Facebook unveiled a remarkable piece of internal infrastructure in the spring of 2020. Called Web-Enabled Simulation (WES), the platform is a detailed replica of Facebook, with artificial user accounts ranging from simple bots that browse the site to machine-learning-based agents that mimic social interactions. The sophistication of the platform is astonishing.
Going far beyond traditional realms of software testing, which is typically directed at system behavior rather than user behavior, WES is like Facebook’s personal copy of itself. It’s a programmable alternate reality, allowing Facebook to better understand users, simulate bad content, uncover weaknesses in its privacy mechanisms, and, of course, refine the underpinnings of its global advertising empire.
On the same day in April that Facebook announced this platform, more pressing developments were taking place where sophisticated technologies like WES might have made a huge difference. The U.S. had experienced its highest number of daily coronavirus deaths since the start of the pandemic. The country was struggling to ramp up mask manufacturing even as the scientific communication surrounding their use remained muddled. And social distancing measures remained controversial and unevenly adopted.
Amid the chaotic backdrop of a global pandemic, Facebook’s simulation platform showcased a significant contemporary challenge: translating the advanced technology developed in the frictionless environment of advertising and social media to critical national infrastructure such as health care. It also highlighted that the U.S. should waste no time in developing analogous infrastructure to simulate the entire health care system.
Simulations can play critical roles as aggregators of data and models, and as computational “thought partners” for planning. In the case of the health care system, a full-scale simulation would make it possible to dynamically characterize the relationship between population health and supply chains, particularly in crisis situations.
As became painfully clear early in the pandemic, the U.S. was critically lacking in supplies such as personal protective equipment (PPE) and swabs. The specific stresses on the health care system would have been different with a different type of virus, such as one that was transmitted through drinking water rather than being airborne, or one that affected the stomach or kidneys rather than the lungs.
A computational simulation of the U.S. health care system would allow analysts to go far beyond the traditional preparedness-response exercises that many advocate for. It would allow them to model the impact of any infectious agent and calibrate the country’s biodefenses, such as PPE stockpiles and medical countermeasures. Simulating the health system would also allow analysts to anticipate and plan for a wide range of threat scenarios, such as fallout from a radiological attack; a cyberattack on closely related critical infrastructure, such as the power grid; or disruption to medical supply chains due to natural disasters.
A detailed simulation of the health care system would also be valuable in ordinary conditions. It could help uncover inefficiencies and capture nuanced aspects of hospital operations that cannot not be characterized through qualitative means or ad hoc data gathering. It could inform the allocation of residency spots for different specialties or optimize the design of clinical trials. It could help clarify the potential effects of policy changes on health care disparities and characterize the intricate relationships between race, geography, and well-being.
And as new technologies are developed that change the practice of medicine, most significantly those driven by machine learning and automation, a simulation of the health system would allow planners to prepare for deep organizational changes proactively and organically rather than reactively.
Is such a simulation even possible? And at what level of detail?
Systems-level simulations are not unheard of in the software industry, particularly for companies that face complicated deployment and integration challenges. Simulations are also used in epidemiology, some of which have been essential for shaping policy during the Covid-19 pandemic.
In recent years, simulations have entered the realm of urban planning, where platforms such as UrbanSim, developed at the University of California, Berkeley, have been used to model greenhouse gas emissions, land-use patterns, and transportation systems.
Remarkably, a platform known as Archimedes was developed nearly 20 years ago at Kaiser Permanente with WES-like ambitions for the health care system. Built in collaboration with the Sandia National Laboratories, Archimedes included models of patients, providers, interventions, policies, protocols, logistics, finances, and more. Moreover, the simulation was so granular that it modeled the relationships between a patient’s heart, lungs, kidneys, and other organs. The fact that such a simulation was possible before digitized health records, machine learning, and cloud computing suggests that its ambitions can certainly be revived today.
What would a modern effort look like operationally, and how do we avoid building an expensive software boondoggle? Modern software ecosystems have developed a sophisticated culture of building cheap prototypes, called minimum viable products (MVPs) that can inform next steps without committing a huge amount of resources. Rather than fully plan out every detail beforehand, software systems often scale up from MVPs to support larger user bases and include more complex features. This practice allows for fluidly changing course when necessary. The same flexibility could apply to a simulation of the health care system.
Who are the parties empowered to pursue this agenda and what should they do next? The deeply fragmented nature of the U.S. health care system suggests that the right thing to do is to start extremely small, with a diversity of simple prototypes that could be executed by individual research labs, small startups, or open-source projects. As a first step, a joint workshop sponsored by multiple health care organizations could invite key software leaders to examine parallels with other industries and broadly characterize the contours of a health system simulation. Thought leaders from the biosecurity world should also be represented to ensure alignment with pandemic preparedness initiatives such as the Biden administration’s proposed National Center for Epidemic Forecasting and Outbreak Analysis.
Another outcome of this workshop should be to identify a single achievable computational task that would require participation from many hospitals and care delivery organizations. The joint nature of the task is the most critical dimension. For example, building a computational model to accurately predict a single variable across many organizations over the course of a week, such as the use of a key resource or service — bags of saline, antibiotic prescriptions, number of surgical procedures, or the like — would be a good minimum viable product that could inform more ambitious iterations.
The overarching reason why the U.S. should pursue an agenda like this is that it represents a vision and ambition to direct cutting-edge technology at one of the country’s most pressing problems: our collective health. As a result of the deep health disparities in our society, discussions about health care reform here are primarily oriented towards structural changes. But in addition to structural changes aimed at advancing access and equity, we also need to continually push the boundaries of what our systems are capable of. Indeed, new systems-level capabilities could enable new ways to understand and resolve disparities.
A full-scale simulation of the U.S. health care system would represent a profound accomplishment for the country. It would be a shared structure that enables coordination across the diversity of care delivery organizations. It would be a key resource for planning for — and preventing — future pandemics and biological attacks. And it would also show the world that the U.S. is a leader in directing its unparalleled science and engineering talent to advance the nation’s well-being and ensuring its safety.
Gopal Sarma is a physician-scientist at the Broad Institute of MIT and Harvard, where he leads strategy and operations for its Machine Learning for Health initiative.