How to better combat COVID-19, a deep learning counterfactual approach

An accurate policy decision engine to combat the epidemic based on deep learning counterfactual estimators

Dr. Yefeng Zheng from Tencent, Yefeng Zheng is an R&D director at Tencent, Shenzhen, China, leading the company’s initiative on Medical AI. His research interests include medical image analysis and deep learning. He has published 100+ papers on top journals and conferences and invented 70+ patents. I am an elected fellow of the American Institute for Medical and Biological Engineering (AIMBE) and have served as an Associate Editor of IEEE Journal of Biomedical and Health Informatics since 2013.

During the COVID-19 outbreak, almost all countries have adopted a series of containment and closure policies at different time points, and some seem to have curbed the virus transmission with varying degrees of success. Many previous studies have evaluated the effect of NPIs on COVID-19 containment but most of them were based on epidemiological or statistical assumptions, which were not able to control the effects of time-varying confounders and thus led to biased effect estimations. Here, we proposed a novel deep learning based counterfactual estimator with balanced representation. Empirical evidence proved that our estimators have successfully controlled the effects of time-varying confounders, leading to unbiased estimates of different NPIs. Based on the estimator with various country-specific characteristics such as population, GDP, and epidemic scenario as inputs, our estimators can be transformed into a real-time, dynamic, and accurate policy decision planning engine. 

Our solution mainly serves government and public health authority policy decision-makers. Decision-makers often rely on limited experience and intuition to fight the epidemic.

Through a data and model-driven approach, our solution can provide decision-makers with accurate and straightforward effect estimates of different intervention policies on COVID-19 or other outbreaks similar to COVID-19, with country-specific characteristics as inputs.

First, our solution provides accurate effect estimates of different intervention policies by country over time. 

Moreover, we will also deploy a demonstration intervention policy effects and planning app, through which policymakers can interact with our model with user-specific attribute values as input and examine the expected epidemic trend and economic loss under different policy implementation strategies. Should the next pandemic similar to COVID-19 arise, our result can be used to advise policymakers to better combat the spread with more informed policy decisions.

Our data analysis and the outcome of counterfactual estimation will serve as a guide to the general public should similar unfortunate situations arise again. For example, the data we’ve collected shows countries with higher GDP per capita has a mobility pattern that is harder to slow the spread. This is validated by higher daily infections in these countries as well as less effective NPIs in counterfactual estimations.

Compare the two plots in the Figure above, we can notice faster "mean-reverting" patterns of mobility change in higher GDP per capita countries, accompanied by a higher daily infections/deaths level.

The above figure shows the two NPIs, School Closing and International Travel, work better for countries with lower GDP-per-capita. Analysis of Workplace Closing, Reduce Public Contact, Internal Mobility also show similar conclusions.

Over the next one year, we plan to collect broader and more granular data, especially the economic and vaccination data of each country, to train a model with a wider range of predicted variables. With higher model accuracy and more country-specific features, we can offer tailored NPI evaluations to specific countries and sub-national regions. As we accumulate more granular data, we will also consider offering sub-national level modeling. In the meantime, we will deploy a demonstration app including more countries and regions. This will not only reach more policy-makers and let them make more informed decisions, but also increase public adherence to those NPIs.

Over the next three years, we will try to expand our impact and get our solution into real-world usage. Tencent will cooperate with Chinese and other governments' public health authorities, encouraging them to make policy decisions with consideration of the recommendations from our model outputs. We will also launch a public health campaign that aims to strengthen the public's understanding and compliance with government intervention policies through the Tencent Wechat platform with 1 billion users.

As a solution that mainly provides an evaluation on various NPIs and tries to raise public understanding of and compliance with intervention policies, the success largely depends on the range of influence we can generate. We plan to measure the success of our solution as follows:

•  The number of countries/regions and the range of factors included in our analysis. One of our objectives is to give information backed by scientific analysis to policymakers. The more countries/regions and influencing factors included in our analysis, the better we are able to give specific and interpretable advice under various scenarios.
• The number of visits to our demonstration app. Higher public awareness can transform to better executed public intervention policies, which helps to contain a pandemic.
• The global COVID-19 epidemic is still spreading and some countries have witnessed a second wave. We will collaborate with policymakers to evaluate the performance of the solution prospectively in a real-world setting

There are two challenges for us to increase the quality and accuracy of our analysis and expand our influence on policymakers and the general public.

• Technology, quality of analysis: The accuracy of analysis and quality of conclusion is vital to the credibility, the applicability of our solution. We will continue to expand our data coverage, increase model capacity to generate higher quality analysis. Being a counterfactual estimation approach, the direct output of our model cannot be validated explicitly, which raises a higher requirement for our quality control. We will update our model regularly and monitor its results accordingly.
• Public awareness: Reaching more people and increasing public awareness is also part of our objective. Should the situation permits, we will utilize Tencent’s platform to launch a public campaign. Such a campaign can potentially let our work known to over a billion people and influence their behavior to better combat the pandemic.

Tencent

"A coalition united by the common aim of better protecting the world against health emergencies, using data-driven research and analytics." The subtitle on The Trinity Challenge’s website addresses two challenges we face in developing and deploying our solution at scale.

1. Data-driven research is the core of our NPI evaluation system. By participating in The Trinity Challenge, we will have access to expertise brought by The Members. Through collaboration among the brightest minds, we can potentially get new insights and methods to improve our model.
2. To deploy our solution at scale and let a larger population benefit from the conclusions and recommendations offered by our solution, we need a platform like The Trinity Challenge that offers a venue for global collaboration under the common aim to contribute to a world better protected from global health emergencies.

Cuebiq and Google, who have more granular human mobility and contact data of the US and European countries.