Tips for navigating through a sea of high-dimensional data and a jungle of powerful analytic tools to arrive at a clinically useful test

Abstract:

A clinically useful test is one that can aid in guiding medical decisions that lead to better health outcomes. Increasingly, high-dimensional data underlie the development and use of these tests, including data generated from omics assays, imaging, and wearable monitoring devices. Powerful data analysis tools such as machine learning methods, which are capable of processing large volumes of high-dimensional data to develop complex predictors of phenotypes or outcomes, have led to optimism that medical test development could be greatly accelerated. While certainly there have been some success stories, the number of clinically useful tests arising from these efforts has been relatively few given the investment of resources and number of publications claiming to have developed useful predictors. Common pitfalls in the design, analysis, and interpretation of studies aiming to develop medical tests based on complex predictors applied to high-dimensional data are highlighted through a series of examples in oncology. The goal of this discussion is to promote more critical thinking in the assessment of existing evidence for medical tests based on these types of predictors and the planning of new studies. Recommendations are offered to improve the design, conduct, reporting, and value of research aiming to develop medical tests based on complex predictors applied to high-dimensional data.