How Can AI Help Fight Coronavirus?

Xin Gao
The Startup
Published in
9 min readMar 24, 2020
Photo by Peter Zhang in Antarctica

The US has declared this rapidly spreading coronavirus outbreak as a “national emergency” and many states are grappling with a rise in confirmed cases. There was a false dichotomy between willful obliviousness and growing panic. Some people compared COVID-19 to the seasonal flu while others were panic-buying amid COVID-19 fears. White House officials summoned tech giants to call for AI solutions to tackle the crisis two weeks ago. As a product manager at an AI startup, I would like to share some of my thoughts here.

Before asking engineers to roll up their sleeves on a prototype, all the product managers should identify different systems at play, potential users within each system and their various pain points. I am going to summarize the pain points I identified in this article and suggest possible AI solutions to COVID-19. Based on my assessment, coronavirus fears mounted on the following aspects:

1. Hospitals will be stretched beyond capacity. Many already strained hospitals will be overwhelmed by coronavirus patients if the virus is in community spread. The number of patients in need of intensive care and ventilation to support breathing will outnumber the hospital beds and ventilators.

  • Wuhan, the epicenter of the coronavirus outbreak in China, has a population of 11.08 million. As of March 24, there were 50,006 confirmed cases in Wuhan. In the first few weeks of the outbreak, the surge in coronavirus patients swamped 26 designated local hospitals and 16 makeshift hospitals with 13,467 open beds were built to hospitalize over 12,000 patients during the past 36 days. Based on my calculation, Wuhan has roughly 1.7 beds/0.08 ICU beds per 1000 people while the US has 2.8 beds/0.14 ICU beds per 1000 people.
  • Hospitals will face a dearth of ventilators, medical consumables, and other respiratory care equipment such as Extracorporeal Membrane Oxygenation (ECMO). According to a study, the US has 20.5 ICU beds with mechanical ventilation capability per 100,000 people. In Italy, the epicenter of the European outbreak, doctors were told to choose whom to save due to the lack of equipment. The ventilator producer Siare Engineering was urged by the government to quadruple the production to 500 per month. Many hospitals in the US have published guidelines to prioritize care and access to ICU beds and ventilators based on the likelihood of survival.

2. Labs/test centers will be stretched beyond capacity. Both state and commercial labs are facing a shortage of testing swabs, reagents and other materials needed to conduct the test as the demand for testing soars. Even people who have mild symptoms or are carrying the virus before developing symptoms should get tested and asymptomatic infection will worsen the community spread. The involvement of commercial labs has increased the capacity but hospital providers and individuals still need to know which lab is the best place to send your specimen to considering the adequacy of the test kit, critical lab materials, and staffing in the lab.

3. No effective medications and vaccines. Currently, there is no proven antiviral treatment for COVID-19. However, Gilead’s Remdesivir has demonstrated experimental efficacy in treating a small number of COVID-19 patients in the US, China, and Italy. Remdesivir is under phase 3 trial assessment, so hospitals have to request compassionate use for severe cases and there is a limited supply available on request now. From the prevention’s perspective, the company Moderna has started testing its vaccine on 45 healthy participants by collaborating with the National Institute of Health. Another six biotech companies are still working through preclinical development. Could AI help accelerate patient access to effective medications and vaccines?

4. COVID-19 has an insidious incubation period and can transmit indirectly. One study shows that indirect transmissions occurred through touching contaminated surfaces, viral aerosolization in a confined space, or asymptomatic infected people. COVID-19 is detectable up to 8 hours on copper, 24 hours on cardboard, 48 hours on steel and 72 hours on plastic and remains viable in aerosols for 3 hours. COVID-19 takes 2 to 14 days to show symptoms after exposure. Iceland carried out large-scale testing in the general population and over 50% of those who tested positive are asymptomatic. Researchers have not drawn a conclusion on how contagious COVID-19 is before symptoms but the latest study from Germany shows that COVID-19 presents before symptoms and virus loads are very high during the first week of symptoms, indicating the virus is contagious during the incubation period. This means that if you have close contact with an asymptomatic infected person, it’s possible for you to contract the virus. Instead of getting out of home for grocery shopping by yourself, could we use AI to reduce the possibility of having close contact with infected people or touching contaminated surfaces?

5. Poor visibility into the demand and supply of medical supplies and food. People are worried about the city running low in food under lockdown due to the lack of end-to-end visibility into the supply chain. Panic-buying and hoarding sprees reflect that uncertainty around COVID-19 and lack of faith are looming large. In order to ease people’s concern over the supply of essential items, we need a consolidated dashboard to visualize and monitor the real-time inventory in local supermarkets. In addition, transparency in the supply chain should be enhanced, so that local business owners or governments can catch the shortage of essentials earlier and take actions in a more proactive manner to manage risks and boost or restore confidence.

6. Information overload and misinformation sow panic. Information overload prompts fears which make a large number of “worried well” patients flood into hospitals. Doctors are stretched thin dealing with patients with mild flu symptoms and hospital supplies of personal protective equipment are dwindling quickly. To make matters worse, fake news is spreading at an unprecedentedly rapid pace. Even though the CDC and WHO have debunked false claims (e.g. COVID-19 only infects the elderly, drinking alcohol can prevent COVID-19, the virus will die off when the temperature rises, etc.), rumors still circulate among the public.

