Machine Learning to Discern Interactive Clusters of Risk Factors for Late Recurrence of Metastatic Breast Cancer

Abstract

Background: Risk of metastatic recurrence of breast cancer after initial diagnosis and treatment depends on the presence of a number of risk factors. Although most univariate risk factors have been identified using classical methods, machine-learning methods are also being used to tease out non-obvious contributors to a patient's individual risk of developing late distant metastasis. Bayesian-network algorithms can identify not only risk factors but also interactions among these risks, which consequently may increase the risk of developing metastatic breast cancer. We proposed to apply a previously developed machine-learning method to discern risk factors of 5-, 10- and 15-year metastases.

Methods: We applied a previously validated algorithm named the Markov Blanket and Interactive Risk Factor Learner (MBIL) to the electronic health record (EHR)-based Lynn Sage Database (LSDB) from the Lynn Sage Comprehensive Breast Center at Northwestern Memorial Hospital. This algorithm provided an output of both single and interactive risk factors of 5-, 10-, and 15-year metastases from the LSDB. We individually examined and interpreted the clinical relevance of these interactions based on years to metastasis and reliance on interactivity between risk factors.

Results: We found that, with lower alpha values (low interactivity score), the prevalence of variables with an independent influence on long-term metastasis was higher (i.e., HER2, TNEG). As the value of alpha increased to 480, stronger interactions were needed to define clusters of factors that increased the risk of metastasis (i.e., ER, smoking, race, alcohol usage).

Conclusion: MBIL identified single and interacting risk factors of metastatic breast cancer, many of which were supported by clinical evidence. These results strongly recommend the development of further large data studies with different databases to validate the degree to which some of these variables impact metastatic breast cancer in the long term.

Keywords: Markov Blanket and Interactive Risk Factor Learner (MBIL); causal learning; machine learning; metastasis; metastatic breast cancer; risk factors.

Figure 1

A BN DAG model illustrating the Markov blanket. The Markov blanket of T consists of nodes X11, X12, X13, X14 and X15. These nodes are the direct risk factors of T and separate T from the influence of the noisy variables X1–X10, X16 and X17 (adapted from [8]).

Figure 2

MBIL-generated causal sets of 5-, 10- and 15-year breast cancer metastases.

Figure 3

The MBIL generated an output of clinical interactions related to 5-, 10- and 15-year breast cancer metastases. HER2, ER, grade, race/ethnicity, TNEG, smoking/alcohol and surgical margins are represented. Each bar plot indicates the number of counts in which each of these variables was identified as a risk factor of metastasis at different values of alpha.