Performance Evaluation of COVID-19 Proximity Detection Using Bluetooth LE Signal

Area of Science:

• Computer Science
• Electrical Engineering
• Public Health

Background:

• The risk of COVID-19 transmission increases with close, prolonged contact, termed "Too Close for Too Long" (TCTL).
• Detecting TCTL situations is crucial for maintaining social distancing and mitigating disease spread.
• Bluetooth Low-Energy (BLE) Received Signal Strength Indicator (RSSI) is explored for social distance monitoring, but faces challenges due to signal variance.

Purpose of the Study:

• To evaluate the effectiveness of Machine Learning (ML) algorithms against classical estimation theory for detecting TCTL situations.
• To compare the proximity classification accuracy and confidence levels of different detection methods.
• To assess the performance of ML algorithms using the Mitre Range Angle Structured (MRAS) Private Automated Contact Tracing (PACT) dataset.

Main Methods:

• Utilized the MRAS PACT dataset containing BLE RSSI measurements.
• Applied classical estimation theory techniques for proximity detection.
• Implemented and evaluated ML algorithms including Support Vector Machines (SVM), Random Forest, and Gradient Boosted Machines (GBM).
• Extracted thirteen features (spatial, time-domain, frequency-domain, statistical) from BLE RSSI data for ML models.

Main Results:

• ML algorithms achieved comparable results to classical estimation methods.
• ML algorithms demonstrated 3.60% to 19.98% better precision in proximity estimation.
• Machine learning approaches provide enhanced accuracy in detecting TCTL situations.

Conclusions:

• ML algorithms offer a more precise solution for detecting TCTL situations compared to classical methods.
• The study validates the potential of ML in enhancing social distancing monitoring systems.
• Further research can leverage these findings to develop more robust contact tracing and proximity detection technologies.