Death Predictions from Social Determinants: Holistically and Precisely with Explainable AI

Jiani Yan , University of Oxford

Existing literature often examines death from a narrow perspective, largely at the individual level, leaving a comprehensive, interdisciplinary approach understudied. Using advanced data linkage of ageing including the Health and Retirement Study in the US (HRS), the Survey of Health, Ageing and Retirement in Europe (SHARE) and the English Longitudinal Study of Ageing in the UK (ELSA), we aim to bridge the knowledge gap by assessing the predictability of death using cutting-edge machine learning and explainable AI algorithms which integrates explanation and prediction simultaneously. Specifically, we extract information from all datasets in seven health-related domains including Adulthood Psychological Diathesis, Adulthood Socioeconomic, Childhood Adversity, Adulthood Adverse Experience, Health Behaviours, Social Connections and Demography. We then construct a predictive research design which allows us to accurately consider how predictable death is and surface the most important risk factors at both the single risk factor level and domain levels, which we uniquely re-engineer, across multiple countries. Furthermore, our study delves into two methodological challenges in employing machine learning for social science research, exploring the influence of information scale (variable availability and data scope) and quantifying the ramifications of different seed choices on model efficacy.

See extended abstract

 Presented in Session 2. Machine Learning Approaches for Population Research