TRAIL Score: A Simple Model to Predict Immunochemotherapy Tolerability in Patients With Diffuse Large B-Cell Lymphoma

Abstract

PURPOSE Rituximab with cyclophosphamide, doxorubicin, vincristine, and prednisolone (R-CHOP) represents the standard of care for first-line treatment of diffuse large B-cell lymphoma (DLBCL). However, many patients are unable to tolerate R-CHOP and have inferior outcomes. This study aimed to develop a practical tool to help physicians identify patients with newly diagnosed DLBCL unlikely to tolerate a full course of R-CHOP. METHODS We developed a predictive model (Tolerability of R-CHOP in Aggressive Lymphoma [TRAIL]) on the basis of a training data set from the phase III GOYA trial (obinutuzumab with CHOP v R-CHOP in 1L DLBCL) using a composite binary end point, identifying patients who prematurely stopped or required reductions of R-CHOP. Candidate predictive variables were selected on the basis of known baseline characteristics that contribute to patient frailty, comorbidity, and/or chemotherapy toxicity. TRAIL was developed using an iterative trial-and-error modeling process to fit a logistic regression model. The final model was evaluated for robustness using a GOYA holdout data set and the phase III MAIN (R-CHOP with or without bevacizumab in 1L DLBCL) R-CHOP-21 data set as external validation. RESULTS TRAIL includes four simple predictors available in the routine clinical setting: Charlson Comorbidity Index, presence of cardiovascular disease or diabetes, serum albumin, and creatinine clearance. Model generalization performance estimated by the area under the curve was around or above 0.70 across GOYA training, GOYA holdout, and MAIN data sets. Classifying patients into low-, intermediate- and high-risk categories, the proportion of patients experiencing a tolerability event was 3.3%, 12.4%, and 32.9%, respectively, in GOYA holdout, and 9.7%, 9.7%, and 34.2%, respectively, in MAIN. CONCLUSION TRAIL may be useful as a clinical decision support tool for treatment decisions in patients with DLBCL who may not tolerate standard chemoimmunotherapies.

GameRank: R package for feature selection and construction

“…We evaluated the GameRank algorithm on the real-world dataset used in Harris et al (2022) We compared the time, performance as measured by area under the ROC curve and bias between hold-out and validation sets for GameRank, random, forward, backward and the bidirectional selection algorithms for feature selection ( Table 1 ). We observed that GameRank, while being slower than random search and forward selection for very small variable sizes, maintains a very stable runtime independent of the number of features selected.…”

“…Before variable selection makes sense, a thorough review of each included variable is essential. From our experience, Harris et al (2022) , clinical variables can be degenerate, show high distribution skew or may even be multi-modal. Therefore, we provide methods to evaluate variable transformation (sqrt, cube root, log), e.g.…”

“…Here, we present an R package consisting of a framework for clinical variable selection that includes a novel algorithm, GameRank, which has been previously applied to building a clinical predictive model for the prediction of chemotherapy tolerability ( Harris et al , 2022 ).…”

The role of chemotherapy toxicity scales in decision-making for acute leukemia and lymphoma: a scoping review

“…Contudo, existem muitas semelhanças entre as variáveis utilizadas e este processo de elaboração de indicadores específicos para avaliação de toxicidade neste grupo é importante, pois conseguem ser superiores aos atuais preditores de prognóstico e funcionalidade ao menos ao predizer a toxicidade da quimioterapia. (HARRIS et al, 2022) O tratamento das leucemias e linfomas de uma forma geral utilizam tratamento de primeira linha com protocolos de quimioterapia com potencial de toxicidade elevado, sendo esse tipo de tratamento evitado em grupos de pessoas idosas, que acabam realizando esquemas alternativos e de menor intensidade, mas com menor potencial de cura. Isso ocorre por uma avaliação de risco de toxicidade que muitas vezes é intuitiva levando em consideração apenas a idade.…”

“…Desse modo, a elaboração de indicadores específicos para avaliação de toxicidade têm mostrado resultados superiores aos atuais preditores de prognóstico e funcionalidade, ao menos em predizer a toxicidade da quimioterapia (HARRIS et al, 2022) .…”