Leveraging Rich Data and Machine Learning to Facilitate Policy Making on Public Health Interventions in Fighting to Contain the COVID-19 Pandemic

This is a short article, focusing on promoting more study on SEIR modeling by leveraging rich data and machine learning. We believe that this is extremely critical as many regions at the country or state/provincial levels have been struggling with their public health intervention policies on ﬁghting the COVID-19 pandemic. Some recent published papers on mitigation measures show promising SEIR modeling results, which could shred the light for other policymakers at diﬀerent community levels. We present our perspective on this research direction. Hopefully, we can stimulate more studies and help the world win this “war” against the invisible enemy “coronavirus” sooner rather than later.

The compartmental SEIR (Susceptible, Exposed, Infectious, and Recovered) model (Fig. 1) and the like in epidemiology prove effective in facilitating policymaking on intervention measures enforcement and healthcare planning to fight a pandemic at different control levels, regionally, nationally, or internationally (1).However, the notorious reproduction number (R0), an epidemiologic metric used to describe the contagiousness or transmissibility of infectious agents, in an SEIR model has been challenging to get accurately estimated due to the complexity of measuring and estimating R0 (2-3).R0 for COVID-19 is a function of all the parameters (e.g., λ, δ, µ, γ, p, and time) in Fig. 1.Without a vaccine or effective treatment, R0 dictates the curve shape of a pandemic, while public health interventions determine the change of R0 over time, i.e., when and how much R0 drops (1-3).The recent study from Dehning et al. (2) has clearly shown that appropriate public health interventions enacted and enforced promptly in Germany proves effective in containing the COVID-19 pandemic in Germany.They show that their approach using Bayesian inference with Markov-Chain Monte Carlo (MCMC) by leveraging the available rich pandemic data can help quantify R0 and other key parameters in SEIR modeling, which further helps assess the effectiveness of public health interventions.It is noteworthy that the proposed approach proves to be practically useful in terms of quantifying the time and rates of R0 dropping owing to different restrictive levels of intervention measures that have been adopted over their study period.
Flattening the new case curve with intervention while increasing healthcare capacity over time and considering social and economic impacts has become the essential strategy around the world (1).R0 for COVID-19 is highly related to the spreading rate of the coronavirus and the recovering rate of COVID-19 patients.The former rate changes along with the mutation of the virus and the timing, enforcement, and types of public health interventions as time goes; the latter rate varies with the communities where population and healthcare systems differ considerably (1-3).Thus, weaponizing and implementing public health interventions face tremendous challenges, substantially varying with communities, locally, nationally, and internationally (1-3).Unfortunately, this is particularly true in the US.As of today, the outcomes of fighting COVID-19 across the US are substantially different from state to state as witnessed since the beginning of March when the outbreak started in the US.The northeastern corridor centered at New York city had been the epidemic center since March until the end of May.Gratifyingly, we have seen how the COVID-19 across the region has been contained due to the enforcement of strict intervention measures.However, since this past Memorial Day, new coronavirus infections and hospitalizations are rising in many other states.The latest data clearly shows that COVID-19 hospitalizations have begun to swell in places like Texas, Arizona, Florida, and California.On June 18, 2020, more states, including Arizona, North Carolina, Texas, and Florida finally mandated face coverings in public to get a grip on spiraling coronavirus cases as those states continue to set daily records since this past Memorial Day.
Although the reasons for this unfortunate consequence might be associated with political and economic considerations and sometimes individual beliefs, the lack of data-driven approach that can better measure and estimate R0 in developing effective SEIR and corresponding social and economic models that can accurately show the consequences in a scientific and timely manner can be a key contributing factor.Even worse, R0 has been quite misleading due to too much unknown about the COVID-19.Consequently, policymakers tend to be influenced by their partisan stance rather than the science and reality because of lacking the needed scientific evidence and support.
We understand that inadequate intervention could result in a resurgence of COVID-19 pandemic locally and nationally (Fig. 1).One-size-fits-all lockdown might not be an optimal policy either.In the US, COVID-19 hot spots have shifted to the south where population is with more retired people who tend to have some chronic underlying conditions.As a result, more infected persons would become more seriously ill, leading to more hospitalizations and heavily taxed hospital systems.This will be surely more "worrisome" in the fall and winter when flu viruses circulate routinely, jeopardizing the healthcare systems across the country."History never repeats itself but it rhymes," said Mark Twain.Dehning et al. (2) reveal that intervention measures have been the greatest "weapon(s)" to fight COVID-19 effectively in Germany.We believe that more research on improving SEIR modeling will play a critical role in facilitating the policymaking on public health interventions.As the COVID-19 pandemic data is daily updated state by state, the currently available rich COVID-19 pandemic data and machine learning algorithms can thus be leveraged to extend and enhance SEIR modeling with parameter auto-adjustment capability weekly if not daily.R0, a time-and location-dependent epidemiologic metric, along with other parameters (Fig. 1), can be more accurately estimated at the state level and the local community level if needed.More promisingly, this kind of study can be further enhanced and used to facilitate policymaking on the timing and restrictive levels of intervention measures, types and lifting or hardening, needed for a controlled area, at the state, region, or country level as time goes.Rich data and machine learning work well in reducing discrepancies.Because of knowing all truth that is highly context-dependent, promisingly and confidently, we can win this "war" and be well prepared for the future.

Fig. 1
Fig. 1 An Extended SEIR model (λ: expected number of persons an infected person infects per day, δ: 1/days to get infected, also called incubation time, µr: 1/days to recover without hospitalization, µc:1/days to be hospitalized from infection, ph : percentage of Infected persons to be hospitalized, γr : 1/days to recover from hospitalization, pd : percentage of Infected persons to die, γd: 1/days of hospitalized patients to die.R0 is contextual, a function of time and these above-mentioned parameters.)