A PERSONALIZED risk-prediction model for myelodysplastic syndromes (MDS) has been developed through the use of a machine-learning approach that analyzes genomic and clinical data for an individual patient. According to lead investigator Aziz Nazha, MD, of the Cleveland Clinic, the model provides survival estimates at different time points, predicts for transformation, and outperforms current models in clinical practice, such as the Revised International Prognostic Scoring System (IPSS-R) tool.
Aziz Nazha, MD
“We wanted to build a personalized prediction tool that can provide insights about a specific outcome for a specific patient,” Dr. Nazha said at a press briefing during the 2018 American Society of Hematology Annual Meeting & Exposition.1 “The algorithm extracts important variables, which are not easily done by a treating clinician.”
In addition to his work as a clinician treating patients with MDS, Dr. Nazha is a programmer with expertise in artificial intelligence and machine learning. He has estimated that current models misclassify risk in up to 30% of patients, which leads to overtreatment and undertreatment.
“The survival of patients with MDS can sometimes be measured over years, but for others, it is over a matter of months. You can find some patients with low risk [whose disease course is] worse than for patients with high risk. This has made us ask the question of whether we can build a model that provides personalized prediction for a given patient,” he said. “Understanding a patient’s prognosis allows us to more appropriately develop a treatment plan and counsel patients. We are optimistic that an improved prediction will lead to more personalized care.”
Building the Machine-Learning Model
THE MODEL was developed with data from patients treated at the Cleveland Clinic and the Munich Leukemia Laboratory (n = 1,471) and was validated in a separate cohort from the Moffitt Cancer Center (n = 831). The developers fed demographic, clinical, and genomics data into the algorithm and incorporated next-generation targeted deep sequencing of 40 gene mutations common to myeloid malignancies. The aim was to accurately predict an individual’s overall survival probability and risk for transformation to acute myeloid leukemia (AML), ie, leukemia-free survival.
In the IPSS-R mutational model, 3 mutations are considered to be significant, but the machine-learning model identified 12. The developers found mutations in 11 genes to be prognostic of survival, in the following order: TP53, RUNX1, STAG2, ASXL1, SF3B1, SRSF2, RAD21, NRAS, NPM1, TET2, and EZH2.
The variables incorporated into the model included IPSS-R cytogenetic risk categories, platelet count, white blood cell count, hemoglobin level, percent bone marrow blasts, and mutational number and status for relevant genes. The final prediction model was based on the most important variables that affected outcomes and the least number of variables that produced the best prediction.
New Model vs IPSS-R
IN A HEAD-TO-HEAD comparison using patient medical records, the new model correctly predicted a patient’s likelihood of survival for a given period relative to another patient 74% of the time (concordance index [c-index] = 0.74). By comparison, the IPSS-R correctly predicted a patient’s likelihood of survival 67% of the time (c-index = 0.67). For leukemia-free survival probability, the c-index with the new model was 0.81, compared with 0.73 for the IPSS-R, he reported.
Similarly, in the 831 patients included in the validation cohort, the algorithm predicted overall survival 80% of the time and leukemia-free survival 78% of the time. The new model was also prognostically superior to models based on genetic mutations alone; mutations plus cytogenetics; and mutations, cytogenetics, and age, he added.
Dr. Nazha is now developing a Web application that will be physician- and patient-friendly. This model can be used to provide probabilities for overall survival and AML transformation for an individual patient at different times during the course of MDS. He and his team plan to improve the model by eliciting feedback from clinicians and by incorporating additional outcomes, such as quality-of-life information.
The Web-based tool is not yet commercially available, but the researchers are working toward that goal. ■
DISCLOSURE: Dr. Nazha is a consultant for Karyopharm and Tolero and is on the data monitoring committee for MEI Pharma.
1. Nazha A, Komrokji RS, Meggendorfer M, et al: A personalized prediction model to risk stratify patients with myelodysplastic syndromes. 2018 ASH Annual Meeting & Exposition. Abstract 793. Presented December 3, 2018.
PRESS BRIEFING moderator Joseph Mikhael, MD, Professor of Applied Cancer Research and Drug Discovery, Translational Genomics Research Institute, City of Hope Cancer Center in Phoenix, noted that traditional models are based on simplicity, and the scoring system contains few variables. “In an era...