Time series classification using a modified LSTM approach from accelerometer-based data: A comparative study for gait...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Gait event detection (GED) is crucial for analyzing gait abnormalities and developing assistive devices.
• Current GED models require improvement for accuracy across varied scenarios and environments.

Purpose of the Study:

• To introduce a novel, robust method for detecting heel strikes (HS) and toe offs (TO) in the gait cycle.
• To enhance gait analysis using modified Long Short-Term Memory (LSTM) networks.

Main Methods:

• Utilized a modified LSTM network incorporating oversampling, composite accelerations, and optimized architecture.
• Tested the model on the MAREA database (20 healthy subjects) using accelerometer data from waist, wrist, and ankles.
• Evaluated performance in indoor and outdoor environments during walking and running.

Main Results:

• The modified LSTM model outperformed six state-of-the-art GED algorithms.
• Achieved high median F1 scores: 0.98 for HS and 0.98 for TO in indoor steady walking.
• Demonstrated robust performance with median F1 scores of 0.94 for HS and 0.68 for TO in outdoor walking/running scenarios.

Conclusions:

• The proposed single modified LSTM model offers a potential alternative to multiple existing GED models.
• Highlights the model's adaptability and effectiveness in diverse gait detection conditions.