Healthcare is a complex and technologically rich industry, full of innovations: in therapeutics, in devices, in processes of care and now in software. The key questions healthcare leaders face is whether these innovations are safe, whether they work and whether they add value. The established approaches to answering these questions (clinical trials or long term real world studies) are costly and difficult to scale. In this article, we argue that simulations, or digital models that imitate the operations or processes within a system, can be used in healthcare to develop new innovations and scale them.
However, to realize these gains, partnership is needed between technology organizations who have access to the technical expertise required and healthcare organizations and the patients they serve. Though the field is nascent and the investments are costly, the benefits are considerable. We believe that a health system that makes use of simulation in the ways that we describe below will not only be more effective and efficient in its daily operations but also more responsive and safer – the holy grail for health systems globally.
Simulation has demonstrated two powerful benefits
It enables the testing of strategic and operational decisions in the digital world where there is a low cost of failure.
And it creates synthetic data to train machine learning models at a scale and resolution that may be impossible to obtain otherwise.
Due to foundational advances in machine learning, simulation is now possible even when the underlying mechanics of the process to be simulated are not known a priori. “Digital twins” can be created to replicate machines in silicon, for example digital twins of jet engines can be created and used to monitor performance and predict maintenance and safety needs in a simulated environment and these inferences used to improve performance and safety in-flight. Car manufacturers use digital twins of factory processes to identify discrepancies between the observed operations of the factory and those predicted (from the simulation). Photorealistic simulations can even be used to train self-driving cars to respond to traffic, people and even the influence of the weather. In addition, at sufficient scale, rare events are statistical certainties and therefore simulation can reveal edge cases typical of the long tail distributions that describe complex processes. There is now substantial evidence that a machine learning model trained on a combination of real world and simulated data can perform better than an equivalent model trained only on real data and at a much lower cost of development. This is a fundamental shift from model centric machine learning to data centric machine learning.
Creating a simulation involves two steps:
1. Capturing data from the real world and managing trade offs
Creating a simulation begins by capturing data from the situation or process to be simulated. The guiding principle is that the closer the simulation is to the real world, the better the end results. Data can now be captured with video cameras or a wide variety of sensors with dedicated privacy preservation technology where necessary. The counterpoint is that the higher fidelity the simulation is, the more expensive it will be to develop. One key design factor for simulations is that they should be rich enough to work for the specific problem that they will be used for, but as cheap to produce and use as possible. Simulations can use just numerical data, or they can include graphical or 3D geometry data – the type and amount of data needed is dependent on the intended outcome of the simulation. A digital twin of a machine in a factory may require only data from sensors with bands indicating when the machine is behaving out of specifications. A simulation for a self-driving car would need high resolution 3D models of the environment and predefined behavioral rules regarding traffic signs and probabilistic behavior of other vehicles and pedestrians. A simulation of an airborne agent spreading in an indoor space would require a highly detailed 3D model of the environment and physical information such as air temperature and humidity, and fairly complex rules for the movement of the simulated peoples inside this environment.
2. Applying simulations in the real world
This step involves using the inferences from the simulation to improve processes or anticipate faults in the real world or as previously mentioned to train agents that can then perform in the real world. Often simulation data can be decorated with both uncertainty ranges and partially available real data to produce output that correlates much better to reality, creating agents able to predict and operate in the real world with more efficacy. A key insight from recent research in simulation is that training machine learning models on a combination of real and simulated synthetic data can produce better results than if the model was trained only on an equivalent amount of real data, and at much lower cost. This is because the synthetic data can be carefully set up to maximize the learning of the model on edge cases, or in circumstances in which it’s hard to obtain real data which is sufficiently diverse.
Simulations in healthcare
So how can simulation be deployed in healthcare?
There are two kinds of processes in healthcare: those that are well understood (for example surgical procedures that have checklists such as eye surgery, hernia surgery or minimally invasive procedures) and those that are not understood or less well understood (inpatient ward based care and outpatient face to face consultations) where the work is less structured and often reactive to emergent patient needs. Additionally, healthcare is broader than just medicine. Healthcare spans all points from prevention to palliation, involves the collaboration of multi-disciplinary teams of experts with patients as pivotal members and operates from the molecular to the planetary scale. Healthcare is replete with interdependent and often recursive processes only some of which are formally described. The necessary infrastructure for simulation is less well developed than in other industries. Healthcare is also notoriously variable with substantial practice variation between clinicians and between healthcare facilities due to differences in resources, norms of professional practice, and policy. Furthermore there are necessary privacy considerations that limit the collection of data in healthcare that are not present in other industrial settings where processes can more easily be captured.
