Can we approach AI in medicine the same way as we do in other industries?

“What’s natural is the microbe. All the rest-health, integrity, purity (if you like)-is a product of the human will, of a vigilance that must never falter.”  

The Plague by Albert Camus, French author

Introduction

SARS. MERS. Ebola. These are familiar names of recent pandemics (see figure at left from The Visual Capitalist) that strike fear even amongst seasoned global healthcare workers even though the combined mortality (774, 38, and 11325 deaths respectively for a total of 12,137) was less than the number of people who have already succumbed to the current coronavirus pandemic (18,605 worldwide including over 700 in the US of as March 24th). The ongoing coronavirus disease 2019 (COVID-19) is a serious respiratory disease as a result of infection from the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)(Note: the former describes the disease entity from the latter which is the name of the virus). COVID-19 as a disease is manifested by fever, fatigue, cough, and shortness of breath and the main pathology is ground-glass lesions in the subpleural areas of the lungs with progression to consolidation.

The coronavirus SARS-CoV-2, a very large RNA virus, is similar to the SARS-CoV that was responsible for the SARS pandemic. SARS-CoV-2 is covered in a lipid bilayer with protein spikes, which bind to the host cell membrane via the ACE2 surface receptor for entry and replication. TMPRSS2 is an enzyme that aids the virion to enter the host cell.

This accelerating pandemic, among the worst since the Spanish flu of 1918-1919, has turned the world into a surreal apocalypse; the media hype with its multi-colored graphics of the virus and accompanying bleak counts of cases and deaths (recently accompanied by pictures of patients on the floor in the hospital gasping for air) has been difficult to watch. As a comparison, the most recent pandemic of the novel influenza A virus H1N1 in 2009-2010 (the so-called “swine flu”) lead to about 60 million cases with about 275,000 hospitalizations and 12,469 deaths in the United States alone (151,700 to 575,400 people worldwide). The following is a multi-part discussion of how artificial intelligence will need to be part of global health in its fight against present and future pandemics. First, an epidemiology and pandemic primer frames the ensuing three-part discussion. The first segment focuses on current strategies for overcoming a pandemic and how countries have performed to date. The second section discusses the early lessons learned during this pandemic and how these lessons are relevant to data science and artificial intelligence. The third and last part delineates how artificial intelligence will be leveraged as an essential partner for human clinicians and global healthcare in the future for pandemics.  

Pandemics: An Epidemiological Primer

A broad testing protocol at an early stage (especially if some individuals are relatively asymptomatic) coupled with contact tracing and surveillance is essential for discovering the true number of new cases that is the underpinning of a successful containment strategy.

There are several important epidemiological terms and concepts in the context of a pandemic, which is an epidemic (outbreak of a disease that affects many in a population in a particular region or country within a short time) that has spread to involve an entire country or several countries or regions (there is no set number of countries or regions for this definition). 

The testing of the virus, an essential part of early management of an epidemic or pandemic, has been a continual discussion and controversy in the news. Singapore had an aggressive testing program with over 6,800 people tested per million as of March 20 (compared to only 1 per million in the United States in early March). A broad testing protocol at an early stage (especially if some individuals are relatively asymptomatic) coupled with contact tracing and surveillance is essential for discovering the true number of new cases that is the underpinning of a successful containment strategy. This is probably the singular reason for high level of success seen in both Singapore and Taiwan.

There is an incubation period (or lag time) of about 2-14 days between time of actual SARS-CoV-2 infection to time of symptoms (which then leads to a positive test). Even though the lockdown in Wuhan had an immediate impact in retrospect, the number of new cases in the news at that time did not reflect this downward trend until 12 days later (due to the incubation time)(see figure below from: Wu Z and McGoogan JM. Characteristics of and Important Lessons From the Coronovirus Disease 2019 (COVID-19) Outbreak in China. JAMA (February 24, 2020) and Blog post authored by Pueyo T. Coronavirus: Why You Must Act Now. March 10, 2020).

