This article is more than 5 years old. Information may no longer be current.
Novel risk stratification model outperforms standard prediction tools for myelodysplastic syndrome
Click Here to Manage Email Alerts
SAN DIEGO — A personalized risk prediction model developed using machine learning showed utility as a predictive and prognostic tool among patients with myelodysplastic syndrome, according to findings presented at ASH Annual Meeting and Exposition.
Results showed the new model better predicted mortality and transformation to acute myeloid leukemia than the standard International Prognostic Scoring System (IPSS) and Revised IPSS (IPSS-R).
Appropriate risk stratification is essential for this heterogenous patient population, Aziz Nazha, MD, medical hematologist and oncologist at Cleveland Clinic, said during his presentation.
“The survival for these patients can sometimes be measured over years but, for some of them, it’s over months,” he said.
However, he acknowledged that current risk stratification models may be insufficient.
“You can find some patients with low risk who behave worse than other patients with high risk,” he said. “This has made us ask the question of whether we can build a model that can provide personalized prediction for a given patient.”
It is also important to note the relationship between risk stratification and treatment plan, with chemotherapy and hematopoietic stem cell transplantation used for patients with very different risk profiles, according to Nazha.
Nazha and colleagues “harnessed the power of artificial intelligence and machine learning” to develop a model that included individual patient genomic and clinical data, Nazha said. The researchers then entered the data into a machine-learning algorithm and a web application with the aim of investigating OS and transformation to AML probability at various time points.
The analysis included 2,302 patients, including 1,471 (median age, 71 years; range, 19-99) in the training cohort and 831 in the validation cohort.
Sixteen percent of patients in the training cohort had progressed to AML, 11% had secondary myelodysplastic syndrome and 9% underwent HSCT.
According to IPSS criteria, 27% were low risk, 43% intermediate-1 risk, 19% intermediate-2 risk and 7% high risk. According to IPSS-R, 51% were very low/low risk, 23% intermediate risk, 13% high risk and 6% very high risk. Seven percent of patients had missing risk data.
Using their model, researchers identified the following characteristics — ranked in order —as most impactful on OS: cytogenetic risk categories by IPSS-R; platelets, mutation number, hemoglobin, bone marrow blasts percentage, 2008 WHO diagnosis, white blood count, age, absolute neutrophil count, absolute lymphocyte count, TP53, RUNX1, STAG2, ASXL1, absolute monocyte counts, SF3B1, SRSF2, RAD21, secondary vs. de novo myelodysplastic syndrome, NRAS, NPM1, TET2 and EZH2.
The group compared the ability of the new model to predict OS and transformation to AML compared with the IPSS and IPSS-R.
Results showed that the concordance index for the novel scoring system was 0.74 for OS and 0.81 for transformation to AML.
By comparison, the IPSS carried a concordance index of 0.66 for OS and 0.73 for AML transformation. For the IPSS-R, the C-index was 0.67 for OS and 0.73 for AML transformation.
“Compared with IPSS and IPSS-R, our model significantly outperformed these models,” Nazha said.
Nazha further reported that for OS, their model was prognostically superior to models based on genetic mutations only (C-index = 0.64), mutations plus cytogenetics (C-index = 0.68), and mutations, cytogenetics and age (C-index = 0.69).
Regarding AML transformations, the novel model bested models that considered mutations only (C-index = 0.72), mutations plus cytogenetics (C-index = 0.74), and mutations, cytogenetics and age (C-index = 0.75).
In the validation cohort, the novel method yielded a C-index of 0.8 for OS and 0.78 for AML transformation.
“With a machine learning approach and artificial intelligence, we are able to achieve much higher accuracy in terms of OS prediction as well as leukemia-free survival,” Nazha said. – by Rob Volansky
Reference:
Nazha, et al. Abstract 793. Presented at: ASH Annual Meeting and Exposition; Dec. 1-4, 2018; San Diego.
Disclosures : Nazha reports consultant roles with Karyopharma and Tolero, and being on the data monitoring committee of MEI Pharma. Please see the abstract for the full list of author disclosures.
Perspective
Back to Top
Anna B. Halpern, MD
The authors of this abstract created a new prognostic model for patients with myelodysplastic syndrome that incorporates molecular and genomic features, which previously have not been included in our mostly commonly used prognostic model, the IPSS-R score. The results are quite interesting and important for clinical practice.
When I think about the utility of risk prediction models, I think about two things; one is whether they are clinically applicable/practice changing — given the availability of a useful, and widely used, model already — and easily applied in clinic. This depends on the number of variables the model includes as well as how widely available those variables are to clinicians. For example, some of the mutational data they included are readily available in large, academic institutions, but clinicians in community centers may not have access to a 40-gene panel.
The second feature I look for in a prediction model is how good it is at predicting the outcome it’s trying to predict. In this case, the model is predicting OS and transformations to acute leukemia.
In terms of how this model fits these two criteria, what they are trying to do is improve upon our most standardly used models: IPSS and IPSS-R scores. These models are widely used in clinic and in research studies.
This model from Nazha and colleagues strives to incorporate molecular data, which are increasingly more available and data are suggesting they contribute strongly to prognosis. So, I think this model is very useful in personalizing molecular data into a prognosis for each patient. The second criteria for a good model — how well does the model predict what it actually aims to — is measured by the C-statistic, which looks reasonable for this model for both survival and transformation to AML. My one concern regards how easy it is to use, as based on the abstract it is a little unclear exactly which data are included in this model.
Perspective
Back to Top
Joseph Mikhael, MD
How is that I can walk around San Diego and look down at my cellphone, and it suggests to me where I should eat dinner based on where I’ve eaten dinner over the last 3 months. Yet, when I’m in my clinic and I’m trying to choose what’s best for my patients, we have these archaic-like methods. That concept is going to change. It’s already changing. We’re going to be taking that big-data approach to information technology and applying it to provide the best care of our patients. I can be in my clinic and look at two patients who, on paper, might be very similar. They may be the same age or have the same diagnosis, but we know the disease is vastly different in those patients. It does us a disservice if we treat them in the exact same way. Using this machine learning, Dr. Nazha and his team were able to develop a system that outperforms our current prognostic score, the IPSS. It accurately predicts the patient’s likelihood of surviving a given the length of time and the progression to acute leukemia. Most of our scoring systems are based on two or three variables. What we’re trying to demonstrate today is if scoring can be enhanced by a web-based tool or app, we can include more variables.
Click Here to Manage Email Alerts