Harmony Alliance Provides a Machine Learning Researching Tool to Predict the Risk of Relapse after First Remission in AML Patients Treated without Allogeneic Haematopoietic Stem Cell Transplantation

• Published in: Blood 2021-11, Vol.138 (Supplement 1), p.4041-4041
• Main Authors: Sobas, Marta, Martinez Elicegui, Javier, Villaverde Ramiro, Angela, González, Teresa, Hernandez-Sanchez, Alberto, Azibeiro Melchor, Raúl, Benner, Axel, Sträng, Eric, Gastone, Castellani, Heckman, Caroline A., Heredia Casanoves, Ana, Versluis, Jurjen, Abáigar, María, Jamilis, Laura, Valk, Peter JM, Metzeler, Klaus H., Ayala, Rosa, Martinez Lopez, Joaquin, Dombret, Hervé, Sierra, Jorge, Preudomme, Claude, Damm, Frederick, Mills, Ken I, Mayer, Jiří, Thiede, Christian, Voso, Maria Teresa, Amadori, Sergio, Sanz, Guillermo, Calado, Frederico, Döhner, Konstanze, Gaidzik, Verena I, Heuser, Michael, Haferlach, Torsten, Turki, Amin T., Reinhardt, Dirk, Villoria Medina, Rubén, Van Speybroeck, Michel, Schulze-Rath, Renate, Butler, John E, Huntly, Brian J.P, Dohner, Hartmut, Ossenkoppele, Gert J., Bullinger, Lars, Hernández-Rivas, Jesus Maria
• Format: Article
• Language: English
• ISSN: 0006-4971; 1528-0020
• DOI: 10.1182/blood-2021-149521

Background: The decision to perform allogeneic haematopoietic stem cell transplantation (alloHSCT) in acute myeloid leukemia (AML) is based on the risk-benefit ratio (non relapse mortality vs reduction of relapse risk). In 2017, the European LeukemiaNet (ELN) proposed a risk score based on cytogenetic and molecular genetic characteristics to facilitate this decision. Despite this improved classification of the genetic landscape of AML, the assessment of risk of relapse should be more precise. However, large cohorts are needed to analyze the clinical outcome of specific genetic alterations. Within the HARMONY alliance, we have now collected harmonized clinical and analytical data for a large number of AML patients. Aims: This study focuses on AML patients who achieved first complete remission (CR1) that, according to ELN risk (low/intermediate) assessment are not classical candidates for alloHSCT as consolidation therapy. The aim of this study is to create a more accurate risk prediction in this setting based on an on-line tool that can visualize the likelihood of relapse and thereby help to determine in which patient alloHSCT should be performed in CR1. Methods: The data included in the HARMONY alliance database was provided by 100 organisations in 18 European countries. In order to be accepted, they passed through quality control, anonymisation and harmonisation processes before being included in the database. Harmonisation is carried out according to the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM), which is specially designed to accommodate both administrative claims and medical records, making it possible to bring together all the information from different data sources and to speed up its subsequent analysis. Through the analysis platform, we selected patients from the ~5700 patients available that matched the target population of the study. We filtered out those patients without sufficient information on their clinical course, those who did not achieve complete remission and patients with a poor prognosis (adverse risk according to ELN2017), as the study focuses on patients who a priori did not have an indication for alloHSCT. This process resulted in a sample of 842 patients. In the next steps, variable selection was performed together with the treatment of incomplete cases by imputation. Multiple Machine Learning (ML) techniques, both parametric and non-parametric, were tested for predictions (Random Forest, Weibull dis