Feature selection for proximity estimation in COVID-19 contact tracing apps based on Bluetooth Low Energy (BLE)

Area of Science:

• Computer Science
• Epidemiology
• Signal Processing

Background:

• Contact tracing apps for COVID-19 often rely on Bluetooth Low Energy (BLE), which has limitations in accurately determining user proximity.
• Current apps typically use only one signal feature (mean signal attenuation), leading to debatable performance in close contact detection.

Purpose of the Study:

• To investigate the use of machine learning models with multiple features extracted from BLE signals to enhance proximity detection accuracy.
• To assess the impact of environmental factors (indoor vs. outdoor) on the performance of these enhanced models.

Main Methods:

• Extracted multiple features from BLE signals and other smartphone sensors.
• Developed and evaluated machine learning models using these feature sets.
• Analyzed model performance in both indoor and outdoor environments.

Main Results:

• Machine learning models incorporating multiple features significantly improved proximity detection accuracy compared to single-feature methods.
• Environmental awareness (indoor/outdoor) further boosted accuracy by approximately 10%.
• Achieved up to 83% accuracy in indoor settings and 91% in outdoor settings with feature-rich, environment-aware models.

Conclusions:

• Future contact tracing applications should integrate machine learning with multi-feature analysis and environmental awareness.
• This approach can substantially improve the accuracy of close contact detection, aiding in better risk assessment and pandemic control.