues from centers across Germany and Austria, set out to validate which variables are clinically relevant prognostic factors and which act as treatment-effect modifiers in DLBCL. The distinction is technically crucial. A prognostic factor predicts how patients fare overall regardless of which therapy they receive, such as advanced age or high tumor burden. A treatment-effect modifier, by contrast, changes how well a specific therapy works, meaning the treatment’s benefit itself differs depending on the patient’s status. Confusing the two can distort indirect treatment comparisons, the family of statistical methods, including matching-adjusted and simulated treatment comparisons, that health technology assessment bodies increasingly rely on when randomized evidence is absent. If an effect modifier is left unadjusted, the comparison may attribute differences in outcome to the drug when they actually stem from the patient mix.

To build the variable list, the team first drew on a prior systematic literature review of prognostic factors for efficacy and safety outcomes of CAR-T therapy in DLBCL, then refined the candidates through clinical review. Six experienced hematologists and oncologists from Germany and Austria, institutions spanning Cologne, Berlin, Essen, and Vienna, then participated in a structured, two-round expert elicitation process modeled on consensus methodologies such as the Delphi approach. Each expert rated the importance of every candidate variable as a prognostic factor, a treatment-effect modifier, or both, using a three-point Likert scale. Variables that crossed predefined consensus thresholds were assigned to relevance tiers and then mapped to specific clinical endpoints: progression-free survival, overall survival, response outcomes, and CAR-T-related toxicities such as cytokine release syndrome and immune effector cell-associated neurotoxicity syndrome, the two signature adverse events of this therapeutic class.

The findings on prognostic factors were remarkably coherent. Thirteen variables reached the top relevance tier: Eastern Cooperative Oncology Group performance status, age, disease histology, the International Prognostic Index, disease stage, elevated lactate dehydrogenase, bulky disease, extranodal involvement, tumor burden, primary refractory disease, number of relapses, time to relapse, and the number of prior lines of therapy. Six of these achieved unanimous agreement among all panelists as prognostic factors for survival outcomes: ECOG performance status, the IPI, disease histology, disease stage, elevated LDH, and primary refractory disease. These are, for the most part, the venerable workhorses of lymphoma prognostication. The IPI, developed in the early 1990s, combines age, performance status, stage, LDH, and extranodal sites into a single risk score, and its components dominated the panel’s top tier, suggesting that decades of clinical experience have produced a stable, shared understanding of what drives outcomes in aggressive B-cell malignancy.

The logic behind each variable is biologically plausible. ECOG performance status captures a patient’s functional fitness, which influences both tolerance of intensive lymphodepleting chemotherapy before CAR-T infusion and the ability to survive complications. Elevated LDH signals high tumor turnover and burden. Primary refractory disease, meaning lymphoma that never responded to initial therapy, and short time to relapse both mark biologically aggressive, treatment-resistant clones. Number of prior therapy lines reflects cumulative exposure to cytotoxic regimens and often correlates with diminishing responsiveness. Bulky disease and high tumor burden strain not only efficacy but also safety, a point the panel emphasized: tumor burden-related variables were judged important for both efficacy endpoints and toxicity outcomes, including the risk of cytokine release syndrome, in which engineered T cells trigger a systemic inflammatory cascade, and neurotoxicity, which can range from mild confusion to life-threatening cerebral edema.

Where consensus weakened was precisely where the science is least mature: treatment-effect modifiers. Compared with the strong agreement on prognostic factors, substantially lower agreement emerged for effect modifiers, particularly for response-based outcomes such as complete response and overall response rates. Only a handful of variables reached the top tier as effect modifiers: primary refractory disease, time to first relapse, best response to prior therapy, bridging therapy, CAR-T product type, and CAR-T-related response characteristics. Bridging therapy, the treatment given between leukapheresis, when a patient’s T cells are harvested for engineering, and infusion of the final CAR-T product, was the sole modifier to achieve unanimous agreement for complete and overall response rates. That makes intuitive sense: bridging therapy both controls disease during the manufacturing window and may itself select for or against certain patient profiles, entangling its effect with the therapy being evaluated.

