Infectious diseases have imposed a heavy burden on the global population throughout human history [1,2]. The COVID-19 pandemic has brought the threat of contagious diseases into sharp focus. With 6.95 million deaths; 769 million confirmed cases globally as of August 8, 2023 [3]; and an estimated >US $16 trillion in lost economic activity [4] for the United States alone, the COVID-19 pandemic has been one of the defining global crises of the 21st century [5].

Epidemiological models date back over a century [6-8] but have become more useful through leveraging sophisticated algorithms [9,10] and increasing computing power [11,12]. Epidemiological models to predict, plan, and respond to pandemics and outbreaks can inform decision-making to mitigate the impact of infectious diseases [6-8]. Epidemiological models require well-grounded physical and behavioral factors to provide reasonable estimates of disease spread [13]. Linking a population’s spatial behavior to infectious events that aids or inhibits the spread of disease is particularly difficult.

For airborne contagious diseases such as measles [14] or COVID-19 [5], a key enabler for disease spread is colocation (being in the same location at the same time). The effective spatial volume for COVID-19 is determined by aerosol dynamics and is often approximated at 2 m (6.56 ft) [15]. Measuring colocation can be conducted by self-reporting, as is commonly used in classic contact tracing and direct observation and counts [16], or by electronic means [17-19].

A total of 2 primary modalities for determining colocation using electronic devices exist: measurements based on estimating the distance from one person to another directly (beaconing) and measurements based on estimating the location of each person of interest within a coordinate system and calculating distances (localizing). Devices can be bespoke, such as the sociometric badge [20-22], or can leverage existing technologies such as Bluetooth(BT)–enabled phones, beacons, or dongles [23-26]. Localization techniques use systems such as GPS to place every user at a specific location at a specific time [27,28] and can be piggybacked on existing smartphones or mined from some social media platforms [29-31].

To estimate the probabilities of disease transmission, the total number of interactions, dwell times, and spatial proximity must be measured to properly baseline the parameter estimates. Techniques from companies such as Ethica Data (Avicenna Research) [32] and other companies made possible by a Google-Apple partnership [33] can be used to obtain these data for target populations under transparent and ethical data acquisition practices. However, the underlying physical processes and mathematical treatment of beaconing and localization data are substantially different and have different failure modes. Previous research had not elucidated the disparities in the estimated contact patterns for the same population between techniques. It is simple to hypothesize that colocating and beaconing will yield different contact patterns, but it is less apparent how the differences will interact with disease dynamics and impact the overall simulation outcomes.

In this study, we examine the contact patterns derived from 3 previously collected data sets using both BT beaconing and GPS localization on smartphones running the Ethica Data app. We demonstrated that while the underlying contact patterns generated from colocating and beaconing are broadly similar, they contain salient differences. For each of the 4 pathogens marked by different dynamics, we compared the results of an agent-based simulation of a communicable disease outbreak for the pathogen parameterized with beaconing- and localization-derived contact patterns. The results demonstrated that the method used to estimate contact patterns can result in significant differences between estimates of the key outbreak parameters. We showed that GPS-based contact patterns estimate significantly fewer and less-severe outbreaks than BT-derived contact patterns for the same participant and device.

Transmission models for communicable diseases are based on the characterization of natural history of a condition and contact networks [13]. In addition to traditional population-based nonspatial approaches, agent-based transmission models can use individual-level contact records and behaviors to identify emergent patterns using a bottom-up approach [9].

Real-world proximity tracking has applications in contact tracing, location-based risk assessment, mobility tracking, and outbreak detection [34]. Deriving real-world proximity contact mainly falls into 2 categories: calculating the delta of measured absolute positions—with, for example, GPS-assisted and Wi-Fi network–assisted locationing [35,36]—and directly measuring the relative distance with, for example, BT [19,37] or radio frequency identification [38].

Exemplars of each of these 2 approaches—GPS and BT—have been studied for digital contact tracing and transmission simulation [15,34]. Recent comparisons between GPS- and BT-inferred proximity contact collection approaches focus on privacy preservation, adoption, and compliance rates [15]. In contrast, the accuracy of simulations with GPS- and BT-derived proximity contacts is yet to be quantified across different underlying populations and pathogens [37].

Advances in digital contact tracing have also contributed to disease parameter estimation. For example, at the beginning of the COVID-19 pandemic, researchers focused on estimating the basic reproduction number R₀ from limited and highly regionally dependent infection data. As the pandemic spread, data collection and reporting standards enabled the daily reporting of incident cases, active cases, and mortality for various geographic scales over time, allowing the estimation of the effective reproduction number R_e [39-42].

