Abstract
<title>Abstract</title> <p> <italic> <bold>Objective—</bold> </italic> Wearable inertial sensors monitor locomotion, but most deviation metrics use a single feature or supervised models needing labeled abnormal data. We introduce Gait Perturbation Burden (GPB), an unsupervised consensus multifeature index quantifying the proportion of locomotor windows that deviate from a person's own walking, evaluated using stair negotiation as a perturbation. <italic> <bold>Methods—</bold> </italic> In a public dataset (32 healthy adults; four body-worn accelerometers), each 3-s window was described by four directional features referenced to an individualized walking baseline. A window was flagged when at least three features deviated beyond the baseline interquartile range; GPB is the percentage of flagged windows. GPB was compared with single-feature and amplitude-only rules by discrimination, false-positive rate, and a threshold-swept operating characteristic. <italic> <bold>Results—</bold> </italic> GPB was near zero during walking (0.2–0.4%) and rose sharply at the ankles during stairs (left-ankle ascent 43.6%, 95% CI 32.3–54.6), with a steep proximal-to-distal gradient (Friedman P<0.001 at hip and both ankles; wrist not significant). GPB's walking false-positive rate was significantly lower than every comparator except the harmonic ratio (paired Wilcoxon, Holm-adjusted). On the operating characteristic a tuned single feature was more sensitive at matched false-positive rate, but GPB was confined by design to a low-false-positive region. <italic> <bold>Conclusion—</bold> </italic> A consensus rule over individualized-baseline features yields a highly specific, label-free index of locomotor perturbation, trading some sensitivity for a large reduction in false positives during unperturbed walking. <italic> <bold>Significance—</bold> </italic> The framework is task-agnostic and applicable wherever deviation from habitual movement must be flagged from wearable inertial data with minimal false alarms. </p>