Using Bayesian Networks to Predict Survival Outcomes: New Case Study
Earlier this month, my colleagues at Cytel Canada published a paper in JCO Clinical Cancer Informatics, offering a proof-of-concept for a Bayesian Network Model, that successfully predicts safety and survival outcomes in patients with metastatic renal cell carcinoma (mRCC). Their work demonstrates that a Bayesian Network Model can make predictions with accuracy comparable to that of other machine learning methods, but with greater explanatory power, thus avoiding the concerns that often swirl around the proverbial "black box." In doing so it also innovatively uses Bayesian Network Models to show how such a quantitative tool can offer clinicians another tool to predict patient outcomes.
Understanding Bayesian Network Models
A Bayesian Network Model can help clinical researchers understand the statistical relationship between a large number of variables including patient characteristics, treatment characteristics, patient outcomes and more. These relationships are organized into a visual format called a directed acyclic graph (DAG) with nodes that reflect the variables and lines that encapsulate conditional probabilities. These probabilities are determined using a dataset (in this case the CheckMate 025 Data Set) and essential machine learning techniques. (The Bayesian Network Model developed for the JCO Clinical Cancer Informatics paper is depicted below.)
Oftentimes machine learning techniques will establish relationships between variables but struggle to explain these relationships, leading to cynicism about their scientific validity. In the published study, scientists at Cytel then tested the ability of the Bayesian Network Model to make predictions about patient-level outcomes, using another dataset, the IMDC Data Set.
Identifying Prognostic Variables
The Bayesian Network Model condenses a number of statistical relationships into a single mathematical entity, including information about biomarkers, treatment and outcomes. In doing so, these models can potentially also identify interpretable prognostic variables; prognostic variables that can be used by clinicians for improved treatment. Therefore, in addition to constructing a scientifically valid model, there is potential for immediate benefit within clinical practice and medical decision-making.
The Broader Context of Clinical Research
Clinical trials in oncology are oftentimes the most complex in terms of clinical trial design. Advances in statistical genetics and biomarker-driven research means that new treatments can often apply to only very small subsets of a population. Any information a clinical trial designer has to understand the underlying relationships between any and all of these biomarker subgroups, and their responses to treatment, can inform more efficient and ethical medical discovery. While Bayesian Network Models are not currently in popular use in clinical trial design, the potential to leverage insights from such models is highly promising.
About the Author:
Dr. Esha Senchaudhuri is a research and communications specialist, committed to helping scholars and scientists translate their research findings to public and private sector executives. At Cytel Esha leads content strategy and content production across the company's five business units. She received a doctorate from the London School of Economics in philosophy, and is a former early-career policy fellow of the American Academy of Arts and Sciences. She has taught medical ethics at the Harvard School of Public Health (TH Chan School), and sits on the Steering Committee of the Society for Women in Philosophy's Eastern Division, which is responsible for awarding the Distinguished Woman in Philosophy Award.