Patients with myelodysplastic syndromes (MDS) have widely variable survival outcomes, ranging from months to more than a decade. Oncologists often use prognostic scoring systems such as the Revised International Prognostic Scoring System (IPSS-R) to predict how long patients may survive. The IPSS-R categorizes patients into one of five groups, from very low risk to very high risk, based on risk of mortality and transformation to acute myeloid leukemia (AML).
Cleveland Clinic is a non-profit academic medical center. Advertising on our site helps support our mission. We do not endorse non-Cleveland Clinic products or services. Policy
Such models, however, are often not very precise. “We often found a significant gap between what we predicted in terms of survival for our patients, based on those models, and what actually happened to our patients,” says Aziz Nazha, MD, hematologist and medical oncologist at Cleveland Clinic Cancer Center.
“So that triggered us to think, ‘Can we do a better job predicting the actual survival of the patient?’ Because prognosis in oncology, I would argue, is the most important part of our job. All oncology patients want to know how long they are going to live.”
Dr. Nahza and colleagues decided to create a new system that incorporates individual patient genomic and clinical data using a machine-learning algorithm to better predict an individual patient’s outcome.
Their system, which they presented recently at the American Society of Hematology meeting, outperformed other models for overall survival (OS) and acute myeloid leukemia (AML) transformation.
To create the new model, Dr. Nazha and his colleagues gathered data from the records of a cohort of 527 Cleveland Clinic patients with MDS. They applied next generation sequencing to the patients’ DNA to find 60 gene mutations commonly mutated in myeloid malignancies. The top commonly mutated genes were: SF3B1 (14 percent), ASXL1 (13 percent), TET2 (12 percent), SRSF2 (11 percent), DNMT3A (11 percent), STAG2 (9 percent), TP53 (8 percent), and RUNX1 (8 percent).
From there, they combined all clinical variables and mutations that were present in at least five patients into a “random survival forest” algorithm. The algorithm has many decision trees that sort through data and decide into what survival category the individual patient should be sorted. Finally, the researchers used data from a cohort of 448 patients from five large institutions to validate their model.
The model had better predictability than other models. For instance, the C-index for it was 0.71 for OS and 0.76 for AML transformation compared to 0.67 and 0.73 for the IPSS-R, 0.65 and 0.72 for WHO prognostic scoring system and 0.65 and 0.70 for the MD Anderson prognostic model.
Dr. Nazha says the next step is to build a website to make it easier for clinicians to use their system. “This will allow them to input a patient’s clinical and genetic characteristics and get back the patient’s survival curve as well as the patient’s probability of surviving at different time points such as six months, 12 months, and 18 months, etc.”
First-of-its-kind research investigates the viability of standard screening to reduce the burden of late-stage cancer diagnoses
Study demonstrates ability to reduce patients’ reliance on phlebotomies to stabilize hematocrit levels
Findings highlight an association between obesity and an increased incidence of moderate-severe disease
Cleveland Clinic Cancer Institute takes multi-faceted approach to increasing clinical trial access
Key learnings from DESTINY trials
Gene editing technology offers promise for treating multiple myeloma and other hematologic malignancies, as well as solid tumors
Study of 401,576 patients reveals differences in cancer burdens as well as overall survival
Enfortumab plus pembrolizumab reduced risk of death by 53% compared with platinum-based chemotherapy