Laura McPherson, DPT, Ph.D. Joshua R. Oltmanns, Ph.D. 

ABSTRACT

Using the SPAN dataset, this project applies transformer-based survival models and large language models to predict mortality. It integrates structured data (e.g., biomarkers, personality) and transcribed life narratives to identify predictive features. The approach emphasizes scalable, non-invasive predictors and aims to improve risk stratification. Future directions include biomarker expansion and R01 submission.

Lay Summary:
Accurately predicting individual differences in mortality risk remains a critical public health challenge. Accurately predicting individual differences in mortality risk remains a critical public health challenge. As the global population ages, identifying reliable predictors of mortality can help target interventions, inform personalized care strategies, and enhance quality of life in later years. For over 17 years, the St. Louis Personality and Aging Network (SPAN) has collected a wide variety of person-level (e.g., interviews, personality measures, blood-based biomarkers) and environmental-level (e.g., geocoding) variables.

This project uses this richly phenotyped longitudinal dataset to develop predictive models that integrate biological, psychosocial, and environmental factors to assess mortality risk. It applies state-of-the-art transformer models—advanced AI tools—to analyze both traditional variables and transcribed life narratives. By combining structured data with language-based insights, the study aims to uncover new risk signals and improve the accuracy of mortality prediction. Ultimately, this research could lead to more effective prevention and care strategies for older adults, especially those underserved by traditional healthcare systems.