Is Temporal Variability a Standalone Predictor in Medical Data? An Actigraphy Study in Bipolar Disorder

In medical data, measures of central tendency (e.g., mean) and temporal variability (e.g., standard deviation, SD) are widely used as biomarkers to quantify physiological states and disease dynamics. However, both metrics are sensitive to skewed distributions, which can obscure their true predictive value and thereby reduce interpretability. This study investigates whether temporal variability in actigraphy data carries outcome-relevant information about clinical states in bipolar disorder (BD), or whether its apparent predictive capacity is confounded by mean activity levels. Actigraphy recordings from 326 individuals with BD were analyzed, and a subset of 34 participants who experienced both manic and remission periods was used for statistical modeling. Daily features were power-transformed (Box-Cox, Yeo-Johnson) and aggregated weekly by mean (μ7) and SD (σ7). Mixed-effects logistic regression models were fitted to distinguish manic weeks from remission. Power transformations reduced μ7-σ7 correlations, and variability remained a retained outcome-relevant predictive information for most features; however, for some features (27%), the effects were partially explained by mean-variance coupling.