Predicting Homebound Status in US Older Adults Using Data From the National Health and Aging Trends Study

TL;DR

This study uses health data to predict which older adults may become homebound, helping to plan interventions to support their independence.

Contribution

The study introduces a machine learning model to predict homebound status in older adults using a nationally representative dataset.

Findings

The model achieved 0.75 accuracy and 0.82 AUC in predicting homebound status.

Top predictors included physical performance, mobility, and income.

Identifying modifiable factors can help design interventions to prevent homeboundness.

Abstract

At least 5% of U.S. older adults are homebound —defined as primarily confined to their residence— a condition associated with adverse health outcomes and increased costs. Timely and accurate prediction of homebound status can ensure effective implementation of interventions to support aging in place and long-term care planning. While several risk factors have been identified for homeboundness (e.g., dementia, low income), there is limited work investigating prediction. We aggregated data from Round 1 and Round 5 of the National Health and Aging Trends Study (NHATS), a nationally representative panel study of U.S. Medicare beneficiaries aged ≥65 years, to predict homebound status. We included 43 sociodemographic, clinical, functional and environmental factors from the NHATS surveys as predictors. Homebound was defined as “Never/rarely leaves home”. We developed and evaluated CatBoost machine learning models for homebound prediction, using nested cross-validation for hyperparameter tuning and class weights to address outcome imbalance. Our final dataset included 8,651 not-homebound participants at baseline (age mean (SD): 76.6 (7.6) years, 56.2% female). 11.2% became homebound within the following four years). The final model achieved 0.75 accuracy, 0.75 specificity, 0.76 sensitivity, 0.28 precision, 0.82 AUC, and 0.41 F1-score. Short Physical Performance Battery (SPPB) score, driving, ability to walk 3 blocks, ADL dependency, and income quartile were the top 5 ranking features. These findings illustrate the feasibility of predicting homebound status among older adults, enabling focused interventions before homeboundness occurs. Moreover, identifying modifiable predictors enables the design of data-driven targeted interventions to enhance mobility and promote functional independence.