A Machine Learning Model Based on Tumor and Immune Biomarkers to Predict Undetectable MRD and Survival Outcomes in Multiple Myeloma
Camila Guerrero, Noemí Puig, María‐Teresa Cedena, Ibai Goicoechea, Cristina Pérez, Juan‐José Garcés, Cirino Botta, Marı́a José Calasanz, Norma C. Gutiérrez, María-Luisa Martín-Ramos, Albert Oriol, Rafael Ríos, Miguel-Teodoro Hernandez, Rafael Martínez-Martinez, Joan Bargay, Felipe de Arriba, Luis Palomera, Ana Pilar González, Adrián Mosquera-Orgueira, Marta Sonia González, Joaquín Martínez‐López, Juan José Lahuerta, Laura Rosiñol, Joan Bladé, María‐Victoria Mateos, Jesús F. San Miguel, Bruno Paiva
Abstract
PURPOSE: Undetectable measurable residual disease (MRD) is a surrogate of prolonged survival in multiple myeloma. Thus, treatment individualization based on the probability of a patient achieving undetectable MRD with a singular regimen could represent a new concept toward personalized treatment, with fast assessment of its success. This has never been investigated; therefore, we sought to define a machine learning model to predict undetectable MRD at the onset of multiple myeloma. EXPERIMENTAL DESIGN: This study included 487 newly diagnosed patients with multiple myeloma. The training (n = 152) and internal validation cohorts (n = 149) consisted of 301 transplant-eligible patients with active multiple myeloma enrolled in the GEM2012MENOS65 trial. Two external validation cohorts were defined by 76 high-risk transplant-eligible patients with smoldering multiple myeloma enrolled in the Grupo Español de Mieloma(GEM)-CESAR trial, and 110 transplant-ineligible elderly patients enrolled in the GEM-CLARIDEX trial. RESULTS: The most effective model to predict MRD status resulted from integrating cytogenetic [t(4;14) and/or del(17p13)], tumor burden (bone marrow plasma cell clonality and circulating tumor cells), and immune-related biomarkers. Accurate predictions of MRD outcomes were achieved in 71% of cases in the GEM2012MENOS65 trial (n = 214/301) and 72% in the external validation cohorts (n = 134/186). The model also predicted sustained MRD negativity from consolidation onto 2 years maintenance (GEM2014MAIN). High-confidence prediction of undetectable MRD at diagnosis identified a subgroup of patients with active multiple myeloma with 80% and 93% progression-free and overall survival rates at 5 years. CONCLUSIONS: It is possible to accurately predict MRD outcomes using an integrative, weighted model defined by machine learning algorithms. This is a new concept toward individualized treatment in multiple myeloma. See related commentary by Pawlyn and Davies, p. 2482.