A Deep Learning Approach to Predicting Age-Related Gait Speed Decline

Gait speed is a critical measure of healthy aging outcomes due to its strong association with mortality and brain age. Understanding the hallmarks and predictors of healthy aging is essential to increasing population health span. Many attempts to predict gait speed decline and identify its key biomarkers use traditional regression-based models. However, these statistical models are only capable of testing a limited set of potential predictors. Deep learning (DL) techniques have been successful in a variety of biomedical applications and are well-suited to overcome this limitation, handling numerous input variables across multiple domains. Thus, we developed a 3-layer, DL binary classifier which predicts incident slow gait across several time frames. The neural network was trained on 15 biomarkers collected from 13,796 observations of the Baltimore Longitudinal Study of Aging. The biomarkers encompassed physiological and lifestyle variables previously identified as potentially relevant to gait speed decline. Model performance was evaluated by precision-recall statistics and compared to linear and logistic regressions. Sensitivity studies were performed on multiple domains to optimize performance. Sobol Indices identified high impact input variables. This model offers an advanced tool for prediction of incident slow gait which has the potential to aid in targeted care of individuals at risk of poor aging trajectories.