Developing Recurrent Neural Networks to Predict Gait Speed with Longitudinal Clinical Information

Gait speed is recognized as a significant indicator of biological aging. Through the development and use of tools that identify accelerated aging trajectories and their associated biomarkers, clinical decisions can be better-informed to include earlier and more effective interventions. Recurrent deep learning models allow for the investigation of longitudinal, nonlinear relationships between clinical variables and outcomes. The purpose of this work is to develop a recurrent neural network (RNN) to predict aging-related incident slow gait and its determinants across various timeframes from a basic set of health measures. By comparing the longitudinal analysis of an RNN with the analysis of a non-longitudinal neural network (NN), we intend to determine the relevance of longitudinal information in predictions of aging-related decline. We are utilizing the 3,821 gait speed measurements from 1,363 unique subjects in the Baltimore Longitudinal Study of Aging (BLSA) and a clinically relevant gait speed cut-point (1.0 m/s) to investigate the prediction of both current and future (2-year and 6-year timeframes) slow gait. Currently, the NN has outperformed the RNN for each of the current and future predictions. The best performing NN classifier achieved a sensitivity and specificity of 80.6% and 73.3%, respectively, for a 2-year prediction following 3 consecutive years of data. Going forward, we are developing new RNN architectures and exploring new variables to identify the determinants of gait speed and determine the significance of longitudinal information in the BLSA.