Citation

Langsetmo L, Schousboe JT, Taylor BC, Cauley JA, Fink HA, Cawthon PM, Kado DM, Ensrud KE. JBMR Plus 2023; 7(8): e10757.

Copyright

(Copyright © 2023, John Wiley and Sons)

DOI

10.1002/jbm4.10757

PMID

37614297

PMCID

PMC10443071

Abstract

Targeted fracture prevention strategies among late-life adults should balance fracture risk versus competing mortality risk. Models have previously been constructed using Fine-Gray subdistribution methods. We used a machine learning method adapted for competing risk survival time to evaluate candidate risk factors and create models for hip fractures and competing mortality among men and women aged 80 years and older using data from three prospective cohorts (Study of Osteoporotic Fractures [SOF], Osteoporotic Fracture in Men study [MrOS], Health Aging and Body Composition study [HABC]). Random forest competing risk models were used to estimate absolute 5-year risk of hip fracture and absolute 5-year risk of competing mortality (excluding post-hip fracture deaths). Models were constructed for both outcomes simultaneously; minimal depth was used to rank and select variables for smaller models. Outcome specific models were constructed; variable importance was used to rank and select variables for inclusion in smaller random forest models. Random forest models were compared to simple Fine-Gray models with six variables selected a priori. Top variables for competing risk random forests were frailty and related components in men while top variables were age, bone mineral density (BMD) (total hip, femoral neck), and frailty components in women. In both men and women, outcome specific rankings strongly favored BMD variables for hip fracture prediction while frailty and components were strongly associated with competing mortality. Model discrimination for random forest models varied from 0.65 for mortality in women to 0.81 for hip fracture in men and depended on model choice and variables included. Random models performed slightly better than simple Fine-Gray model for prediction of competing mortality, but similarly for prediction of hip fractures. Random forests can be used to estimate risk of hip fracture and competing mortality among the oldest old. Modest gains in performance for mortality without hip fracture compared to Fine-Gray models must be weighed against increased complexity. © 2023 The Authors. JBMR Plus published by Wiley Periodicals LLC on behalf of American Society for Bone and Mineral Research. This article has been contributed to by U.S. Government employees and their work is in the public domain in the USA.

Language: en

Keywords

HIP FRACTURE; MACHINE LEARNING; RANDOM FOREST