There is much confusion and consternation with the number of new cases: this is often a reflection of number of people who had testing that turned out to be positive (in the past 24 hours) rather than the true number of new cases (which includes sometimes a much larger number of people with infection who are not yet tested). Often the number of new cases is increasing fast but in actuality, more people are getting access to the tests and therefore there are more positives simply based on more people having access to get tested (this is the case in the US). Of course the total cases is the cumulative number of cases to date (including those who have recovered from the infection). 

The case fatality rate is the number of people dying from the disease (total deaths from disease) divided by the number of people diagnosed with the disease (total cases); it is not number of people dying from the disease divided by the number of people in the entire population, as that is the mortality rate. Hence the number of deaths from COVID-19 disease is much more reliable as an index of disease than the number of people diagnosed with the disease (as this relies heavily on access to testing). The case fatality rate for pandemics range widely between the seasonal flu of about 0.1% (with about 500,000 deaths per annum worldwide) to 2.5% for the Spanish flu of 1918 (that resulted in 50-100 million deaths worldwide), and is most lethal at about 50% for Ebola. The case fatality rate for COVID-19 has ranged from about 0.5% or less (Germany) to an astonishing 9.5% (Italy) and depends on not only the demographics of the population (as it is more lethal for the senior population) but also how capable any region’s health system is to accommodate the relatively large and sudden influx of critically-ill patients. So the case fatality rate can be high because of: 1) relatively low level of testing (smaller denominator of the case fatality rate so the final number is bigger); 2) relatively high number of deaths from the disease mainly from an overwhelmed health system (larger numerator so the final number is also bigger) or 3) both (as in the case of Italy). 

Finally, the contagiousness of an infectious agent can be measured by R0 (pronounced R-naught), which is the number of people that any infected person can transmit the infectious agent. While the typical influenza has an R0 of about 1 and measles has an R0 of about 16, R0 for COVID-19 is estimated to be about 2.0-2.5 (so more contagious that the average flu but less contagious than SARS, MERS, or Ebola, all with R0 of 4 or greater). Contagiousness, however, needs to be in the context of time of manifestation of the disease as a major challenge of COVID-19 has been the lack of obvious symptoms for many especially during early phases of the infection. In other words, if there is no obvious signs and symptoms of the disease, one can infect many without knowing. In addition, contagiousness can be mitigated with measures such as aggressive testing, hand washing and sanitizing, contact tracing, temperature checkpoints, travel restrictions, and bans of gatherings above a certain size. More stringent measures include: closing sports events and bars and restaurants, closing of schools, and home quarantines except for food and urgent services (which can close as well). Finally, the ultimate “bad” virus would have the following characteristics: high case fatality rate like Ebola, high contagiousness (R0) like measles, and long incubation time with majority of hosts with little or no early symptoms like SARS-CoV-2.

Part I. Current Strategies for Pandemics

Although one hears about containment and mitigation most often in the media, other strategies for controlling an epidemic or pandemic also include anticipation (the first strategy), suppression, and eradication. Once an individual with infection is present, containment is a multi-dimensional strategy that quarantines (or isolates) infected or exposed individuals from the rest of the population as well as traces all the infected individuals’ contacts. In addition, screening and monitoring of travelers at a higher risk for infection are more effective methodologies than simple travel bans. Community spread occurs when new cases lack clear identifiable travel history related to the disease or exposure to infected individuals; when this occurs, containment has failed and mitigation is added as a strategy.

The strategy of mitigation (also labelled as “flattening the curve” which is not entirely accurate) aims to slow the further spread of the virus in hopes of not overwhelming the capacity of the healthcare system by reducing the need for hospitalizations, especially intensive care. This mitigation is executed by case isolation, home quarantines, and social distancing of those most at risk. It is akin to an “epidemiological retreat”: a maneuver to preserve lives and to buy time; although mitigation can potentially achieve these goals (like any retreat), its effectiveness and outcome are unpredictable (see figure when mitigation is not as effective as one would like).

A more stringent (and effective) strategy is suppression, which advocates (in addition to the aforementioned mitigation measures) extreme social distancing (the entire population as a whole) and school/business/event closures for a lengthy period of a few months in hopes to lower R0 to less than 1; this strategy is potentially capable of not only stopping but even reversing the pandemic and therefore is much more reliable as a measure to flatten the curve (as compared to mitigation). In short, the earlier the intervention and the stronger the measures, the more effective the strategy and the less time required for these interventions; on the other end of the spectrum, the later the intervention and the weaker the measures, the less effective the strategy and the higher the number of deaths and the higher the case fatality rate (since critically-ill patients overwhelm the health system). To make matters worse, the latter horrifying scenario will require extremely stringent measures for a relatively lengthy period of time just to stabilize the very bad situation. Finally, eradication (“epidemiological nirvana”) is the elimination of the infection and disease, and this was achieved in 1980 with smallpox (also caused by a virus, variants of variola).

The earlier the intervention and the stronger the measures, the more effective the strategy and the less time required for these interventions; on the other end of the spectrum, the later the intervention and the weaker the measures, the less effective the strategy and the higher the number of deaths and the higher the case fatality rate.

All of these aforementioned strategies (which can be “on” or “off” for varying time periods) are aimed at reducing the death toll with varying effectiveness and outcomes, and the outcomes of these strategies are also related to the timing and availability of effective therapeutic options (such as vaccines and antiviral medications).

While China imposed a historic harsh quarantine after failure of its initial containment strategy, other countries like Singapore, Hong Kong, Japan, Taiwan, and South Korea have successfully reached an acceptable equilibrium: implementing containment and mitigation strategies (and less draconian quarantines) while maintaining economic stability.

Thur far, different countries have had varying successes with different versions of these strategies. While China imposed a historic harsh quarantine (of an astounding 60 million people) after failure of its initial containment strategy, other countries like Singapore, Hong Kong, Japan, Taiwan, and even South Korea (which had a slow start due to super spreader issue) have successfully reached an acceptable equilibrium: implementing containment and mitigation strategies (and less draconian quarantines) while maintaining economic stability. Countries with less success in maintaining this delicate balance of disease burden (and capacity mismatch) and social freedom included Italy and Spain, where heart wrenching scenarios and ethical discussions had to take place in the midst of severely stressed health systems.
The number of confirmed COVID-19 cases in the United States today is now close to 55,000 (with the true number probably 10 times higher or more due to lack of access to testing in some or most areas) with well over 700 deaths. The overall performance of the US leadership and health system during this pandemic has been dysfunctional and poor. There was the perfect storm of failures: denial of the seriousness and urgency of the pandemic; lack of large-scale testing at early stages during which this practice is critical; and lack of unified and effective containment and mitigation strategies. This situation was made even worse by three populations: the uninsured, the illegal immigrants, and those without paid sick leave as these subgroups all lack desire to be tested for the virus (positive test has dire consequences). The imbroglio has made the leaders look overwhelmed and even visibly shaken (the US Surgeon General just announced with a frightened countenance “This week is going to get bad” as his public statement). There were even surreal quasi-comical news tidbits related to the pandemic (“cast of Grey’s Anatomy decided to donate masks to hospitals”; “Tom Hanks tested positive for COVID-19 but tweets his well being”; and “lines at various stores with the most popular items emptied from shelves being-eggs and toilet paper”). Despite a prior pandemic (H1N1) that started on US soil a decade ago, the US remains utterly unprepared and is near chaos. With relative conservative estimates based on an R0 of 2, an infection rate of 25-50% (depending on how much the epidemic curve is “flattened” by relatively weak and inconsistent mitigation measures), and a mortality rate of 0.5-1.0% (assuming adequate hospital staffing, space, and supplies), there is a good possibility for US to have a range of a low of 100,000-250,000 to as many as 2 million deaths in the US (potentially more deaths than American lives lost during all the wars combined, including the Civil War). There is the possibility that the virus may have a premature exit during the summer months or become attenuated due to unfavorable mutations (both characteristics of RNA viruses) so that the death toll will be significantly less than the aforementioned bleak numbers. In short, the US is paying a heavy price (economic and human) for an inadequate and underfunded public health infrastructure to deal with pandemics as well as excessive tolerance for the individualism ethos (that resulted in soft and ineffective interventional measures early on), and now playing a near futile “catch up” compounded with fragmented leadership that has not engendered trust in the public. The hope is that the overloaded healthcare system can partly take on the upcoming burden and that the virus may dwindle on its own in the ensuing months to minimize the fatalities.

