Using Mobility Data for Disease Modeling

The use of mobile phone data has been on the rise in areas like Public Health, city planning, and dealing with natural disasters for over ten years. The COVID-19 pandemic pushed this trend even further, as authorities needed to track movements and make decisions about travel restrictions and lockdowns. During the pandemic, mobility data helped model how the virus spread, allowing experts to monitor or predict the transmission of COVID-19.

#What is Mobility Data?

Mobility data from mobile phones show how people's movements change over time. This data helps us understand how people interact, see where cases of diseases originate, and forecast how a virus might spread geographically. Researchers analyze this data, often collected for billing or through digital platforms, to gain insights into human behaviors. For instance, mobility patterns have been used to study the seasonal patterns of diseases like dengue and rubella in countries like Pakistan and Kenya.

#The Role of Mobility Data in Disease Modeling

During the COVID-19 pandemic, mobility data became crucial. Researchers used it to create models that showed how human movement affected the spread of the virus. These models also predicted how the epidemic would unfold and estimated how effective measures like lockdowns and social distancing would be.

Despite the usefulness of these datasets, privacy remains a big concern. Even when data is deidentified and aggregated, people worry about their personal information being used without their consent. Currently, there aren't standardized agreements or guidelines to ensure privacy while also benefiting public health efforts.

#Privacy and Data Sharing

As technology continues to evolve, the amount of data available grows quickly. This makes it easier for companies or individuals to re-identify data previously considered anonymous. To address these privacy issues, various frameworks have been developed. One noteworthy approach is called Differential Privacy (DP), which helps balance privacy and the utility of the data.

DP works by adding random noise to data, which makes it harder for someone to identify specific individuals from the dataset. This way, researchers can still obtain useful information while protecting individual privacy. The concept of DP is being adopted by many companies and government agencies, but how it should be applied to mobility data in disease modeling is still not fully clear.

#Applying Differential Privacy to Disease Modeling

In this article, we look into how DP can be used in models that predict the spread of infectious diseases. We examine how different levels of noise affect important features of an epidemic by running simulations. Our method builds on an existing model and analyzes how adding noise impacts key outputs related to disease spread.

We used real mobility data from New York State during the early COVID-19 pandemic to showcase how applying differential privacy can influence certain metrics associated with disease spread. Our findings suggest that using DP may alter some estimates but can still provide significant privacy protection.

#Examining Mobility Data

The mobility data we analyzed was collected from around August 15 to November 15, 2020. This data included a large number of transitions (more than 800,000) between counties in New York, averaging about 9,000 transitions each day. The number of transitions varied greatly between counties, with some having very few and others having many. After adding DP, the total number of reported transitions changed, but the general ranking of travel routes remained consistent.

#Scenarios Analyzed

We looked at several scenarios to see how adding noise through DP affected the spread of diseases. Our first scenario considered starting outbreaks in both large and small counties. Large counties, like Kings and Queens, have populations of around 2 million, while the smallest counties, like Allegany and Essex, have populations around 46,000 and 37,000, respectively.

When starting an outbreak in large counties, we saw that it began around day 50 and peaked around day 75, with about 1% of the population infected. In the smaller counties, however, the outbreak started later, around day 60, and peaked around day 150, affecting about 5% of the population.

We tested various combinations of scenarios and noise levels over 1,000 iterations. When the epidemic started in larger counties, metrics related to the size of the outbreak and the number of counties with cases were higher than in smaller counties. However, when we introduced significant noise, many of the initial estimates dropped, while rates for spread and other factors increased.

#Connectivity and Disease Dynamics

Next, we looked at how the level of connectivity between counties affected the predictions. We simulated outbreaks in pairs of counties with either low, medium, or no connectivity to a larger county (Kings County). In areas with low connectivity, the number of infections remained quite low, while medium connectivity displayed a slightly increased infection rate.

We found that increased noise generally led to a decline in important metrics about the disease spread. This pattern continued across various scenarios, including when outbreaks started in smaller counties and in counties that were poorly connected within the mobility network.

#Variations in Epidemic Characteristics

To delve deeper into the nature of spreading diseases, we tested several potential changes in the trajectory of outbreaks in Kings and Queens counties. We looked at how a faster epidemic spread due to increased transmission rates affected the metrics. We also considered scenarios with high versus low numbers of asymptomatic individuals.

When we increased the transmission rate, the peak of the epidemic came much sooner, and more individuals got infected than in previous simulations. However, as noise increased, the metrics showed either conservative behaviors or more erratic results.

#Metapopulation Models and Their Importance

Throughout the COVID-19 pandemic, researchers developed various metapopulation models to inform decisions, anticipate the disease's spread, and identify weaknesses in healthcare systems. Mobility data has been vital in these models, providing insights into different geographical and behavioral factors among populations. However, there is still concern that such data could potentially reveal specific individuals' travel behaviors, which is why privacy measures are critical.

Our research indicates that when metapopulation models utilize mobility data, applying noise can help maintain the validity of important metrics across various noise levels. Up to a certain point, the addition of noise appears to protect individual privacy while still allowing for good estimates of public health metrics.

#Future Directions

While our findings focus on a specific combination of mobility data and modeling techniques, they present a flexible framework. This can help researchers assess the trade-off between privacy and accuracy as mobility data becomes more widely used in disease modeling.

As a next step, researchers could use our findings to evaluate how different levels of privacy-preserving noise affect their specific models and data sets. This could ultimately contribute to better privacy protection while ensuring public health data remains useful for analysis.

#Conclusion

The use of mobile phone data presents both opportunities and challenges in fields like public health and epidemiology. While these data sets can reveal important insights, especially during Epidemics, it is crucial to ensure the privacy of individuals. Differential privacy offers a promising solution by providing a method to analyze mobility data while protecting personal information. As methodologies evolve, the balance between privacy and utility will remain an essential consideration for researchers and decision-makers alike.