How to better combat COVID-19

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 a balanced representation. Empirical evidence proved that our estimators have successfully controlled the effects of time-varying confounders, leading to unbiased estimates of different NPIs. The accurate estimates can be transformed into a real-time, dynamic, and accurate policy decision engine which can assist governments in making policy planning beforehand in the face of a second coronavirus pandemic wave to prevent large numbers of people infected and deaths.

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. 

To fight the spread of COVID-19, governments around the world have adopted various social distancing policies, such as school closing, workplace closing, and cancel public events, yet their effectiveness has not been well assessed. In this solution, a novel counterfactual model based on the time-series COVID-19 epidemic data to quantify the effects of different interventions and their potential interactive dependencies is proposed. Firstly, we employ an RNN-based deep learning model to t and simulate the development trend of COVID-19. Then, to investigate the effects of different interventions, we make counterfactual estimates of interventions by changing the model inputs. However, estimating the effects of time-dependent intervention for different countries is affected by time-dependent confounders and the counterfactual estimation of a hypothetical situation will be biased accordingly. To get an unbiased counterfactual estimation, we introduced a novel adversarial training module with an adversarial training objective, implemented through Gradient Reversal Layer (GRL). Finally, by considering differences in medical resources, cultural and political systems, and economic development levels of different countries, we further analyzed the effectiveness of different intervention policies in different countries.

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.

Our solution mainly investigates and quantifies the effectiveness of different intervention policies of different countries and regions to further explore how to reduce the transmission and spread of the COIVD-19 epidemic more effectively. Our solution aligns with the Challenge purpose of slowing or halting future disease outbreaks. For other similar pandemics, accurate estimates also can be transformed into a real-time, dynamic, and accurate policy decision engine, slowing and halting future disease outbreaks effectively.

Our solution hopes to investigate different interventions through a model based on deep learning and give specific guidance to different countries and regions. At present, our solution belongs to the stage of proof of Concept. First of all, our solution proposes a counterfactual estimation model based on deep learning. We verify the correctness of the model through theory and a large number of experiments and prove that our solution can investigate and quantify the effectiveness of different intervention policies of different countries and regions. Then, from the original data mining process and model results, our solution shows how to investigate and quantify the effectiveness of different intervention policies from multiple perspectives. Finally, our solution has been perfected through experiment, and we are going to launch a simulation website to publicly test our solution.

To the best of our knowledge, we are the first to propose an RNN-based deep learning framework that can be used in counterfactual estimation to quantify the effectiveness of various NPIs.

Moreover, we propose a "balanced representation" to reduce the bias in the counterfactual estimation. Our model has proved to successfully eliminate the effects of time-varying confounders, leading to unbiased estimates.   

 The training procedure can be seen as fitting a function h(a_t, m_t, y_t, S) that can predict the daily infection in a future date.

During counterfactual estimation, we will alter the history of NPI and compare the model’s counterfactual output to the real case. As we do not have access to the implicit error function, the counterfactual estimation will be biased by the actual history of a_t. The introduction of "balanced representation" Φ(m_t, y_t, S) through adversarial training will solve this problem. As Φ doesn’t contain information about the actual history of a_t, Equation 1 can be rewritten as:

As the reform plan is in the preliminary stage of research and development, there are currently no serve people. But we plan the number our solution be serving in one year will be 1m and the number our solution be serving in five years be more than 100 million.

We can measure our progress toward the impact goals by the number of users and visits to our website, the number of governments or countries that cooperate with us.

Jichao Sun, Tencent 

Zhihao Ye, Tencent 

Yanpei Tian, Tencent 

Yifan Yang, Tencent 

Qianyi Wang, Tencent 

Yuan Luo, Tencent 

Sixiang Peng, Tencent 

Richard Li, Tencent 

Yue Huang, Tencent 

Xinjie Shi,  Tencent 

Yadi Wang, Tencent 

Xiaojuan Zhu, Tencent 

Yefeng Zheng, Tencent 

Alexander Ng, Tencent 

The members of the team are from Tencent Healthcare and Tencent Jarvis Lab. First, Tencent is the largest Internet company in China using technology to enrich the lives of Internet users. The mission of Tencent is "Value for Users, Tech for Good", accommodating the mission of the Trinity Challenge. Then Wexin app of Tencent is a messaging and social media app with over 1 billion active users, providing an optimum platform for a public campaign.  Tencent Jarvis Lab is a laboratory focusing on artificial intelligence in the medical field. Tencent Healthcare has published over one hundred academic papers in the field of machine learning, epidemiology, and statistics on first-class AI conferences and medical journals. One of the publications is Sun et al. (2020), which is a long-term forecast of the COVID-19 trend using a dynamic model.

In addition, our solution is also supported by the Chinese top pulmonologist Zhong Nanshan and his team. Zhong Nanshan is the top academician of the Chinese Academy of Engineering, leader of the senior expert group of the National Health Commission. His team formally reached out for Tencent to jointly establish a "big data and artificial intelligence joint laboratory". 

We will discuss and communicate about the progress of this solution every week or even a few days, and our team will discuss the different opinions of different members one by one. For the different skills of different members, we will allocate tasks reasonably. For example, some team members are good at designing algorithms and some members are good at business analysis. A reasonable division of labor allows the project to be effectively carried out and improved.

1. Data-driven research is the core of our NPI evaluation system. By participating in Health Security & Pandemics 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 Health Security & Pandemics Challenge that offers a venue for global collaboration under the common aim to contribute to a world better protected from global health emergencies.

Since our data is not comprehensive enough, we need data to further improve the model. Besides, since our plan is still at a relatively early stage, how to better cooperate with more decision-makers or different governments to cooperate is a further thing that needs to be done.

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