Awardees
Project
Multiscale PET data metrics for improved early detection and response to epidemics
Awardees:
Imperial College London: Kris Parag
University of Oxford: Ben Lambert
About the Project:
Epidemiological data are noisy and uncertain. Reporting biases, missingness, and sparseness, particularly at fine spatiotemporal or group-level scales, hinder our ability to track spread and make informed public health decisions that are tailored to those scales. These biases and errors are then carried forward limiting the effectiveness of wider-scale measures such as population-level interventions.
Leveraging multiple different data sources can help provide more robust estimates of epidemiological quantities, especially at granular scales, offering the opportunity to overcome the above issues. Here, we investigate the inclusion of privacy-enhancing transactional data with an aim to improve estimates of critical epidemiological quantities, in particular the time-varying reproduction number, Rt, across multiple scales. We will then use our estimates to probe the constituent drivers of disease transmission, including how mobility and contact patterns affect transmission. With this knowledge, we can then better assess the effects of interventions and optimize their targeted deployment while preserving anonymity.
Project
Joint deep learning and epidemic transmission model for public health analyses with differential privacy
Awardees:
University of Virginia: Anil Vullikanti, Zihan Guan, Dung Nguyen
Georgia Tech: B. Aditya Prakash, Leo Zhao
University of Arizona: Ravi Tandon, Payel Bhattacharjee, Fengwei Tian
About the Project:
We develop a powerful framework, DPEpiNN, for performing the public health analyses in this challenge, using a combination of diverse public and private datasets. Instead of just attempting to forecast epidemic incidence with privacy (which corresponds to well-studied problems of time series forecasting, and could be done using existing differential privacy techniques), DPEpiNN combines deep learning and epidemic transmission models, which are jointly trained in a private manner for providing epidemic forecasts. In prior (non-private) work by our team members, this type of joint forecasting, which includes an epidemic model, has been shown to improve the performance of epidemic forecasting tasks. Additionally, the epidemic model (which is learned with privacy), can be used in two of the tasks in the challenge (who-infects-who and unconstrained scenarios). Further, the deep learning method is very flexible, and allows diverse kinds of datasets to be incorporated, including private credit card transaction data.
Project
Privacy-preserving anomaly detection for early outbreak discovery
Awardee:
Technical University of Munich (TUM): Dmitrii Usynin
About the Project:
In this project, we consider the task of early outbreak detection. During a disease outbreak, there are often detectable shifts in consumer behavioral patterns. One such pattern is a sudden change in the transaction timings. Additionally, previous pandemic-related data shows that there are surges in the transactions located near locations that can be broadly classified as ‘healthcare-related’. Thus, real-time financial data coupled with prior epidemiological knowledge can be effectively used to identify anomalous patterns often associated with early-stage outbreaks. Such data is, however, inherently sensitive making it challenging to perform algorithmic processing on such datasets. A number of additional privacy-enhancing technologies (PETs) could be leveraged to mitigate these issues, one of the most commonly used ones being differential privacy (DP). We are planning to leverage the existing solutions in the realm of DP model training in order to permit the use of sensitive financial data for the task of early outbreak detection.
Project
Privacy-enhanced models for transactional data
Awardee:
Indian Institute of Technology Kanpur (IIT Kanpur): Shubham Kumar, Milan Anand Raj, Divya Gupta
About the Project:
Our objective is to come up with privacy-enhanced models or tools utilizing transactional data combined with open-source data to help epidemiologists in their research and policymakers to make well-informed decisions during a pandemic. We are focused on harnessing the spatio-temporal features of transactional data through differential privacy to get real-time insight into people’s behavioral patterns during a pandemic. Through privacy-enhanced analysis of the transactional behavior of people, we will try to estimate the spread of the pandemic, the current stage of the pandemic, potential hotspot identification, and the effect of policies on people’s behavior during a pandemic.
We aim to estimate the contact patterns among social groups by analyzing the location, and age groups associated with the merchandise involved in transactions. This data will help us infer social interactions and inform disease transmission models.
The association of expenditure on transportation can be related to mobility. Therefore, we are planning to utilize transportation expenditure to develop a statistical model i.e. Poisson regression model and one Deep Learning-based model for forecasting new cases.