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

Area of Science:

• Social and biological sciences
• Population health research
• Gerontology

Background:

• Mortality prediction is crucial in social and biological sciences.
• Traditional models often analyze linear associations between single risk factors and mortality.
• Transformer models offer a novel approach by capturing long-term dependencies across multiple variables.

Purpose of the Study:

• Introduce a transformer-based model for mortality prediction.
• Apply the model to data from the Health and Retirement Study (HRS).
• Evaluate the model's performance against traditional and machine learning benchmarks.

Main Methods:

• Analyzed data from 38,193 US adults aged ≥50 years from the HRS (longitudinal, biennial surveys since 1992).
• Utilized linked mortality data from the National Death Index and postmortem interviews.
• Modeled changes in 126 financial, physical, and mental health risk factors over 29 years using a transformer architecture.

Main Results:

• Over a median 9-year follow-up, 17,448 deaths occurred.
• The transformer model consistently outperformed traditional and machine learning methods.
• Achieved a twofold improvement in average precision scores (APS) for next-wave mortality prediction.

Conclusions:

• Transformer-based models, like BEHRT, significantly enhance mortality prediction.
• These models offer a superior alternative to traditional approaches.
• Highlight the potential of transformer neural networks in population health research on aging.