The asymmetry between prognostic factors and effect modifiers is not a failure of the panel but a reflection of the evidence base. Identifying a prognostic factor requires only observing that outcomes vary with a characteristic; identifying an effect modifier requires evidence that the magnitude of a treatment’s benefit differs across levels of that characteristic, which demands comparative data that are often simply unavailable. Recent real-world analyses comparing tisagenlecleucel and axicabtagene ciloleucel, the two most widely used CD19-directed CAR-T products in DLBCL, have begun to probe such interactions, but the literature remains thin and heterogeneous. The panel’s hesitancy, in other words, encodes genuine scientific uncertainty rather than disagreement about method. For analysts constructing indirect comparisons, this means adjustment strategies for effect modification must be justified case by case, with sensitivity analyses exploring how conclusions change under different assumptions.

The study’s methodology also carries lessons for the broader field of comparative effectiveness research. Structured expert elicitation has become an accepted tool when empirical data cannot settle a question, and the two-round design with predefined consensus thresholds guards against the dominance of a single loud voice. The approach mirrors frameworks developed by health technology assessment bodies, including the German Institute for Quality and Efficiency in Health Care, which has promoted systematic confounder identification in indirect comparisons. By allocating variables to specific endpoints rather than treating confounding as a generic problem, the panel acknowledged a technical subtlety: a variable may be essential to adjust for when comparing progression-free survival but irrelevant for response rates, and a variable that predicts toxicity, such as tumor burden’s link to cytokine release syndrome, may need separate consideration in safety analyses.

There are caveats worth noting. The panel comprised six clinicians from two neighboring countries with broadly similar treatment landscapes, and their judgments, however experienced, are ultimately structured opinion rather than empirical proof. The study was funded by Miltenyi Biomedicine, a manufacturer of CAR-T technology, with several authors employed by the company or a contract research organization, although the funder reported no role in study design, data collection, or interpretation. Participants were compensated, and the elicitation captured physician judgment rather than patient-level data. These limitations do not undermine the central contribution, but they mean the resulting variable list should be treated as a validated starting framework, to be updated as real-world registries and comparative studies accumulate.

For patients with relapsed or refractory DLBCL, the practical stakes are considerable. CAR-T therapy has transformed outcomes for a population that once had few options, but the therapy is expensive, logistically demanding, and carries real risks, so payers and clinicians need trustworthy evidence about which product or sequence works best for whom. Indirect comparisons are often the only bridge across the gaps in randomized evidence, and this study supplies the engineering specifications for that bridge: a consensus-backed set of prognostic factors that must be balanced, a shorter and more tentative list of effect modifiers demanding caution, and an explicit mapping of variables to endpoints. The next step, the authors and the field suggest, is to test these expert-identified variables empirically in large, multinational real-world datasets, turning structured clinical intuition into statistically verified adjustment models that can withstand regulatory scrutiny.

Subject of Research: Confounder and effect-modifier identification for indirect treatment comparisons of CAR-T therapy in relapsed or refractory diffuse large B-cell lymphoma

Article Title: Confounder Identification in Diffuse Large B-Cell Lymphoma: Findings from an Expert Panel of German and Austrian Hematologists

Article References: Heger, J.-M., Gödel, P., Habringer, S., Jäger, U., Kutsch, N., von Tresckow, B., Zhang, R., Rungaldier, S., Oddsdottir, J., Zacharioudaki, M., & Mahlich, J. (2026). Confounder Identification in Diffuse Large B-Cell Lymphoma: Findings from an Expert Panel of German and Austrian Hematologists. Advances in Therapy. https://doi.org/10.1007/s12325-026-03820-z

Image Credits: AI Generated

DOI: 10.1007/s12325-026-03820-z

Keywords: diffuse large B-cell lymphoma, CAR-T therapy, confounders, prognostic factors, treatment-effect modifiers, indirect treatment comparisons, expert consensus, bridging therapy, cytokine release syndrome, International Prognostic Index, health technology assessment, hematology