Sport-Inspired Risk Model Improves Cancer Risk Prediction

Liam Davenport

July 16, 2019

A sport-inspired dynamic risk assessment model that combines multiple risk factors to predict a cancer patient's outcome may be able to determine the patient's risk of progression better than current methods, say researchers reporting an international study.

Ash Alizadeh, MD, PhD, associate professor of medicine (oncology) at Stanford University, California, and colleagues took their inspiration from the kinds of 'win-probability' assessments used to predict the outcome of a soccer match or an election.

Using a Bayesian approach to combining not only multiple individual risk factors into a single model but also their conditional probabilities, they developed a Continuous Individualized Risk Index (CIRI) model, which they then tested in three types of cancer.

Focusing initially on patients with diffuse large-B cell (DLBC) lymphoma, the researchers were able to improve significantly on predicting event-free survival at 24 months over any factor alone, including an established risk prediction score.

The team also tested the model in patients with chronic lymphocytic lymphoma (CLL), and found that as well as improving on predictions of progression-free and overall survival, it even indicated which patients would respond best to an individual therapy.

Similar results were obtained when the researchers looked at historical data from patients with breast cancer.

The research was published online July 4 in the journal Cell.

Coauthor Maximilian Diehn, MD, PhD, from the Institute for Stem Cell Biology and Regenerative Medicine at Stanford, explained that currently used prediction models are based on one set of findings.

"What we're doing now is somewhat like trying to predict the outcome of a basketball game by tuning in at halftime to check the score, or by watching only the tipoff," he commented in a statement.

"In reality we know that there are any number of things that could have happened during the first half that we aren't taking into account."

Coauthor David M. Kurtz, MD, PhD, from the division of oncology at Stanford, said, "Our standard methods of predicting prognoses in these patients are not that accurate."

"Using standard baseline variables it becomes almost a crystal ball exercise," he said, adding that the new model is "markedly better than we've done in the past."

Alizadeh agreed, saying, "What I didn't expect was that aggregating all this information through time may also be predictive."

"It might tell us 'you're going down the wrong path with this therapy, and this other therapy might be better.' Now we have a mathematical model that might help us identify subsets of patients who are unlikely to do well with standard treatments," he said.

The CIRI is available online for clinicians to use.

However, Alizadeh cautioned that, because the team "did not study a prospective design to use these measures to change direction," they were initially "very resistant" to putting the tool online. But during the review process, they were persuaded that there was a case for doing so, he told Medscape Medical News.

Alizadeh believes that "with tools like this in clinicians' hands, and this website being available, I'm hoping we can do those studies to change therapy based on projected outcomes or perhaps even more importantly to consider giving patients less therapy than might be harmful."

Linking Data From Several Time Points

Alizadeh, who is both a cancer geneticist and medical oncologist, explained that diffuse large-B cell lymphoma was the starting point for their study.

"It's a fascinating disease, he said. "It's among the most common blood cancers; it's also among the most curable, with the majority of patients treated with a combination regimen that we administer over the course of about 5 to 6 months."

Nevertheless, one third of patients will likely succumb to the disease, and the issue of identifying patients who will do well vs those who will progress has been the subject of a great deal of work.

"We have been working on this problem for some time," Alizadeh said, "building 'crystal balls' that we apply at diagnosis, and then at defined milestones — let's say after a certain duration of time has passed — but not connecting them to each other."

Taking inspiration from 'win-probability' models developed to refine sports, political, and even meteorological predictions, the researchers decided to "take more of a gestalt statistical approach" to cancer outcomes, he said.

Alizadeh continued: "We have a crystal ball at time of diagnosis, we have crystal balls at milestones during therapy, so what if we link these the same way as when you're watching, say, a football match, or an election, so that accumulative data over the course of time can be used to refine that prediction for the future?"

"We don't really do that in medicine or oncology in any exercise, despite it being particularly popular in other fields."

To investigate further, they constructed a dynamic risk model for integrating diverse outcome predictors into a single quantitative risk estimate over the course of disease.

They used two approaches to risk modeling:

  • A naive Bayesian framework to estimate the probability of a clinical outcome at a defined endpoint in time

  • A personalized probability of survival over time based on Cox proportional hazard modeling

The first approach, which combines the initial probability of an adverse outcome with the conditional probabilities of a series of risk factors, was initially tested on 2558 patients with diffuse large-B cell lymphoma from 11 previously published studies.

The second approach involving giving Cox modeling coefficients a "prior probability" derived from available univariate analyses to develop a so-called Bayesian Cox model.

This was also tested on the diffuse large-B cell lymphoma population, as well as being applied to data from three phase 3 clinical trials on chronic lymphocytic leukemia, and a study of neoadjuvant chemotherapy in patients with resectable breast adenocarcinoma.

