Detecting Gait Events from Accelerations Using Reservoir Computing.

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Signal Processing

Background:

• Gait event detection (GED) is crucial for analyzing human movement, but existing algorithms are often too complex for wearable devices.
• Wearable sensors require computationally efficient and robust algorithms for real-time gait analysis.

Purpose of the Study:

• To develop and evaluate a computationally simple Echo State Network (ESN) algorithm for robust gait event detection (GED) on resource-constrained hardware.
• To assess the ESN's performance against state-of-the-art algorithms using diverse gait datasets.

Main Methods:

• Implemented a reservoir computing (RC) algorithm, specifically an Echo State Network (ESN), for numerical gait event detection.
• Trained and tested the ESN using inertial measurement unit (IMU) and ground force sensor data from 28 healthy adults across 15 walking conditions.
• Utilized ridge regression for fast training adapted to large IMU datasets for real-life environment GED.

Main Results:

• The ESN achieved robust gait event detection performance comparable to state-of-the-art algorithms, even with its low computational complexity.
• Mean absolute errors (MAE) for 6 gait events ranged from 40-120 ms (95th percentile) compared to force sensors.
• ESN performance was competitive with other algorithms, showing a MAE within 10 ms for normal walking and achieving the 2nd lowest MAE for outdoor walking/running.

Conclusions:

• The Echo State Network (ESN) offers a robust and computationally efficient solution for gait event detection (GED).
• ESN's performance is suitable for real-time applications in wearable sensors, facilitating long-term patient monitoring.
• Training the ESN with diverse data ensures good detection results across various real-life walking conditions.