Background: Identifying features for gait classification is a formidable problem. The number of candidate measures is legion. This calls for proper, objective criteria when ranking their relevance. Methods: Following a shotgun approach we determined a plenitude of kinematic and physiological gait measures and ranked their relevance using conventional analysis of variance (ANOVA) supplemented by logistic and partial least squares (PLS) regressions. We illustrated this approach using data from two studies involving stroke patients, amputees, and healthy controls. Results: Only a handful of measures turned out significant in the ANOVAs. The logistic regressions, by contrast, revealed various measures that clearly discriminated between experimental groups and conditions. The PLS regression also identified several discriminating measures, but they did not always agree with those of the logistic regression. Discussion & conclusion: Extracting a measure's classification capacity cannot solely rely on its statistical validity but typically requires proper post-hoc analysis. However, choosing the latter inevitably introduces some arbitrariness, which may affect outcome in general. We hence advocate the use of generic expert systems, possibly based on machine-learning.