Part II. Global Health Lessons Learned

Three important global health lessons learned during this pandemic based on performance of the international community are discussed below with relevance to data science and artificial intelligence:

Lesson #1: Early mass testing for the infection and organization of database need to be accurate and complete for a successful containment strategy. Testing needs to be early and broad as many or most of the infected people are not even tested in certain countries (so it is exceedingly difficult to follow true number of new cases or to calculate the case fatality rate for this virus). Whether there will be a breakdown of the health care system or not depends on this data as it reflects a real-time status of the disease and its burden. This collected data ideally should be made available to the international community in real-time for analysis. As elaborated earlier, the earlier a country institutes these relatively heavy containment measures, the less number of people will get infected and therefore the less time that these containment measures will need to be in place. 

The more proactive one is about mass testing and the smarter one is about data and data organization, the less one needs interventional measures for a longer period of time (and therefore the lower your economic burden will be).

Taiwan and Singapore executed aggressive surveillance coupled with smart use of data and databases to give real-time feedback to measure effectiveness of public health interventional strategies. Taiwan was exemplary in its proactive case identification approach in gathering data into a large database (combining native population with the visitor database) based on travel history (even supported by a QR code) and clinical status. This aggressive pursuit of accurate data even included public health officials boarding planes traveling back from Wuhan back in December to examine patients. This pandemic data program was under the direction of Taiwan’s National Health Command Center and included a list of 124 action items, with an aggressive tracing program (to identify all the people the infected person has come in contact with). A combination of aggressive testing and tracing has also decreased the initial exponential growth of the virus in South Korea after a super spreader infected many at the early stage. Counter to our American experts’ opinions, we need to test not just everyone but potentially everyone serially (those who become symptomatic after an initial negative test) in order to contain absolutely everyone with the infection.

 

The lesson here is that we need to proactively test everyone and follow new cases during the early stage of disease with a robust database and couple this strategy with an early aggressive containment, mitigation, or suppressive strategy to minimize mortality.

Lesson #2: Continual disease predictive modeling is essential for real-time accurate information to predict resource allocation and to minimize mortality. 

There has been excessive uncertainty and guesswork in projections during this pandemic, especially in the US; the best graphic the American experts can elaborate on is the now well-demonstrated one with two curves (see figure): a curve with no interventional measures that is a taller curve vs one with some measures which is a curve with lower amplitude (flattening of the first curve). These curves, usually accompanied by a dotted red line signifying hospital capacity, simply illustrate that mitigation as an intervention can result in less mortality over a longer period of time, but there is usually no defined timeline on the x-axis nor number of people on the y-axis. Data science can be coupled, therefore, to global health crises so that there is more certainty and less chaos (the latter promulgates public hysteria). Discrepancy between projected numbers and real on-the-ground numbers can be continually reconciled as there is a myriad of moving elements. Everyone should learn to appreciate that each hour or day can matter greatly as the growth of the number of infected people is not linear but exponential; any period of time has immediate sequelae of more cases and more deaths, and this bad situation is further compounded by insufficient healthcare resources (which leads then to unnecessary deaths). These models have to accommodate many nuances such as geography and climate, population demographics, early herd immunity, number of travelers, healthcare resources, interventional measures and compliance, degree of clustering, etc; these factors will need a nonlinear approach and more modern techniques (including deep reinforcement learning) to analyze the data in real-time mode. A quick search into publications on the use of deep learning in COVID-19 pandemic yielded publications on chest CT imaging rather than decision support.

If we project a 20% infection rate and a case fatality of 1% (both conservative estimates and assuming the health system is not overwhelmed), then China should have had 2.5 million fatalities; China is not even close to this number of fatalities (3,281 deaths to date) since an aggressive mitigation/suppression strategy was implemented as soon as containment failed. Of note, while the sight of Chinese citizens being forced into quarantine may seem unacceptable to Westerners, the sight of large groups of wanton American youths partying on the beach during a supposedly mitigation phase may be equally disturbing to some public health observers. Several Asian countries like Taiwan and Singapore are particularly good at gathering healthcare data with modern technology and organizing this data into databases to effectively follow healthcare interventions with robust data analytics. These countries have learned hard lessons from the 2003 SARS pandemic (especially the epicenter of that pandemic, Hong Kong). In Western countries with ample expertise in data science and artificial intelligence, there is not only lack of sophisticated collection of data during a pandemic and insufficient tracing of infected individuals, but also no clear evidence that large medical institutions or governmental agencies routinely implement more modern artificial intelligence methodologies during a health crisis. 