The authors said that they chose these cancers because they "have diverse, established outcome predictors, including biomarkers assessed after therapy by noninvasive or invasive means."

Lymphoma Study

For diffuse large-B cell lymphoma, the team relied on three established clinical, radiographic, and molecular risk factors, including the International Prognostic Index (IPI), and three risk factors derived from circulating tumor DNA (ctDNA), or liquid biopsies.

Predicting event-free survival at 24 months (EFS24) in 132 patients with data on all six factors available, they were able to show that the CIRI performed significantly better than any factor alone (P < .05).

The authors note that the strength of the approach is that the probability can be updated over time.

Comparing two typical patients, they say that, if one of them achieves a favorable early molecular response (EMR) at cycle 2 of treatment, their probability of an EFS24 increases.

If the other does not achieve an EMR at the same time point, their probability of an EFS24 decreases.

While the tool was initially designed to predict only EFS24, the team found that the CIRI was able to significantly improve on the prediction of overall survival at 24 months vs the IPI (P = .009).

Chronic Leukemia Study

When the researchers turned to chronic lymphocytic leukemia, they combined pretreatment risk factors, interim risk factors during treatment, and those seen at the end of the treatment period.

Examining the ability of the CIRI to predict progression-free survival at 12, 24, 36, and 48 months, their model was found to outperform the International Prognostic Index for Chronic Lymphocytic Leukemia (CLL-IPI) and interim minimal residual disease (MRD) status at all time points (P < .05).

The team was also able to use the results to stratify patients into risk groups based on their likelihood of progression-free survival, across all timepoints and specifically at the end of treatment.

Again, the CIRI was able to improve on the prediction of overall survival obtained with the CLL-IPI and MRD assessments alone at all times after 12 months, increasing the C-Statistic from 0.72 to 0.80 (P < .001).

Breast Cancer Study

For breast cancer outcome prediction, the team used clinical stage, tumor grade, estrogen receptor and HER2 status, and pathologic response to neoadjuvant chemotherapy.

As for the other diseases, the CIRI model outperformed predictions based on pretreatment or pathologic response factors separately for predicting distant relapse-free survival at 12, 24, 36, 48, and 60 months (P < .05).

The model was also able stratify patients into risk groups for distant relapse-free survival both postsurgery and across all timepoints.

Predicting Which Patients Would Benefit From Therapy 

The authors found that, in the case of chronic lymphocytic leukemia, the CIRI model may even be able to predict which benefits will benefit from a treatment following initial induction therapy.

For example, MRD was able to distinguish between patients who had superior and less optimal outcomes with fludarabine, cyclophosphamide, and rituximab (FCR). However, it suggested that all patients would benefit from the combination over alternatives.

In contrast, CIRI "provides each individual patient with a quantitative estimate of the probability of disease progression at 36 months, based on the combination of CLL-IPI and interim MRD," the authors write.

The model suggested that patients identified as high risk on CIRI, defined as a greater than 20% risk of progression at 36 months, would have a significantly increased progression-free survival if they were given FCR than those who were not (P < .0001).

In contrast, the model suggested that CIRI low-risk patients would derive no benefit from FCR therapy.

Combine With Info From Liquid Biopsies?

One area that could potentially be explored in the future for adapting the CIRI model is to expand on the use of information gathered from information collected in so-called liquid biopsies, such as levels of circulating tumor ctDNA in blood samples.

Currently, the model incorporates pretreatment ctDNA levels, early molecular response, and major molecular response.

In doing so, "we're looking at the amount of disease that's left at each milestone, and projecting the future based on those measurements," Alizadeh explained, asking: "How quickly did it change?"

"For lymphoma for example, we're looking…about one sixth or one third of the way through the course of treatment before forecasting the future. In breast cancer, we're looking at the end of treatment. In leukemia, we're looking in the middle of treatment."

He continued: "What we're not looking at, which I think is an interesting idea to try to incorporate into these models, is not just how much did the disease change, but how did the disease change?"

"Were there clonal selections or evolution observed during that time that you could then add on top of [current measures]?"

Alizadeh said, "My sense is that if you observe those kinds of things that they will make these predictions even more refined."

The research was supported by the National Institutes of Health, the Damon Runyon Cancer Research Foundation, the American Society of Hematology, the Leukemia and Lymphoma Society, the V Foundation for Cancer Research, the Conquer Cancer Foundation of the American Society of Clinical Oncology, the Emerson Collective Cancer Research Fund, a Stinehart/Reed Award, the Shanahan Family Fund and the Ludwig Institute for Cancer Research.

Alizadeh, Diehn, and coauthor Aaron Newman are cofounders of Palo Alto-based CiberMed Inc, a biomarker discovery company.

Cell. Published online July 4, 2019. Full text

For more from Medscape Oncology, follow us on Twitter:  @MedscapeOnc


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as: