Characterizing Dynamic Walking Patterns and Detecting Falls with Wearable Sensors Using Gaussian Process Methods.

Area of Science:

• Wearable technology
• Machine learning
• Biomedical engineering

Background:

• Wearable devices with sensors and machine learning are increasingly used in healthcare.
• Detecting falls during walking is crucial for preventing injuries.
• Existing methods require prompt and automatic fall detection during daily activities.

Purpose of the Study:

• To investigate Gaussian process (GP) methods for characterizing dynamic walking patterns.
• To develop an automatic fall detection system using wearable sensors in smartphones and smartwatches.
• To leverage GP's uncertainty handling for robust pattern recognition.

Main Methods:

• Proposed a novel auto-encoded Gaussian process dynamical model for low-dimensional latent feature space characterization.
• Combined GP-based state space modeling with nonlinear dimensionality reduction.
• Utilized GP-based novelty detection by recycling latent samples for fall detection.

Main Results:

• The auto-encoded GP dynamical model effectively characterized dynamic walking patterns.
• GP-based novelty detection demonstrated efficient and seamless fall detection.
• Experimental results show promising performance for both pattern characterization and fall detection.

Conclusions:

• Gaussian process methods offer a robust approach for analyzing wearable sensor data.
• The proposed integrated method provides an effective solution for fall detection in wearable devices.
• This research advances the safety and utility of wearable health monitoring systems.