7. Lack of robust models that predict the spread of COVID-19. Carnegie Mellon University had the best flu forecasting model with near-perfect accuracy over the last five years. However, due to the lack of historical data and unseen data patterns caused by panic around coronavirus, Professor Roni Rosenfeld hesitated to take on the COVID-19 prediction. If a model can accurately predict the progressive trends in prevalence and the peak a few weeks in advance, it’ll buy the CDC time to take action to control the spread.

8. Small businesses are affected by Coronavirus. According to the Small Business Administration 2019 report, there are 30.7 million small businesses in the US, creating 1.8 million new jobs. The COVID-19 outbreak significantly weakened customer demand and a survey on 1000 small business owners in Seattle showed that 60% are considering wage cuts or staffing cutbacks and 34.5% concerned that they may have to close. Small businesses are facing an unprecedented challenge to survive amid Coronavirus. Although many restaurants and retail stores are still open to take-out and delivery, some worry that pick-up and delivery are not economically feasible for them and they have to deal with inventory and supply chain shortfalls. What are the possible solutions to mitigate the impact of COVID-19 on small businesses?

For each above-mentioned pain point, there are multiple ways to solve them (see Table 1). In some situations, immediate solutions such as ramping up production of essentials and medical supplies are sought to address urgent needs. In others, AI-empowered solutions are needed to distill insights from data to drive well-informed decisions amid coronavirus uncertainty. For AI solutions I proposed to tackle each pain point, they can be categorized into two buckets — (1) Prediction and Optimization, and (2) Conversational AI.

Table 1. Potential Solutions to Fight Coronavirus

1. Prediction and Optimization

The following use cases fall under the first category. All predictive models are data-hungry. The absence and inadequacy of historical data is the biggest challenge that the entire AI community is facing right now. According to Professor Rosenfeld, two preconditions will affect the success of a forecast model — sufficiently strong theory (being able to incorporate all the factors) and adequate data [1]. It’s understandable why his team at CMU collected crowdsourced data (collective human judgment) to complement surveillance data to forecast COVID-19 spread. Another idea is to collaboratively train a shared model over silo data centers or mobile devices with federated learning (FL). FL is a promising approach to learn the real patterns from the person-to-person transmission and simulate the future spread.

In terms of epidemic disease forecasting, there are three main approaches — mechanical modeling, statistical modeling, and dynamic modeling.

  • Mechanical models include compartments models and agent-based models. Compartment models partition a population into different compartments (e.g. susceptible, exposure, infectious, recovery) based on the disease transmission process. Agent-based models simulate a population by using heuristics or additional data (e.g. population demographics).
  • Statistical models include regression models and time-series models
  • Dynamic models embed mechanical models into a probabilistic framework and system discrepancy modeling gives a handle on the mismatch between the disease transmission process and data-generating process [2]

An ensemble approach has been applied in epidemic forecasting with demonstrated capability to further improve forecasting model performance upon the best results of individual components [3].

Supply chain/inventory/customer/patient load management and staff planning all boil down to demand forecast and supply optimization problems. Time-series models are a common approach to address these problems. A CMU professor Christos Faloutsos has summarized the most important classical and contemporary approaches to time-series forecasting in his recent tutorial at SIGMOD 19. The tutorial provided a detailed review of classical linear and non-linear modeling, scalable tensor methods and deep learning methods for forecasting.

2. Conversational AI

Chatbots can be used to consolidate information from reliable sources such as WHO, CDC, and state health department websites and provide answers to frequently asked questions regarding COVID-19 and situation reports at one stop. They can help reduce anxiety caused by fake news and information overload to the public. Chatbots can also be used for symptom self-check to support initial screening for hospitals. Different usage scenarios require different levels of capability from chatbots. For example, to build a symptom checker, basic chatbot such as a menu-based or keyword-based chatbot is adequate. In some scenarios, a smarter chatbot is needed to tell language nuances and identify the most relevant answer from the search space. With natural language, there might be a lot of structural or non-structural variation even in the same question asked by different people. A chatbot with advanced capabilities is needed in this situation to handle more complicated cases.

I believe in AI’s capabilities to make a difference while admitting its limits to address immediate needs during this special period of time. Data collected in the US and from other countries has only revealed the tip of the iceberg due to limited test capability. I am here calling for data curation and the release of more data, not only scientific literature on COVID-19, to the AI community who is committed to deepening insights for coronavirus decision making.

Acknowledge:

I would like to thank my friend Jennifer Trinh for proofreading this article and giving valuable feedback; Najmeh Sadoughi and Xinyu Li who shared great ideas through the lens of Machine Learning.

Reference:

[1].https://delphi.cmu.edu/files/PredictingThePredictable_13-04-03.pdf

[2].Osthus, Dave, et al. “Dynamic Bayesian influenza forecasting in the United States with hierarchical discrepancy (with discussion).” Bayesian Analysis 14.1 (2019): 261–312.

[3].Brooks, Logan C., et al. “Nonmechanistic forecasts of seasonal influenza with iterative one-week-ahead distributions.” PLoS computational biology 14.6 (2018): e1006134.

--

--