Rethinking the Defaults: Exploring Sample Entropy Parameters for Human Movement Data

Sample Entropy (SampEn) is widely used to quantify the predictability of movement behavior. However, the conventional values for its parameters-embedding dimension and tolerance-were originally developed for discrete and stationary physiological signals like heart inter-beat intervals and may not be optimal for smooth, nonstationary, and cyclic data such as gait kinematics. This study systematically evaluated a broad parameter grid to identify values that maximize SampEn's sensitivity to age-related differences in gait kinematics. We analyzed time series of the right thigh segment angle collected from 2199 overground walking trials across young, middle, and older adults. SampEn was computed on both raw and time-normalized time series. The parameter sweep was performed in two steps: a coarse-scale analysis (m ≈ 1-100% gait cycle; r = 0.05-0.5 standard deviation), followed by a fine-scale analysis (m ≈ 2-20% gait cycle; r = 0.05-0.15 standard deviation). Effect sizes from linear mixed-effects models revealed that conventional parameter values yielded only small-to-moderate effects, whereas time-normalized time series with m = 10% of the gait cycle and r = 0.10 standard deviation led to large group effects. Additional analyses confirmed that the large group effects observed with those values were not sample-specific, suggesting robustness of our results. These findings demonstrate that careful parameter selection, grounded on biomechanically meaningful timescales, improves SampEn's ability to capture inter-group variability in gait. Although SampEn remains a powerful tool for studying human movement variability, researchers should rethink the inherited defaults and tailor parameter choices to the temporal structure of the data in hand.