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

Area of Science:

• Biomechanics
• Machine Learning
• Gait Analysis

Background:

• Classifying slip outcomes during walking is crucial for understanding fall prevention.
• Existing methods for classifying slip types have limitations in accuracy and threshold definition.

Purpose of the Study:

• To apply decision-tree (DT) machine learning models for classifying walking slip outcomes.
• To refine cutoff thresholds for different slip types (no-slip, slip-recovery, slip-fall).
• To compare the accuracy of DT models with existing classification thresholds.

Main Methods:

• Collected kinematic heel data from 50 adults during 516 walking trials.
• Trained two DT models: DT1 (slip distance, velocity) and DT2 (distance, velocity, acceleration).
• Used visually classified slip outcomes (no-slip, slip-recovery, slip-fall) as training labels.

Main Results:

• Both DT models produced distinct slip outcome classification thresholds.
• DT models demonstrated 4.1%–7.6% higher overall prediction accuracy than previous thresholds.
• DT2 generally outperformed DT1, though with reduced sensitivity for no-slip outcomes.

Conclusions:

• DT machine learning models offer improved accuracy in classifying walking slip outcomes.
• The refined thresholds from the DT2 model are recommended for future gait slip response studies.
• While DT models enhance accuracy, their complexity and impact on sensitivity require consideration.