Heel-Strike and Toe-Off Detection Algorithm Based on Deep Neural Networks Using Shank-Worn Inertial Sensors for Clinical Purpose

In clinical practice, detecting heel-strike (HS) and toe-off (TO) events using inertial sensors is often challenging due to patients' foot deformities, which can make standard sensor placements impractical. This paper introduces a novel algorithm for HS and TO detection when sensors are positioned on the lateral malleolus—a more accessible location for patients with foot abnormalities. This placement ensures secure sensor fixation during walking.

The algorithm, based on deep neural networks, is adaptable for analyzing various pathological gait patterns. This is crucial in clinical settings where numerous gait abnormalities exist. The system, integrated into a wearable device, has been validated against a reference treadmill system using a capacitance-based pressure platform.

A total of 117 healthy volunteers (62 males and 55 females, aged 24–55, height 162–183 cm) participated in the study. They walked for 2 minutes at different speeds. The algorithm's performance was assessed for gait cycle, cadence, stance phase, single support, double support, load response, and preswing.

The results showed a mean accuracy of -0.021 ± 0.091 seconds for gait cycle, 0.589 ± 1.144 steps/minute for cadence, -0.051 ± 0.544% for stance phase, -0.37 ± 0.649% for single support, 0.296 ± 0.711% for double support, 0.132 ± 0.561% for load response, and 0.106 ± 0.661% for preswing. The paper also discusses the algorithm's limitations and compares it with existing methods.