Robust Inference and Uncertainty Quantification for Data-Driven Decision Making:

The seminar explores the use of machine learning, optimization, and statistics for data-driven decision-making. It highlights two recent studies: one on statistical inference in the presence of selection bias and distribution shifts, and another on uncertainty quantification in machine learning models. Specifically, it examines how machine learning models can provide reliable information for decision-making in fields such as public health and policy, where observed data may be biased. The first study focuses on statistical inference under selection bias, illustrated through the analysis of COVID-19 hospitalizations across different ethnic groups in the U.S. It shows how hospitalization data are influenced by socioeconomic factors and access to healthcare, which introduces bias in estimating hospitalization risk for certain groups. To correct for this, the study proposes a robust model based on constraints derived from aggregated population data (e.g., census and seroprevalence studies), enabling more trustworthy estimates. The results reveal that even after adjusting for age and income, minorities still faced higher hospitalization risks, highlighting real disparities in health outcomes. The second study addresses uncertainty quantification in clinical prediction models, particularly in the diagnosis of dermatological diseases. It discusses the limitations of poorly calibrated machine learning models that produce outputs difficult for physicians to interpret. The study introduces a Conformal Prediction approach that, instead of providing a single diagnosis, generates a set of possible diagnoses with statistical coverage guarantees. Moreover, the method is optimized to ensure these sets are clinically useful by grouping similar conditions rather than presenting heterogeneous lists. This approach improves the practical usability of machine learning in medicine, offering more robust interpretations and supporting more effective clinical decision-making.