TL;DR
This paper reviews digital contact tracing for COVID-19, focusing on how smartphones and wearables can be used for proximity detection. It highlights the trade-offs between centralized and decentralized approaches and presents experimental results comparing the accuracy of BLE-based proximity sensing between smartphones and smartwatches.
What Problem Does This Paper Address?
The COVID-19 pandemic spurred the development of digital contact tracing applications as a supplement to manual methods. However, widespread adoption has been hindered by concerns about privacy and the reliability of these applications in accurately measuring proximity. This paper aims to address these issues by reviewing the current landscape of digital contact tracing, specifically examining the roles of personal devices (smartphones and wearables), networking technologies, and sensing features, and by experimentally comparing the performance of smartphones and smartwatches in proximity detection.
Methods
The paper adopts a three-component framework for digital contact tracing:
- User Devices: Focuses on smartphones and wearables as personal devices for identifying users while maintaining anonymity.
- Networking Technologies: Discusses both centralized (data stored on a cloud server) and decentralized (data stored locally on devices) approaches for data flow and analysis.
- Sensing Features: Investigates how rich sensing features on user devices can detect interactions and estimate exposure risk.
A significant portion of the paper is dedicated to experimental evaluation of proximity sensing using Bluetooth Low Energy (BLE) signals. The authors collected two distinct BLE datasets:
- Smartphone Dataset: Collected from volunteers holding smartphones in various positions (hand, pocket, backpack) and at different distances.
- Smartwatch Dataset: Collected from volunteers wearing smartwatches on their wrists, simulating different hand-to-hand and crosswise orientations.
These datasets were used to train and evaluate a decision tree classification model to distinguish between "close" and "far" proximity, with a defined cutoff distance of 2 meters.
Datasets
The paper introduces two newly collected BLE datasets for proximity sensing research:
- Smartphone Dataset: Contains 123,718 data points collected from smartphones in various configurations (hand-to-hand, hand-to-pocket, hand-to-backpack, pocket-to-backpack, pocket-to-pocket, backpack-to-backpack). This dataset is publicly available.
- Smartwatch Dataset: Contains 37,644 data points collected from smartwatches, considering direct (left-to-right, right-to-left) and crosswise (left-to-left, right-to-right) views. This dataset is also publicly available.
These datasets are intended to encourage further research in BLE-based applications beyond contact tracing.
Key Insights and Ideas
- Privacy is Paramount: The paper emphasizes that privacy concerns are a major deterrent to user adoption of contact tracing apps, driving the development of privacy-preserving protocols and decentralized approaches.
- Decentralized is Preferred: Users are more willing to adopt contact tracing applications that utilize a decentralized approach, where data is stored locally and erased after a set period, offering greater assurance about privacy.
- Wearables Offer Potential but Face Adoption Hurdles: While wearables can offer good proximity sensing performance, their lower penetration rate compared to smartphones limits their large-scale impact for general public contact tracing. They are better suited for specific, targeted environments like workplaces or healthcare settings.
- BLE Signal Strength (RSS) is Noisy: Received Signal Strength (RSS) from BLE is highly susceptible to environmental dynamics and device variations, leading to potential misclassifications (false negatives) if used as a sole indicator of proximity.
- Data-Driven Approaches are Crucial: Leveraging data-driven methods and training classifiers on diverse datasets is essential for improving the accuracy of BLE-based proximity sensing.
- Device Placement Matters: The way a smartphone is held or placed (e.g., in a pocket or backpack) significantly impacts BLE signal strength and thus proximity detection accuracy, often outperforming direct hand-to-hand measurements.
Key Results
- Smartphone Proximity Sensing Accuracy: The best classification accuracy for smartphones was achieved when both users placed their phones in their backpacks (90.85%). Other configurations like hand-to-pocket also yielded high accuracy (90.75%). Hand-to-hand configurations showed lower accuracy (85.82%).
- Smartwatch Proximity Sensing Accuracy: Smartwatches demonstrated higher overall accuracy, with the direct approach achieving 94.16% and the crosswise approach reaching 90.59%.
- Wearable Advantage: The paper suggests that smartwatches can achieve better performance due to fewer variations in their placement compared to smartphones.
- Adoption Rates are Low: Table I indicates that even for applications launched early in 2020, adoption rates are generally low, with some exceptions in European countries that implemented privacy-preserving features.
Limitations and Future Work
Limitations: * Environmental Dynamics: The paper acknowledges that BLE RSS is susceptible to environmental dynamics, which can affect measurement accuracy. * Device Diversity: Variations in device hardware and software can also impact performance. * Limited Scope of Experiments: The experimental comparison is primarily focused on BLE signals and a specific cutoff distance (2m).
Future Work: * Integration of Solutions: The authors suggest integrating smartphone and wearable solutions to provide continuous contact tracing across both daily life and work routines. * More Robust Sensing: Exploring other sensing features beyond BLE, or combining multiple sensing modalities, could further improve proximity detection. * Advanced Machine Learning: Investigating more sophisticated machine learning models beyond decision trees could enhance classification accuracy. * Addressing User Adoption: Continued focus on privacy-preserving features and clear communication about data usage is crucial for increasing user willingness to adopt contact tracing applications.
Summary updated