Two kinematic data-based approaches for cane event detection

Abstract

Detect cane events in real-life walking situations is needed to assess indirectly human gait without using wearable systems which may be undesirable, uncomfortable, or difficult to wear, especially for patients who are undergoing rehabilitation. This article fosters two reliable kinematic data-based approaches-a machine learning classifier and an adaptive rule-based finite-state machine (FSM)-to detect four cane events that can operate at diverse gait speeds and on diverse real-life terrains in real-time. A comparative analysis was performed to identify the most suitable machine learning classifier and the most relevant subset of features. The FSM only uses the cane's angular velocity and acceleration to facilitate its integration for daily and repeated use. Repeated measures from two groups of seven healthy subjects each were acquired to validate both approaches. The first group (23.29 +/- 1.16 years) performed trials under controlled situations in a treadmill at different speeds (from 1.0 to 1.5 km/h) and slopes (from 0 to 10%). Heterogeneous gait patterns were further collected from the second group (24.14 +/- 0.83 years) during forward walking on flat, rough, and inclined surfaces and climbing staircases. The CNN-LSTM when using the first 32 features ranked by the Relief-F method was more accurate than the FSM. The CNN-LSTM detects cane events accurately with an accuracy higher than 99% under controlled and real-life situations, except for the maximum support moment (MSM) (accuracy > 85.53%). The FSM detects most of the cane events accurately (accuracy > 90.63%). Misdetection was more pronounced in MSM (43.75% < accuracy < 84.91%). The lower computational load, together with high performances, makes these two approaches suitable for gait assessment in the rehabilitation field.