A computer model developed by Cleveland Clinic Cancer Center researchers and their colleagues who applied machine learning to clinical and genomic datasets is able to predict the survival of individual acute myeloid leukemia (AML) patients more accurately than existing models.
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
The researchers presented their preliminary findings at the 2019 American Society of Hematology annual meeting.
“With the dramatic growth of genomic data and profiling in recent years, there have been a number of attempts to incorporate this information into clinical decision tools,” notes hematologist/oncologist Aziz Nazha, MD, a co-author of the report. “However, the challenge is that current risk-classification models undermine the complexity of genomic alterations, and don’t take into account how particular constellations of genomic and clinical risk factors impact patient outcomes.”
Recognizing the limitations of current systems, Dr. Nazha and Cleveland Clinic internal medicine resident Jacob Shreve, MD, MS, sought to develop a more accurate model that could better utilize genomic data. “We took advantage of a large data set to build a machine-learning model that can provide personalized predictions, specific for a given AML patient,” he says. “We talk about personalized treatment all the time, but we don’t spend too much time discussing personalized prognosis. Many in the field are starting to realize the importance of these other concepts of personalized medicine.”
Development and validation
To create the model, the researchers applied machine learning to five AML datasets that contained 792,779 genomic and clinical data points representing 3,421 patients. The datasets included information about patients from the Beat AML Master Trial, Cleveland Clinic, the Munich Leukemia Laboratory (MLL), the German-Austrian Study Group, and the Cancer Genome Atlas (TCGA).
The researchers used a panel of 44 gene mutations commonly associated with AML in their analysis, as well as a number of cytogenetic and clinical variables. These included age, white blood cell count at diagnosis, and disease subtype.
“We utilized a robust, state-of-the-art machine-learning algorithm that can overcome some of the limitations of traditional statistics,” Dr. Nazha explains. “To determine overall survival (OS), clinical and molecular variables were randomly selected for inclusion.
“And then we asked, can we extract the key variables that impacted the algorithm decision? This is very important to ensure the variables are clinically relevant. This is also a crucial step so that we can learn something new from the machine-learning algorithm that we typically aren’t able to when using traditional statistics.”
The researchers used feature-extraction algorithms to isolate the most important variables that impacted decision-making. They used the Concordance index (C-index) to evaluate the new model’s predictive accuracy compared to European LeukemiaNet (ELN) risk classification.
Among the overall study cohort, 1,122 patients (32.8%) had favorable risk cytogenetics according to ELN criteria, 956 (27.9%) intermediate and 1,343 (39.3%) adverse. The most commonly mutated genes included NPM1 (24%), FLT3 (23%), DNMT3A (20%), NRAS (13%), IDH2 (11%), RUNX1 (10%) and TET2 (10%).
The researchers determined that the following variables impacted OS: age, transplant status, white blood cell count, bone marrow blast %, cytogenetics and mutations in ASXL1, CEBPA, DNMT3A, FLT3, KDM6A, KIT, KRAS, NPM1, NRAS, PHF6, PTPN11, RUNX1, TET2, and TP53.
Data showed that the C-index for the new model was 0.80 compared to 0.59 for the ELN classification system. When the new model was applied to the five patient cohorts, the researchers observed that each of their c-indices remained higher than that of the ELN [Beat AML (0.81), Cleveland Clinic (0.85), MLL (0.83), German-Austrian Study Group (0.79), and TCGA (0.80)].
“We found that our model significantly outperformed the ELN classification,” reports Dr. Nazha. “To our knowledge, this is the highest C-index reported of any recently developed model. This finding validates the model’s ability to accurately predict survival outcomes based on clinical and mutational variables.”
Clinical implications, next steps
This new model has significant implications for the future of clinical decision-making, offering oncologists a tool that can more accurately predict an individual AML patient’s risk and outcomes.
“Our treatment guidelines are typically based on risk,” notes Dr. Nazha. “And so, if we identify risk correctly, then we give the right treatment; however, if we don’t, then the chosen treatment approach may not be the most effective.
“Generally speaking, patients with favorable AML are not offered allogeneic stem cell transplants after they achieve their remission,” he continues. “Whereas, transplant is recommended for patients with unfavorable-risk AML.”
Treatment choices for the intermediate-risk patient population are not as clear-cut, according to Dr. Nazha. “By providing this personalized prediction, we hope this model will improve the decision-making process between patients and physicians.”
To further enhance the model, Dr. Nazha and his team are now increasing the number of patient cohorts and adding more analyses.
“We are working to refine the model to find more interactions between the clinical and genomic data that will allow us to further understand not just the impact of those mutations on prognosis, but also on the biology of the disease,” he explains.
In an effort to ease the translation of their model into the clinical setting, the team also is developing a web application that will be available to the public. The goal is to present the final model at the 2020 American Society of Hematology annual meeting and later integrate it into clinical practice.
“By utilizing novel, state-of-the-art artificial intelligence we are addressing an important clinical problem with the development of a clinical decision-making tool that in the future can aid physicians in providing better treatment to their patients,” Dr. Nazha concludes.