For this study, we used 3 previously collected data sets, all of which were collected from the city of Saskatoon, a city in the midwestern Canadian province of Saskatchewan. In all these data sets, additional sensor modalities (eg, accelerometer, gyroscope, and Wi-Fi traces) were also collected, but only the BT traces, GPS traces, and battery data were used in this study. Battery data were used to identify gaps in data collection. If the phone is on and Ethica is running, then battery data were recorded, providing a more reliable way to assess the continuity of data collection than is possible with GPS, where signals can be obscured by the built environment but where the phone is still actively recording. The Saskatchewan Human Ethology Datasets (SHEDs) are a collection of pilot projects and technical trials taking place during the iEpi project—the academic precursor for the Ethica Data system—and associated postprocessing and methodological outcomes [55,56]. The SHEDs were exclusively collected from populations at the University of Saskatchewan in Saskatoon. The SHED7 data set was collected between July 11 and August 8, 2016, and included 61 students. The SHED8 data set was collected between September 25 and October 25, 2016, and included 74 students. The SHED9 data set was collected between October 28 and December 9, 2016, and included 88 students. These participants were part of a social science student study pool that included both undergraduate and graduate students and was weighted toward undergraduates.

Data collection and analysis were conducted under written approval (BEH-14-203) from the University of Saskatchewan Human Behavioral Ethics Review Board. All data were collected with the informed consent of the participants and under the oversight of the University of Saskatchewan Human Behavioral Ethics Review Board.

No experimental manipulations were conducted during data collection. The studies did not undertake stratified sampling according to ethnicity, grade, or gender. The study did not proscribe participation by those connected with the department or research laboratories involved, and the study team informed colleagues in laboratories and the Department of Computer Science first. The awareness of potential study involvement can be assumed to have spread across social networks. For each study, participants joined using their own phones to install the Ethica app and consented to have sensor data collected over the study period. Although for all these 3 studies, both Android and iPhone users were welcome, because BT beaconing did not work reliably on an iPhone due to security settings, iPhone users were removed from the analysis, and all participants reported here are Android users. Each participant receives approximately a CAD $50 (US $38.6) honorarium at the end of a study, and the exact amount varies by study length.

All 3 SHED studies stored sensor data anonymously, and the sensor data associated with a participant were identified by a device-identity number. Despite measures such as encryptions being used to protect sensor data, high-velocity GPS data can allow a skilled practitioner to determine salient information about participants, such as place of work, residence, and daily habits. We circumvented this issue by committing to our own research ethics board and the participants that researchers who access the data must commit to writing requests subject to the review and approval of our ethics boards.

Our results clearly indicate that GPS- and BT-based contact tracking yield disparate results for the same cohort under measurement. The ground truth contact network, while unknown, was the same for each data set—it was the same set of participants carrying a single phone measuring both quantities. Both BT- and GPS-derived contact measurements are estimates of the underlying contact pattern, admitting false positives (eg, BT contacts through a wall) and negatives (eg, a missed GPS contact because it occurred in an area of poor satellite reception). GPS-based contact tracking identifies fewer shorter contacts, leading to a significant decrease in expected outbreak intensity and the number of outbreaks, potentially because both participants need to have a sufficiently good location fix to estimate colocation. The denser contacts reported by BT-based contact tracking led to a higher probability of an outbreak and larger outbreaks, resulting in significantly different attack rates for most data sets and diseases. While there were conditions under which no significant differences were observed across the data collection modalities (particularly for SHED7 BT8-GPS8), differences were often significant enough to encourage caution in the uptake and interpretation of these sensed contact networks. GPS8 tends to underestimate the attack rate relative to the others (BT8, BT20, and GPS20), indicating the general inability of GPS colocating to capture proximate contacts within a short distance. Sensing configurations tend to estimate similar attack rates for infectious diseases without a comparatively high R₀ in a more distant underlying population, except for SHED9-measles. Our study cannot conclusively determine if the higher outbreak frequency and size in BT-derived networks is due to false positives in BT or false negatives in GPS, but based on the precision of commodity GPS receivers and their propensity to lose signal in large buildings, we suspected that the observed disparities are predominantly driven by GPS false negatives. If this suspicion is warranted, GPS location–based proximity measurement should be used in epidemiological simulations with caution and in a fashion that anticipates and accounts for the fact that the data collection modality used may be systemically underestimating contact. This is particularly true for the short contacts outside of normal contact networks that drive mixing.

