The COVID-19 pandemic highlighted critical gaps in our ability to effectively control infectious disease outbreaks, primarily due to widespread shortages of diagnostic testing resources. Mass surveillance testing is crucial for monitoring and mitigating the spread of such diseases, but limitations in cost, availability, and practicality hinder widespread implementation. The high demand for COVID-19 tests, particularly during surges caused by new variants (Delta and Omicron), exacerbated these issues, underscoring the urgent need for innovative solutions. Inefficiencies in testing further complicate emerging disease threats like monkeypox and exacerbate existing health inequities, particularly in rural and underserved communities. This study addresses this challenge by proposing a data-driven approach to intelligently allocate limited diagnostic testing resources, focusing on maximizing the testing positivity rate by identifying individuals most likely to be infected.

Existing literature highlights the importance of mass surveillance testing for controlling infectious disease outbreaks. However, studies also document the global prevalence of diagnostic test shortages and the challenges of implementing widespread testing due to cost, availability, and logistical difficulties. Several research efforts have explored the use of wearable sensor data to predict or detect infectious diseases, showing promise in early detection. Prior work has demonstrated the potential of using physiological data (such as resting heart rate and activity levels) as indicators of illness. This study builds upon these findings to develop a novel method that leverages these data to optimize testing allocation.

The researchers developed the Intelligent Testing Allocation (ITA) model using data from two studies: CovidInMyLife and MyPHd. Data included smartwatch data (resting heart rate and step count), self-reported symptoms, and COVID-19 diagnostic test results. The dataset comprised 15,345 participants, with 1,265 having wearable data and 126 testing positive for COVID-19. The ITA model utilized machine learning algorithms (logistic regression, k-nearest neighbors, support vector machine, random forest, and gradient boosting) to predict COVID-19 infection based on wearable data features extracted from baseline and detection periods. The team performed feature engineering to create various metrics of deviation from baseline values and employed a rigorous process including nested cross-validation to tune hyperparameters and select the best performing model for each cohort. The study also explored the impact of different data resolutions and device types on model performance, analyzing cohorts with high-frequency and all-frequency data. Three cohorts were defined based on data frequency and device type: All-Frequency (AF), All-High-Frequency (AHF), and Fitbit-High-Frequency (FHF). Model performance was evaluated using AUC-ROC and AUC-PR, with AUC-PR prioritized due to the class imbalance in the dataset. The study also analyzed the model's ability to identify asymptomatic individuals and compared the performance of models using RHR, steps, and both features combined.

The ITA model demonstrated promising results in identifying individuals with COVID-19 infection. Resting heart rate (RHR) features showed the ability to distinguish between positive and negative cases as early as 10 days before the diagnostic test, while steps features showed predictive value 5 days prior. Combining both RHR and steps features further enhanced model performance, improving the AUC-ROC by 7–11% and AUC-PR by 36–50% compared to using either feature alone. The best-performing model (logistic regression on the FHF cohort) achieved an AUC-ROC of 0.77 and AUC-PR of 0.24 on the independent test set. Importantly, the model was able to identify asymptomatic individuals (up to 27% of the ITA-identified subpopulation), indicating its broader applicability. Furthermore, the ITA method showed a significant improvement in testing positivity rate, increasing it up to 6.5-fold compared to random testing allocation (RTA), highlighting the efficiency gains in resource allocation.

This study successfully demonstrates the feasibility of leveraging data from commercial wearable devices to improve the allocation of diagnostic tests. The results validate the ITA method's potential to increase the positivity rate of COVID-19 testing, thereby optimizing the use of limited resources. The model's ability to identify both symptomatic and asymptomatic individuals is particularly significant, underscoring its value in enhancing surveillance efforts during disease outbreaks. The findings contribute to the growing body of research highlighting the utility of wearable sensor data for disease surveillance and management. The improved efficiency in testing allocation has broad implications for public health, especially in settings with limited testing resources. Future work should focus on real-world deployment and validation of the ITA model in diverse settings to assess its generalizability and scalability.

The ITA model presented in this paper provides a novel, data-driven method to improve the efficiency and effectiveness of diagnostic testing resource allocation. The model leverages readily available data from commercial wearable devices to identify individuals at higher risk of infection, increasing the positivity rate of testing and optimizing resource utilization. While further real-world validation is necessary, the study's findings suggest significant potential for improving public health outcomes and pandemic preparedness.

The study acknowledges several limitations. The bring-your-own-device study design may introduce bias due to unequal access to wearable technology. Data missingness, particularly related to device usage patterns, could affect the model's performance. Self-reported diagnostic test results may introduce inaccuracies, though the study's scale mitigated this issue to some extent. The class imbalance in the dataset posed a challenge, although the researchers addressed this by prioritizing the AUC-PR metric. Real-world deployment and validation of the ITA model remain necessary to confirm its effectiveness and generalizability in various settings. The generalizability to other infectious diseases was not evaluated explicitly.