Redefining Normal in Clinical Medicine

Abstract

In recent years, a central debate in clinical medicine has concerned how to personalize care - specifically, which patient-level features should inform treatment decisions. This thesis examines the tension between individualized and population-based approaches, leveraging advances in machine learning to enable more robust, data-driven clinical decision-making. Chapter 1 explores the practice of race adjustment in the clinical reference equations that define normal ranges, a practice that has drawn criticism in recent years. Concerns that such adjustments serve as an imprecise proxy have fueled ongoing debate and a shift toward race-neutral equations, which often rely on simplistic approaches that fail to capture the factors driving group differences. To address this, I developed a data-driven framework to evaluate alternative proxies, including anthropometrics, sociodemographic factors, and environmental exposures. Applying this framework to spirometry, I show that anthropometric measures account for a substantial portion of the variation traditionally attributed to race. Replacing race in reference equations with these individual-level factors improves predictive accuracy and generalizability across large observational cohorts, supporting a transition from coarse demographic groupings to directly measured, patient-specific drivers of variation. Chapter 2 examines how the shift toward precision medicine reframes clinical interpretation from population-level norms to individualized baselines, with a focus on routine blood biomarkers. While individual histories capture within-patient variation, purely individualized approaches risk overfitting and generating excess false positives, potentially triggering unnecessary clinical evaluation and patient burden. Using nearly two billion longitudinal laboratory measurements from over 1.5 million adults, I show that individualized thresholds frequently classify results as abnormal without corresponding increases in adverse outcomes, while population-based thresholds miss early change. I developed NORMA, a model that conditions on both patient trajectories and population-level expectations. Across clinical settings, NORMA detected abnormalities months earlier while improving precision for predicting acute events, chronic disease progression, and mortality, demonstrating the value of anchoring individual trajectories to population-level priors. In Chapter 3, I show that these challenges are further amplified in models that incorporate language as part of the input. With the rapid adoption of multimodal foundation models, diagnostic decisions are increasingly mediated by systems that are flexible but less transparent than traditional approaches. I examine the \emph{steerability} of multimodal vision--language models across dermatology, radiology, and histology tasks, finding that diagnostic outputs vary systematically with prompt structure, image characteristics, and patient subgroups, with differences large enough to alter clinical classifications. These findings demonstrate that model outputs are context-dependent and less directly interpretable, shaped by both the data and how the task is specified. Finally, Chapter 4 examines how these dependencies are most consequential in contexts with limited access to care, where the relevant comparison is not idealized clinical practice but existing conditions. I show that evaluation must account for deployment context, as systems that are imperfect in absolute terms may still improve access to diagnosis and treatment. Together, these results show that clinical decision boundaries are shaped by modeling choices and context, and that making these dependencies explicit is necessary to build systems that are accurate, generalizable, and clinically useful in practice.