Transforming Mortality Prediction: A Transformer-based Mortality Prediction Model.

Abstract

Objectives: Mortality prediction and the identification of mortality risks are central to social and biological sciences. Traditional models often assess linear associations between single risk factors and mortality. Transformer models, capable of capturing long-term dependencies across multiple variables, offer a novel approach to mortality prediction. This study introduces a transformer-based model applied to data from the Health and Retirement Study (HRS).

Methods: We analyzed data provided by 38,193 adults aged ≥50 years participating in the HRS, a longitudinal US study surveyed biennially since 1992. Linked mortality data were obtained from the National Death Index and postmortem interviews. Using the transformer architecture, we modeled changes in 126 risk factors spanning financial, physical, and mental health domains manifesting over 29 years. Prediction performance was assessed across multiple settings, with traditional statistical and machine learning models serving as benchmarks.

Results: Over a median follow-up of 9 years, 17,448 deaths occurred (crude rate: 39.6 per 1,000 person-years). The transformer model consistently outperformed traditional and machine learning methods, achieving a twofold improvement in average precision scores (APS) for next-wave mortality prediction relative to the best benchmark model.

Discussion: Transformer-based models, such as BEHRT, significantly enhance mortality prediction compared with traditional approaches. These findings highlight the potential of transformer neural network models in social science-focused population health research on aging.