Statistical foundations for precision medicine

Abstract

Physicians must often diagnose their patients using disease archetypes that are based on symptoms as opposed to underlying pathophysiology. The growing concept of "precision medicine" addresses this challenge by recognizing the vast yet fractured state of biomedical data, and calls for a patient-centered view of data in which molecular, clinical, and environmental measurements are stored in large shareable databases. Such efforts have already enabled large-scale knowledge advancement, but they also risk enabling large-scale misuse. In this thesis, I explore several statistical opportunities and challenges central to clinical decision-making and knowledge advancement with these resources. I use the inherited heart disease hypertrophic cardiomyopathy (HCM) to illustrate these concepts. HCM has proven tractable to genomic sequencing, which guides risk stratification for family members and tailors therapy for some patients. However, these benefits carry risks. I show how genomic misclassifications can disproportionately affect African Americans, amplifying healthcare disparities. These findings highlight the value of diverse population sequencing data, which can prevent variant misclassifications by identifying ancestry informative yet clinically uninformative markers. As decision-making for the individual patient follows from knowledge discovery by the community, I introduce a new quantity called the "dataset positive predictive value" (dPPV) to quantify reproducibility when many research teams separately mine a shared dataset, a growing practice that mirrors genomic testing in scale but not synchrony. I address only a few of the many challenges of delivering sound interpretation of genetic variation in the clinic and the challenges of knowledge discovery with shared "big data." These examples nonetheless serve to illustrate the need for grounded statistical approaches to reliably use these powerful new resources.