The lesson here is that we need to leverage big data analytics and more sophisticated machine and deep learning for an accurate, real-time map of the pandemic to enable a more precise and individualized containment, mitigation or suppression measures with appropriate allocation of valuable resources to save most number of lives.

Lesson #3: Therapeutic interventions, especially if public health measures have failed to have impact, need to be both innovative and expedient. The traditional timelines (in months and years) are no longer acceptable when the velocity of infection towards a pandemic is extremely fast (exponential) due to global connectedness; we therefore need therapies (vaccines or anti-viral medications) in hours and days. Randomized controlled trials (RCTs) with multiple phases are too time-consuming and need to be accelerated in an exponential trajectory (to match that of our viral adversaries). Volunteers with full consent can be recruited to obviate the absolute need for safety trials that are prolonged with many lives lost to pandemics during that interim. Human cognition is still important to guide a therapeutic research program: in addition to vaccines and anti-viral agents, perhaps the observation that children seem to have much less morbidity and mortality can lead to good research questions. For example, is the lack of full maturity of the immune system or lung tissue or their recent vaccinations factors in their decreased disease burden?

 

The lesson is that we need to disrupt the traditional approach of multiple phases of drug trials and bend the trajectory of these trials from linear to exponential while encouraging international and  multidisciplinary collaborative efforts in order to expediently save lives.

Among the promising (but time-consuming) trials are the ones that involve a vaccine; the viruses (especially RNA viruses like SARS-CoV-2), however, can mutate frequently (SARS-CoV-2 has already done so several times) and the solutions such as vaccines or antibodies in recovered patients’ serum are rendered less effective after these viruses mutate enough times. In addition, existing drugs can be called into action: remdesivir, a nucleotide analog antiviral drug initially used with Ebola, is already in a clinical trial. With a genomic map of the virus already online, this promulgated an international collaboration of scientists to explore therapeutic options to treat COVID-19 with various different approaches (such as attacking viral proteins like TMPRSS2 or protecting host proteins like ACE2) using some form of artificial intelligence and three-dimensional protein folding analytics and drug discovery.

 

The lesson here is that we need to disrupt the traditional approach of multiple phases of drug trials and bend the trajectory of these trials from linear to exponential while encouraging international and  multidisciplinary collaborative efforts in leveraging artificial intelligence in order to expediently save lives.

Part III. Future AI-Enabled Strategy for Epidemics

Let’s imagine our strategy against COVID-29 in the near future and how artificial intelligence can be a tour de force in the future management of pandemics:

A novel coronavirus outbreak is detected in southern France with clinical manifestation of bleeding and seizures with an R0 of 7.5 and a case fatality of greater than 50%. The AI-enabled MRI scans of the brain revealed an unusual pattern of brain inflammation and natural language processing as well as unsupervised learning are used to collect data on these patients as index cases. One-shot learning with transfer learning are deployed for ICUs around the world as an alert for these cases. In pursuit of an effective anticipation and containment strategy of the novel virus, mandatory daily testing at home (45 seconds for results) with wirelessly automated data entry is immediately started for all of France and its surrounding countries.

A real-time epidemiological map is made publicly available with proactive approach for case identification and tracing of these individuals using devices for temperature monitoring (including infrared scans now required in all public areas and transportation hubs) and travel history with internet of things and everything (IoT and IoE). Public health measures are immediately implemented in the surrounding countries in a precise format: some areas are in containment with individuals followed via their smart phones while other areas are in surveillance mode so businesses and schools are not disrupted in most surrounding regions. Drones with supplies are dispatched to people who reside in the containment areas.

