Selection of CUT-Off Points for Prognostic Biomarkers in Lymphoma Using Restricted Cubic Splines: Experience of the Grupo De Estudio De Linfoproliferativos De Latinoamérica (GELL) in a Cohort of Diffuse Large B-Cell Lymphoma Patients

Introduction: The platelet/albumin ratio (PAR) and the hemoglobin, albumin, lymphocyte, and platelet (HALP) index have previously been evaluated as prognostic biomarkers of overall survival (OS) in patients with diffuse large B-cell lymphoma (DLBCL) in Latin America (LATAM). Recent studies have primarily relied on ROC curve analysis for selecting cut-off points, which only allows for estimating a single cut-off point. However, restricted cubic splines (RCS) offer a flexible method to model nonlinear relationships in regression models. This facilitates a more precise selection of one or more cut-off points, supplemented by clinical judgment. The nonlinear relationship manifests in different groupings of biomarker values with varying risks, justifying using two established prognostic scores (PAR/HALP) to assess if RCS can effectively establish cut-off points.

Objective: To select and evaluate the prognostic relevance of cut-off points for the PAR and the HALP Index using RCS in patients with DLBCL.

Methods: We conducted a retrospective cohort study using a database (2012-2022) of patients diagnosed with DLBCL treated at various specialized centers across LATAM under the Grupo de Estudio de Linfoproliferativos de Latinoamérica (GELL). Patients with incomplete data (n=84) or could not be contacted (n=3) were excluded. The primary outcome evaluated was overall survival (OS), the time from diagnosis until death from any cause or censuring. The PAR was calculated by dividing the platelet count by the serum albumin level, while the HALP index was calculated as hemoglobin level (g/L) × albumin level (g/L) × lymphocyte count (/L) / platelet count (/L), both measured at diagnosis. RCS with 3 knots was employed in a univariate Cox regression model to determine the nonlinear association between PAR and HALP levels with OS, confirmed by the Wald test. The optimal cut-off point was identified based on a clinically relevant change in the curve. After categorization, the association was analyzed using univariate and multivariate Cox regression, reporting hazard ratios (HR), and 95% confidence intervals (95% CI). Kaplan-Meier survival curves and the Log-Rank test were used for comparison. Analyses were performed using R (version 4.1.2; R Project for Statistical Computing) and relevant R packages (‘rms’, ‘survival’, ‘survminer’, and 'ggplot2'). Missing data was evaluated and imputed using the 'mice' package.

Results: A total of 1,383 patients were included. The median age was 65 years (range: 53-74), and 51% were men. Eighty percent received R-CHOP-based treatments. The median follow-up was 32 months (range: 0.1-152). In the RCS analysis, PAR (p<0.001) and the HALP index (p<0.001) exhibited nonlinear behavior, with two peaks observed in the prognostic relationship of PAR. The selected cut-off points for HALP were <0.2 versus >0.2 (comparison group), and for PAR: <40,000 and >80,000 versus 40,000-80,000 (comparison group). In univariate and multivariate analyses adjusted for gender and International Prognostic Index, a HALP <0.2 (HR: 1.64; 95% CI: 1.36-1.98) and a PAR <40,000 (HR: 1.77; 95% CI: 1.37-2.28) predicted worse OS. PAR >80,000 in the univariate analysis (HR: 1.42; 95% CI: 1.19-1.68) was associated with poor survival; however, this was not observed in the multivariate analysis (HR: 1.05; 95% CI: 0.87-1.27). Survival curves demonstrated a statistically significant difference between HALP >0.20 versus <0.20 (p<0.001) and PAR <40,000 versus >80,000 versus 40,000-80,000 (p<0.001).

Conclusions: The PAR and HALP Index are nonlinear prognostic biomarkers of OS in patients with DLBCL. The use of RCS could be a potentially valuable statistical method for selecting cut-off points, as it allows for the integration of clinical judgment by visualizing risk behavior across all values of the continuous variable. While only one cut-off point was identified for HALP, two potential cut-offs were observed for PAR, highlighting the utility of RCS. This contrasts with the ROC curve, which typically identifies a single cut-off point. Further research is needed to assess the impact of RCS implementation in lymphoma studies.

Perini:BMS, Roche, Abbvie, AsstaZeneca, Beigene: Speakers Bureau. Gomez-Almaguer:Janssen: Consultancy, Other: Advisory board, Speakers Bureau; Astex Pharmaceuticals: Research Funding; Gilead/Forty Seven: Research Funding; Blueprint Medicines: Research Funding; Seattle Genetics: Research Funding; Novartis: Consultancy, Other: Advisory board, Speakers Bureau; Kartos Therapeutics: Research Funding; Incyte: Research Funding; Amgen: Consultancy, Other: Advisory board, Research Funding, Speakers Bureau; ConstellationPharmaceuticals: Research Funding; Sanofi: Speakers Bureau; Takeda: Consultancy, Other: Advisory board, Research Funding, Speakers Bureau; BMS: Consultancy, Other: Advisory board, Speakers Bureau; Roche: Speakers Bureau; Tevas: Speakers Bureau; AbbVie: Research Funding, Speakers Bureau. Malpica:Dizal: Research Funding; Eisai: Research Funding. Castillo:Pharmacyclics: Consultancy, Research Funding; Kite Pharmaceuticals: Consultancy; LOXO: Consultancy, Research Funding; AstraZeneca: Consultancy, Research Funding; BeiGene: Consultancy, Research Funding; Mustang Bio: Consultancy; Janssen: Consultancy; AbbVie: Consultancy, Research Funding; Cellectar Biosciences: Consultancy, Research Funding.