Population and administrative data sources that could be instrumental for studying health disparities are frequently limited by a lack of sophisticated race and ethnicity measurement. Administrative records linking individuals across decades of observation — from school enrollment to death certificates — could tell powerful stories about how racial inequality in risk of premature death evolves over the life course, but sometimes lack basic race information, and even more frequently lack race-adjacent measures, like skin tone, that are known to be instrumental in stratifying health outcomes. Past efforts have addressed this by employing large teams of human coders to assess names and photographs for race and ethnicity. Beyond being prohibitively costly however, this approach is error-prone and overly reliant on point-estimate classifications, a limitation that obscures meaningful variation, given that perceived race is not static but better characterized as a distribution. In this project, we introduce an AI-driven approach to recovering race and skin tone data from archival imagery. The scalability of this method allows us to construct distributions of perceived skin tone and race,  capturing median perceptions alongside rater variation,  rather than collapsing to a single classification. Applied to yearbook photographs linked to mortality records, this approach opens new possibilities for studying how racialization shapes health across the life course.