Information-Theoretic Surprise-Driven Approach to Enhance Decision Making in Healthcare

点击翻译按钮获取中文摘要

PROJECT SUMMARY Traditional guidelines used in clinical decision-making alone are often insufficient in accurately stratifying patients for diagnostic testing. For instance, the National Comprehensive Cancer Network (NCCN) guidelines for stratifying patients for germline genetic testing of breast cancer fails to identify nearly 50% of the patients who might have a BRCA mutation. With advancements in imaging techniques and black-box machine learning algorithms, radiomics has emerged as a promising tool for making predictions in a wide range of health conditions such as breast and ovarian cancer, Alzheimer’s disease, and coronary heart disease. Based on these studies, the underlying hypothesis of this research is that imaging phenotypes obtained from radiomics together with traditional guidelines can screen patients for underlying diseases (in this research, germline BRCA mutation) with a higher positive predictive value than the traditional guidelines alone. Here, we refer to the NCCN guidelines as the traditional guidelines. However, little is known about the causal relationship between these deleterious health conditions and quantitative imaging phenotypes. Together with the lack of standards for quantifying and reporting imaging phenotypes across multiple institutions, it is currently not feasible to integrate them into clinical decision-making. To this end, this research will focus on the following two specific aims to address these challenges and subsequently validate the hypothesis. Specific Aim 1: MRI harmonization via amplitude synchronization to mitigate the scanner-to-scanner variability. Specific Aim 2: Causal inference and information theory to discover the causal relationships between BRCA mutation and imaging phenotypes and subsequently integrate them into clinical decision making. While the proposed research focuses on stratifying patients for germline BRCA testing based on magnetic resonance imaging phenotypes, the methodology and algorithms generalize to other health conditions and imaging modalities. The outcome of this research will lead to a new paradigm of clinical decision making where medical practitioners would be able to link imaging phenotypes with underlying health conditions—akin to how abnormal levels on comprehensive metabolic panels act as indicators of potential health problems—and prepare for appropriate interventions.