Enhancing COVID-19 tracking apps with human activity recognition using a deep convolutional neural network and...

Area of Science:

• Mobile Health
• Machine Learning
• Epidemiology

Background:

• COVID-19 contact tracing apps rely on Bluetooth Low Energy for distance estimation, which is susceptible to interference.
• Inaccurate distance measurements can lead to flawed risk assessments and ineffective pandemic control measures.
• Current social distancing guidelines vary based on activity and environment, necessitating context-aware tracking.

Purpose of the Study:

• To enhance the performance and accuracy of COVID-19 tracking applications.
• To develop a robust human activity classifier for integration into mobile health tools.
• To improve the reliability of proximity detection in pandemic scenarios.

Main Methods:

• Utilized Convolutional Deep Neural Networks (CNNs) for human activity classification.
• Processed raw smartphone accelerometer data into a novel "HAR-Image" format.
• Employed k-fold cross-validation on a real-world dataset for performance evaluation.

Main Results:

• The HAR-Image approach demonstrated high effectiveness for human activity recognition.
• Achieved near-perfect accuracy (close to 100%) in classifying activities using real-world data.
• Validated the HAR-Image features as reliable inputs for tracking applications.

Conclusions:

• The proposed human activity classifier significantly improves the accuracy of distance estimation in mobile health applications.
• Integrating HAR-Images can lead to more informed risk assessments and better pandemic control strategies.
• This method offers a promising advancement for the next generation of digital health tools in public health emergencies.