Death is the great certainty, but what predicts it is not only biology. A new study in the American Journal of Epidemiology uses explainable artificial intelligence (XAI) to revisit the social determinants of health across the U.S., U.K., and Europe, showing that the story of mortality is written as much in wealth, education, and social ties as in blood pressure or cholesterol.
Traditionally, research isolates single risk factors - smoking, exercise, income - and measures their impact in neat lines. But life is not neat. Social disadvantage, psychological strain, and health behaviors intertwine. To capture this, Jiani Yan and colleagues from Oxford University applied machine learning across three major datasets: the Health and Retirement Study (U.S.), the English Longitudinal Study of Ageing, and the Survey of Health, Ageing and Retirement in Europe.
Their self-devised algorithm pulled together seven domains - demography, socioeconomic status, psychology, social connections, childhood adversity, adulthood adversity, and health behaviors. The result was striking: at the domain level, demographic (age and gender) and socioeconomic factors consistently dominated mortality prediction. But when the models drilled down to individual variables, the story became more fragmented and context-specific. Smoking mattered more in some countries, vigorous activity in others, occupational status in yet another.
This duality is crucial. Across nations, broad domains of inequality weigh heavily. But within nations, the precise factors shift - shaped by culture, policy, and history. The data showed, for example, that maternal education predicted mortality more strongly in continental Europe than in the U.S., while income and occupational status carried particular weight in the U.K.
Using explainable AI, the researchers went beyond prediction to interpretation. Shapley values revealed how each factor nudged an individual's probability of death higher or lower. This approach demystifies the "black box" of machine learning, making it possible not only to forecast outcomes but to uncover why those outcomes differ across contexts.
The implications extend beyond demography. If health is socially patterned, then policy is medicine. Education, housing, and social connection become as critical as clinical interventions. Explainable AI makes these connections visible, not as abstractions, but as quantified shifts in life expectancy.
What this study demonstrates most powerfully is that death is not just a biological endpoint but a mirror of lived inequality. Machine learning, when explained rather than hidden, becomes a new lens for social epidemiology - one that respects both the universality of mortality and the local textures of human life.
