PocketCare: Tracking the Flu with Mobile Phones using Partial Observations of Proximity and Symptoms

Wen Dong, Tong Guan, Bruno Lepri, Chunming Qiao

Abstract

Mobile phones provide a powerful sensing platform that researchers may adopt to understand proximity interactions among people and the diffusion of diseases. However, it remains a challenge to track the proximity-based interactions of a whole community and then model the social diffusion of diseases starting from the observations of a small fraction of the volunteer population. In this paper, we propose a novel approach that connects these sparse observations using a model of how individuals interact with each other in terms of a sequence of proximity interactions. We apply our approach to track the spreading of flu in a 3,000-person university campus by mobilizing 300 volunteers to monitor nearby mobile phones through Bluetooth scanning and daily symptom reporting. Our results show that we can predict flu infection two weeks ahead of time with an average precision of 0.24 to 0.35, which is six to nine times higher than a random guess model. At the population level, we can predict the infectious population in a two-week window with an r-squared value of 0.95.

The Challenge of Partial Observations

Traditional epidemiological tracking often suffers from a "missing data" problem. To accurately model how a disease spreads through contact, researchers ideally need to see the interactions of an entire population. However, achieving 100% participation in a "living lab" study is nearly impossible due to privacy concerns and lack of incentives.

The authors tackle the challenge of inference from partial observations: How can we model the social diffusion of a disease across a community of 3,000 people when we only have data from 300 volunteers? The goal is to "stitch together" these sparse data points to infer the latent (unobserved) interactions and health states of the non-volunteers, effectively turning a small group of participants into a "distributed microscope" for the whole community.

The PocketCare Framework

The researchers developed the PocketCare Android app to serve as the primary data collection tool. The app performs three main tasks:

  1. Proximity Sensing: Uses Bluetooth scanning every 5 minutes to detect nearby devices (even those not running the app).
  2. Contextual Tracking: Records WiFi access points and GPS to infer locations and activities (e.g., being in a classroom vs. a gym).
  3. Health Reporting: Collects daily self-reports of flu-like symptoms from the volunteers.
The PocketCare Android application interface, showing the home screen, feature list, settings, and health reporting module.

Modeling Social Diffusion

The core innovation is the use of a Stochastic Kinetic Model (SKM). Usually used in biochemistry to model molecular reactions, the authors apply it to social dynamics, treating "getting infected" or "moving to a new room" as discrete events with specific rate constants.

To handle the sparse data, they employ Particle Filtering and Smoothing. This allows the system to maintain multiple "hypotheses" (particles) about the state of the entire community and update them as new, noisy data arrives from the volunteers. The Bluetooth scans from volunteers act as "probes" that reveal the presence of non-volunteers, allowing the particle filter to reconstruct the unobserved parts of the social network.

Heatmap visualizing the proximity interactions between PocketCare users (volunteers) and the broader community of scanned mobile phones.

Key Insights and Results

  • The Power of 10%: You don't need everyone to participate. Because people in a community share routines (attending the same lectures, eating at the same food courts), the behavior of 10% of the population is highly representative of the other 90%.
  • Proximity is a Leading Indicator: The model shows that knowing who a person was near is a far better predictor of future illness than simply looking at their current symptoms.
  • Individual Prediction: The system predicts flu infection two weeks ahead of time with an average precision of 0.24 to 0.35. This is 6 to 9 times more accurate than a random-guess baseline.
  • Population Tracking: At the community level, the model tracks the total number of infectious individuals with an R-squared of 0.95. Even when the volunteer count was dropped to just 30 people (1% of the population), the model still achieved an R-squared of 0.70.

Conclusion and Future Work

PocketCare demonstrates that mobile sensing can effectively track community health even with limited participation. While battery life and privacy remain technical hurdles, the ability to model interactions as discrete "events" captures the inherent randomness of human behavior more accurately than deterministic models. Future work will focus on automatically learning the "grammar" of social events from data, allowing the system to adapt to different types of social contagions, such as the spread of opinions or behaviors.

Summary updated

Abstract

Mobile phones provide a powerful sensing platform that researchers may adopt to understand proximity interactions among people and the diffusion, through these interactions, of diseases, behaviors, and opinions. However, it remains a challenge to track the proximity-based interactions of a whole community and then model the social diffusion of diseases and behaviors starting from the observations of a small fraction of the volunteer population. In this paper, we propose a novel approach that tries to connect together these sparse observations using a model of how individuals interact with each other and how social interactions happen in terms of a sequence of proximity interactions. We apply our approach to track the spreading of flu in the spatial-proximity network of a 3000-people university campus by mobilizing 300 volunteers from this population to monitor nearby mobile phones through Bluetooth scanning and to daily report flu symptoms about and around them. Our aim is to predict the likelihood for an individual to get flu based on how often her/his daily routine intersects with those of the volunteers. Thus, we use the daily routines of the volunteers to build a model of the volunteers as well as of the non-volunteers. Our results show that we can predict flu infection two weeks ahead of time with an average precision from 0.24 to 0.35 depending on the amount of information. This precision is six to nine times higher than with a random guess model. At the population level, we can predict infectious population in a two-week window with an r-squared value of 0.95 (a random-guess model obtains an r-squared value of 0.2). These results point to an innovative approach for tracking individuals who have interacted with people showing symptoms, allowing us to warn those in danger of infection and to inform health researchers about the progression of contact-induced diseases.
BibTeX
@article{dong2019pocketcare,
  author = {Wen Dong and Tong Guan and Bruno Lepri and Chunming Qiao},
  title = {{PocketCare: Tracking the Flu with Mobile Phones using Partial Observations of Proximity and Symptoms}},
  journal = {Proc. ACM Interact. Mob. Wearable Ubiquitous Technol.},
  year = {2019},
  url = {https://openreview.net/forum?id=i9KygLWs0b},
}

Loading related content...

Figure