Leveraging Deep Learning and Bayesian Networks to Identify Risk Factors and Support Personalized Prediction for Metastatic Breast Cancer
Abstract
Breast cancer is one of the leading causes of cancer death in US women. It is estimated that 40,920 US women will die from breast cancer in 2018. Breast cancer is also one of the major causes of cancer-related death in women globally, and it is estimated that without major changes in prevention or treatment, 846,241 women will die from breast cancer worldwide in 2035. Women do not die of breast cancer confined to the breast or draining lymph nodes; rather, they die mainly due to metastasis, when cancer spreads to other vital organs such as the lung and brain. Metastatic breast cancer (MBC) is the cause of over 90% of breast cancer-related deaths and remains a largely incurable disease. Although most newly diagnosed breast cancer cases are not metastatic, all patients are at risk of developing metastatic cancer in the future, even if they are free of cancer for years after the initial treatment. Being able to predict effectively for each individual patient the likelihood of metastatic occurrence is important because the prediction can guide treatment plans tailored to a specific patient to prevent metastasis and to help avoid under- or over-treatment. Researchers have established a few risk factors for MBC through epidemiologic studies, but these risk factors have not proven to be effective in predicting an individual’s risk of developing metastasis. Therefore, identifying risk factors for MBC and helping improving personalized prediction of breast cancer metastasis continues to be a major research imperative. The purpose of this proposal is to explore new avenues for leveraging artificial intelligence to facilitate the improvement of our capabilities in detecting risk factors and making patient-specific prediction for breast cancer metastasis. When identifying risk factors researchers ordinarily just look for correlation of patient features with the outcome. A problem with this approach is that only some of the risk factors directly affect the outcome. For example, menopause status and age might both be associated with metastasis. However, menopause status might be the direct predictor, while age is associated with metastasis only because it is correlated with menopause status. A second problem is that when we investigate single risk factors, we miss risk factors that only exhibit a substantial influence on the outcome when taken together. That is, they interact. A simple example illustrates this type of interaction. Suppose a strong worker uses ¾ hour to remove a pile of snow, while a weak one uses 2 hours. Then if we assume two weak workers do not interact, we would conclude it would take two weak workers 1 hour to shovel the snow and prefer to hire the strong worker if time is an imperative. However, perhaps the two weak workers can interact effectively by having the faster one dump the snow while having the stronger slower one focus on shoveling the snow to the cart and by so doing complete the job in ½ hour. After investigating this interaction, we would hire the two weak workers. Patient features interact similarly. For example, it is known that HER2 status and ER status interact to affect metastasis, and so their joint effect cannot be learned by looking at them individually. Some interacting features could show no correlation when considered individually. Our research will overcome these problems. First, we will develop methodology that learns from data which features interact to affect breast cancer metastasis, and then treat such interactions as single features. Then we develop a technique that learns which features are the direct predictors of metastasis. Once we have these direct predictors including interacting direct predictors, we will employ state-of-art artificial intelligence to use the predictors to predict breast cancer metastasis for each individual patient. If this project is successful, researchers will have an accurate method that can be used to determine risk factors
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 19, 2019
- Source ID
- W81XWH1910495
Entities
People
- Xia Jiang
Organizations
- United States Army
- University of Pittsburgh