The significance results were relatively insensitive to differences in simulated disease impacting differences in GPS and BT, but the data collection modality induced fewer differences in the results for less contagious diseases, such as seasonal influenza, than for more contagious diseases, such as measles. It is possible that weakly contagious diseases might not demonstrate differences, as outbreaks would be rare and limited in both GPS and BT networks. These findings hold for both a nominal 8-m and 20-m threshold for determining if contact has occurred. The thresholds chosen are already judicious and indicate participants being close enough during a measured portion to have come into close contact during a sensor sleep period, rather than explicitly detecting close contact. Comparing the within-sensor outcomes, the contact threshold impacted the simulated attack rate for most cases, with the exception of SHED8.

We used a stylized, agent-based SEIR model to determine the attack rate using both BT- and GPS-inferred temporal contact patterns. The stylized nature of the simulation implies that the results should be generally correct, but that more detailed models may diverge in the magnitude of the differences observed. SHED7, SHED8, and SHED9 are interesting data sets due to the multiple sensor modalities, but they are also highly biased, being drawn from a university social science participant pool comprised primarily of undergraduate students in the social and physical sciences. GPS or BT data from other demographics will almost certainly have different contact patterns, leading to different outcomes. At one extreme, institutionalized individuals (eg, in incarceration facilities or care homes) have limited mobility and would be expected to have much more convergent GPS and BT contact patterns. Perhaps not surprisingly, some of the worst COVID-19 outbreaks occurred in these institutional settings. Similarly, we analyzed 4 relatively contagious diseases and ignored diseases where a specific type of contact initiates infection, such as sexually transmitted or blood-borne diseases, or where disease propagation is slow or exhibits prolonged latent periods, such as with tuberculosis. As the definition of contact for such excluded diseases is substantially different from those analyzed in this study, the difference between GPS and BT contact patterns may be more or less pronounced. The process we have used to evaluate the differences should generalize to any contagious disease or measured contact pattern and can be used to evaluate the impact of novel contact detection algorithms or other novel diseases such as COVID-19 variants of concern.

While this study has made several meaningful contributions to the literature, particularly in highlighting divergent attack rates for GPS and BT measurements of the same underlying contact network, it is subject to notable limitations. We used 3 data sets drawn from a social sciences participant pool at our institution. These data sets included individuals who were often unknown to each other and likely produced more diffuse data sets than would have been expected had we used snowball or respondent-driven sampling or other socially connected recruiting techniques. Running a similar analysis on other data sets could provide more broadly generalizable or representative results. However, for reasonable privacy reasons, public data sets containing both GPS and BT records are not available, requiring additional measurement effort to extend this analysis. We used an agent-based SEIR model because it provided the most direct link between the data and the simulated diseases. We chose the stylized SEIR model to emphasize the role of evolving contact networks in other disease dynamics. These results could be extended to include more sophisticated disease models and compared against compartmental transmission models grounded in aggregate representations of the underlying contact network. The COVID-19 pandemic has driven innovation in contact tracing, and new measurement techniques based on dongles, beacons, or badges are now readily available. A similar analysis including these data sources could be valuable. Finally, we constrained our analysis to 4 canonical contagious diseases from a historical perspective with relatively well-parameterized behaviors. However, novel diseases will have novel disease parameters. An exploratory simulation study that outlined how diseases might be expected to behave over these contact networks using, for example, a random-walk through parameter space might be valuable in predicting new variants, existing diseases, or new diseases emerging from animal reservoirs.

Epidemiological models of disease propagation are an important tool in controlling and containing epidemic outbreaks. These models rely on the accurate measurement of key biological and behavioral parameters to ground the simulation results. Quantifying the characteristics of dynamic contact networks is a particularly challenging aspect of grounding these simulations. The significant differences in the predicted outcomes for contact networks demonstrated here between GPS- and BT-based contact tracking highlight the difficulty of grounding these simulations. Because of the nature of our data, we know that the contact networks being sought via measurement by BT and GPS should have been identical, as they corresponded to the same device held by the same individual as they went about their lives. The fact that the resulting contact networks and predicted attack rates were different indicates that these modalities are not interchangeable and that caution should be exercised by modelers employing these measures. While BT and GPS data provide more precise measurements than traditional surveys, they are still prone to error and disparate estimates of the underlying network structure and dynamics.