Simulations of disease models using emulators (deep emulator network search, or DENSE) and AI are deployed to speed up simulations many times over of this small outbreak. The projected and confirmed numbers of new cases and case fatalities are reconciled using AI in the form of deep reinforcement learning to minimize the number of fatalities and take into account many changing nuances such as climate and demographics. Using crowd-sourced AI (including high school and college AI student championship teams as well as startups and NIH), and providing genomic sequencing and protein folding with structure predictions, the novel coronavirus and its complex quaternary biomolecular structure is successful delineated within 2 hours by this collective intelligence and a list of top 10 anti-viral agents with highest benefit-risk ratios (using generative by design algorithms) is collected within 6 hours for use in critically-ill ICU patients. The candidate drugs are designed as well as repurposed and are immediately approved by the FDA, which had representatives as part of this process. In addition, a new vaccine is made available in 12 hours as there was already ongoing work on a universal coronavirus vaccine (following the success of the universal flu vaccine in 2025). This work is necessary as coronaviruses now mutate on an hourly basis. 

After 2 months of this small outbreak, a total of 47 patients were infected with 2 deaths and AI (including training on synthetic data generated from generative models) is widely utilized in the management of these patients from a global database in the ICU and hospitals for the COVID-29 patients. The workers in the hospitals had access to AI-enabled 3D-printed masks and gowns (without shortages of the past) and robots attended the COVID-29 patients while they were infectious. A group review of COVID-29 at the international Biomedical Research and Intelligence Center (iBRAIN) and its Global Pandemic Prevention Task Force (collaborative international center formed after COVID-19 that claimed over 25 million lives, with WHO and CDC as well as representatives from 109 countries with a rotating directorship) include a discussion of the last pandemic of the current era, COVID-19, as a case history; no mitigation or suppression measures are necessary as surveillance and immediate containment obviated the need for such historic strategies.  

 

Of note, most of the aforementioned technology is already available but we need to work diligently and relentlessly towards this idealized scenario to reduce the universality of suffering for generations to come.

In conclusion, we need a proactive case identification and tracing strategy by mass screening coupled with sophisticated real-time data science-driven modeling as well as an innovative AI-derived therapeutic program. Viruses are the near perfect complex adaptive system (CAS) as these machine-like automata self-organize, pursue a common goal (finding a live host to replicate), and do this without a central leader. Albert Camus described epidemics (and even more so with pandemics) as “a shrewd, unflagging adversary; a skilled organizer, doing his work thoroughly and well.” Future viral pandemics may very well be even more dangerous adversaries as these become even more contagious and lethal. We can surpass their capabilities with passion, inspiration, and creativity but we humans also have greed, selfishness, and hubris.

Going into battle with viruses without a sound public health strategy is like going to battle without armor, and

Going into war with viruses without artificial intelligence is akin to going to war without weapons;

In both situations, the human toll is unacceptable.

We do need, however, machines to arm us with artificial intelligence to combat these viruses. We need artificial intelligence to help guide us to execute an intervention that is effective and to devise novel therapies with a much shorter timeline. This AI-inspired strategy-outcome coupling using deep reinforcement learning as well as human swarm intelligence will minimize mortality while concomitantly preserving economy (akin to how ICU doctors titrate blood pressure and cardiac output with inotropic medications). Just as we work towards synergy between clinical medicine and artificial intelligence, there also needs to be such a union between global health and data science. COVID-19, potentially the biggest pandemic since the 1918 Spanish flu, is the current generations’ world war. Going into battle with viruses without a sound public health strategy is like going to battle without armor, and going into war with viruses without artificial intelligence is akin to going to war without weapons; in both cases, the human toll is unacceptable. As Alan Turing so presciently stated: “One must design machines to fight machines”.

The protagonist in Camus’ The Plague, Dr. Bernard Rieux, devoted himself selflessly to resist the plague and to reduce the suffering of those around him towards an ethos of solidarity. We must emulate him and do the same during a very difficult time, but we humans cannot combat emerging viruses now or in the future by ourselves. We need to implement effective public health strategies immediately and concomitantly recruit artificial intelligence as our partner in this long and complex war against viruses.

Recommended Posts