Related Experiment Video
Updated: Jan 10, 2026
A Machine Learning Approach to Design an Efficient Selective Screening of Mild Cognitive Impairment
Published on: January 11, 2020
Machine Learning Algorithms for Predicting Injurious Fall Risk Among Older Adults With Depression: A Prognostic
Grace Hsin-Min Wang1, Yao-An Lee1, Amie J Goodin1
1Department of Pharmaceutical Outcomes & Policy, College of Pharmacy, University of Florida, Gainesville, Florida, USA.
Background:
Falls and related injuries (FRI) pose a large burden among older adults with depression. Proactively identifying individuals at high FRI risk enables timely and tailored interventions, reducing unnecessary health care resource utilization. However, prior prediction models relied on fixed time intervals and failed to capture dynamic changes in health status over time.
Objectives:
To develop and validate machine-learning algorithms (i.e., elastic net, random forest, and gradient boosting machine) for predicting 3-month FRI risk among older adults with depression.
Methods:
This prognostic modeling study included fee-for-service Medicare beneficiaries aged 65 years or older with a depression diagnosis in 2017. Beneficiaries were followed in 3-month episodes from the first depression diagnosis until the earliest of death, hospice services or nursing facility utilization, switching to Medicare Advantage plans, or the end of the study period (i.e., December 31, 2019). A total of 261 time-varying predictors, spanning patient-, provider-, health system- and region-related factors, were updated every 3 months to predict incident FRI risk in the subsequent 3 months. We assessed prediction performance using c-statistics and stratified patients into different risk subgroups using the best-performing model.
Results:
Among 274,268 eligible beneficiaries, the mean age was 74.6 (standard deviation [SD] = 7.2) years, 32.0% were male, 85.2% were White, and 15.1% experienced at least one FRI event throughout the study period. Using the random forest model (c-statistics = 0.68), 68.9% of the actual FRI cases were captured in the top three deciles of predicted risk. Individuals in the bottom seven deciles had a minimal FRI incidence (< 1.7%). Key predictors included frailty, age, prior FRI history, and daily dose of antidepressants.
Conclusion:
Using a nationally representative cohort and time-varying predictors, our model offers a practical approach for efficiently identifying older adults at high FRI risk, which can be updated over time. This approach can inform clinical decision-making and optimize the allocation of fall prevention resources.
