(A11) COVID-RED: A randomized trial demonstrating earlier COVID-19 remote detection using wearable technology and artificial intelligence

Publication Number: 295

Background: In 2020, many governments implemented test-and-trace programs for SARS-CoV-2 infections. These systems were often fraught with incompleteness and delays. Studies have shown that epidemic waves can only be stopped by testing-and-tracing if delays in the system are minimized. The COVID-19 Remote Early Detection (COVID-RED) trial investigated whether recent advances in wearable sensor technology and artificial intelligence in combination with symptom reporting through mobile applications could improve the potential detection of SARS-CoV-2 infections prior to symptom development.

Objectives: To determine if an algorithm ingesting wearable-measured physiological signals and symptom-reporting (experimental arm) could outperform symptom-reporting standard of care (control arm) in the early detection of first-time SARS-CoV-2 infections.

Methods: This fully decentralized, randomized, single-blind, two-period, two-sequence crossover trial randomized 17,825 subjects to experimental and control arms and followed them for up to 9 months in 2021 and 2022 in the Netherlands. Subjects wore a medical device and synced it with a novel complementary mobile app in which they also reported symptoms. Using machine learning algorithms, real-time infection indications asked subjects to get tested for SARS-CoV-2 (PCR or antigen) and to report results in the mobile application and periodic surveys. Additionally, SARS-CoV-2 serology (antibody) testing was performed periodically. The overall and early detection performance of both arms was evaluated and compared using measures of diagnostic accuracy and survival analyses.

Results: The experimental arm had better adherence and was superior in the overall and early detection of first-time SARS-CoV-2 infections (experimental: up to 162 infections reported; control: up to 143) with sensitivity up to 99% compared to 47% in the control arm and up to 7 days earlier detection. Due to an increased number of (false) positive indications to get tested this came at the cost of inferior specificity (experimental: up to 4.2%; control: up to 67.2%). Positive and negative predictive values did not differ significantly.

Conclusions: Our findings highlight the potential role of wearable devices in early detection of pre-symptomatic respiratory illnesses. Although the algorithm tended to overestimate SARS-CoV-2 infections, future iterations could be fine-tuned to increase its specificity and its ability to differentiate between various respiratory illnesses. As one of the largest tests of a SARS-CoV-2 wearable-based infection detection algorithm to date, the COVID-RED study generated a unique, generalizable dataset and demonstrated the potential of personalized biofeedback during a global pandemic.