SONAR, a nursing activity dataset with inertial sensors

Area of Science:

• Biomedical Engineering
• Nursing Informatics
• Human Activity Recognition

Background:

• Accurate nursing documentation is crucial for quality patient care and operational efficiency.
• Current documentation methods can be time-consuming and may not fully capture the complexity of nursing tasks.
• There is a need for objective methods to monitor and document nursing activities in real-world settings.

Purpose of the Study:

• To introduce SONAR, a novel, publicly available dataset of nursing activities recorded using inertial sensors.
• To provide machine learning models for recognizing nursing activities from sensor data.
• To establish benchmarks for deep learning architectures in this domain.

Main Methods:

• Collected 61.7 hours of data from 14 caregivers using five inertial sensors (measuring acceleration, angular velocity, etc.).
• Recorded 23 distinct nursing activities performed during daily tasks in a nursing home setting.
• Developed and benchmarked three deep learning model architectures for activity recognition.

Main Results:

• The SONAR dataset provides rich, multi-stream sensor data for 23 nursing activities.
• Evaluated deep learning models using various metrics and sensor placements, demonstrating feasibility.
• Established performance benchmarks for different model architectures and sensor configurations.

Conclusions:

• The SONAR dataset enables research in sensor-based human activity recognition in authentic healthcare environments.
• This work has the potential to enhance nursing documentation accuracy and efficiency.
• Insights from sensor data can lead to improved patient care and identification of areas for workflow optimization.