Automated Classification of the Consequences of a Slip During Walking Using a Machine Learning Approach

This study applied decision-tree (DT) machine learning models to determine whether this approach is more accurate when classifying slip outcome during walking, and to refine the cutoff thresholds of each slip type. Kinematic data of the heel were collected from 50 adults (23.1 [3.6] y) during 516 walking trials. The first DT model (DT1) was trained with heel slip distance and heel slip velocity as predictor variables; the second model (DT2) added heel slip acceleration as the third predictor variable. Walking trials were first classified as a no-slip, slip-recovery, or slip-fall outcome based on visual observation, and these classifications were used as response labels to train the DT models. Results indicated that both DT models yielded different thresholds in classifying slip outcomes and were similar to thresholds suggested in previous studies. However, both DT models resulted in 4.1% to 7.6% greater overall prediction accuracy compared with previously suggested thresholds, with DT2 generally performing better than DT1. Although the improved performance was offset by a ∼7% lower sensitivity when classifying no-slip outcomes and greater model complexity, future studies examining slip responses during gait should incorporate the thresholds derived from the DT2 model to most accurately classify the type of slip outcome.