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

Abstract

During the COVID-19 pandemic, contact tracing apps based on the Bluetooth Low Energy (BLE) technology found in smartphones have been deployed by multiple countries despite BLE's debatable performance for determining close contacts among users. Current solutions estimate proximity based on a single feature: the mean attenuation of the BLE signal. In this context, a new generation of these apps which better exploits data from the BLE signal and other sensors available on phones can be fostered. Collected data can be used to extract multiple features that feed machine learning models which can potentially improve the accuracy of today's solutions. In this work, we consider the use of machine learning models to evaluate different feature sets that can be extracted from the received BLE signal, and assess the performance gain as more features are introduced in these models. Since indoor conditions have a strong impact in assessing the risk of being exposed to the SARS-CoV-2, we analyze the environment (indoor or outdoor) role in these models, aiming at understanding the need for apps that could increase proximity accuracy if aware of its environment. Results show that a better accuracy can be obtained in outdoor locations with respect to indoor ones, and that indoor proximity estimation can benefit more from the introduction of more features with respect to the outdoor estimation case. Accuracy can be increased about 10% when multiple features are considered if the device is aware of its environment, reaching a performance of up to 83% in indoor spaces and up to 91% in outdoor ones. These results encourage future contact tracing apps to integrate this awareness not only to better assess the associated risk of a given environment but also to improve the proximity accuracy for detecting close contacts.