Inertial sensor-based heel strike and energy expenditure prediction using a hybrid machine learning approach.

Area of Science:

• Biomechanics
• Sports Science
• Healthcare Technology

Background:

• Gait analysis is crucial for estimating energy expenditure (EE).
• Traditional VO2 measurement systems are cumbersome and uncomfortable.
• Accurate, portable methods for VO2 estimation and gait event detection are needed.

Purpose of the Study:

• To develop and evaluate a hybrid machine learning model for VO2 estimation and heel strike (HS) detection.
• To assess the efficacy of integrating Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Transfer Learning (TL).
• To compare the hybrid model's performance against an LSTM-only approach.

Main Methods:

• A hybrid model combining CNNs for spatial and LSTMs for temporal feature extraction was developed.
• Transfer learning (TL) was utilized to leverage pre-trained model weights.
• A single 9-axis inertial measurement unit (IMU) was used for data collection.
• Model performance was benchmarked against a clinical-grade VO2 machine and an LSTM-only model.

Main Results:

• The hybrid model reduced VO2 prediction error from 20% to 3% compared to the LSTM-only model.
• Heel strike (HS) detection accuracy reached 93.53%.
• The IMU-based system demonstrated effectiveness as a practical alternative to traditional VO2 measurement devices.

Conclusions:

• A hybrid ML approach using IMU-based systems shows significant potential for accurate VO2 estimation and HS detection.
• The developed system offers a lightweight and practical solution for physiological monitoring.
• Further validation with diverse datasets is recommended to enhance model generalizability.