Abstract
Classical gait speed tests rely on the existence of a flat surface of 4 to 15 meters, distance that is not always available in clinical settings. To address this space constraint and to preserve the well–reported clinical outcomes associated with the gait speed test, the present work investigates the potential of a set of machine learning models to predict the gait speed in older adults, but by means of a short–distance test. To enhance screening, the present framework adopts a risk–stratification model, categorizing gait speed into four strata rather than treating it as a continuous variable. Motion data were obtained with a smartphone from over two hundred older adults. Feature extraction consisted of time– and frequency–domain quantities derived from the acceleration and angular velocity signals based on the discrete wavelet transform and power spectral density. Feature selection was investigated by the ReliefF algorithm. Among the several model setups evaluated, the k–Nearest Neighbors classifier is the one that provided the best predictions (median [95% CI] – accuracy: 0.969 [0.963–0.974]; F1-score: 0.971 [0.964–0.975]; AUC: 0.980 [0.975–0.984]). Based on these figures, the proposed strategy is a promising approach to provide clinicians with a reliable and straightforward gait speed stratification in cases where the classical test is not feasible due to space constraints.