With these considerations in mind, we believe that there are three levels at which simulation can add value in healthcare:
1. At the individual level to simulate processes occurring inside the body
Computer simulations have been used in chemistry for a long time. The level of detail and resolution of these simulations has been expanding as we gain access to increasingly powerful computing resources. This is the area where we have seen the most extensive use of simulation techniques. For example microfluidic simulations for the creation of medical devices, or molecular dynamics simulations of molecules interacting with each other.
These types of simulations are very computationally intensive. This limits the number of molecules, the size of these molecules in terms of the number of atoms, as well as the time span that we can run these simulations. Recently we have seen how AI may enable a substantial leap. Neural networks such as AlphaFold or RoseTTAFold leverage data previously collected from x-ray crystallography to estimate with very high probability the tertiary structure of proteins, potentially offering a better prediction of the effects of enzymes and drugs on metabolic processes of interest for healthcare treatments.
In-silico processes for drug discovery are only going to increase in importance. The increased availability of data, more powerful hardware and better algorithms are going to enable larger simulations across all dimensions. This, in turn, may potentially enhance preclinical drug discovery by making experimentation cheaper. We are still far away from being able to design a drug entirely in-silico, but the quality of computational tools for preclinical drug discovery has increased drastically over the years.
2. At the individual level to simulate disease processes including the interaction between individuals and their environment
Chronic diseases, the greatest burden of disease in health systems globally, by definition lack a definitive cure and need sustained and co-ordinated action to manage – where the majority of this work is done by patients themselves in between interactions with the formal health system. The growth of digital health has, for the first time, instrumented the management of chronic diseases at scale. Although it is early and progress is by no means uniform, data generated from digital health platforms is already used to proactively identify care needs and even automate clinical interventions. This same data with the techniques previously mentioned can be used to simulate the future health trajectories of individuals and therefore anticipate demand for healthcare services enabling more informed organization of healthcare resources.
Furthermore there is the potential to use simulation to personalize clinical guidelines. Now that data is available concerning the intermediate health states of individuals (from digital health platforms), it will be possible to simulate and derive ideal guidelines for an individual and use these to inform the daily actions of the patient in managing their chronic conditions and also any reciprocal actions needed from the health system. In addition, simulation of clinical trials is an active field of innovation in the life sciences industry where the proposed agility and cost advantages are appealing. But there’s a long way to go for simulated trials to meet the mark for regulatory approval of a new therapeutic.
3. At the organizational level to simulate operations in healthcare organizations
Healthcare facilities such as hospitals or clinics can be thought of as factories whose purpose is to co-create health outcomes with patients. The number of different health outcomes to be manufactured and the many ways of doing so make healthcare facilities particularly complicated factories and as such simulation is probably best restricted to a subset of processes within a healthcare facility, for example, simulating the operations of a specific clinical service line, or the flow of patients and clinicians in a hospital in order to minimize wait times and maximize productivity. Simulation can even be used to model the flow of airborne pathogens in a hospital in order to minimize the impact of hospital acquired infections or future pandemics.
The Path to Simulation
Simulation is an emerging technology and has not yet been applied at scale in any of the domains mentioned. If you are are leading a healthcare organization or working in the health technology industry, and are interested in realizing the potential of simulation, three investments are necessary:
1. An investment in collecting data
The amount and nature of data necessary for simulation depends on the process to be simulated and the level of accuracy necessary to attain a useful outcome. For self-driving cars, where the cost of failure is high as is the complexity of the environment, a lot of data is needed. For training of robots working in controlled industrial environments less data is necessary to achieve compelling results.
2. An investment in enabling technology
Creating simulations is complex, and leveraging them to perform change analysis and train machine learning models is computationally intensive. Organizations cannot expect to leverage these capabilities from day one, or to simply bring in a vendor to do it for them. Making effective use of simulation will require the development of expertise inside the organization, including the identification of key use cases and what technology is required to execute on those use cases. It will typically include using cloud infrastructure to store real and simulated data, and data processing and AI tools to run and extract insights from simulations.
3. An investment in identifying a near term return on investment
Whilst the majority of the value that will be delivered from simulation in healthcare will be realized in the long term, in order to create the necessary momentum for realizing the potential of simulation there is a clear and unequivocal value story in the short term that makes the process of transformation self-sustaining. This can be either in improvement in business results (such as finding new drugs or developing new processes), or efficiency or safety enhancements, properly measured.
Advancing healthcare through simulation requires partnership
The field of simulation has advanced rapidly due to foundational advances in machine learning and cloud computing. Simulation is now widely used in many industries to enhance safety and performance and to create data for machine learning applications. Healthcare has only relatively recently been considered a complex industrial process and has also incurred a considerable amount of technical debt as it has lagged many other industries in the adoption of digital tools.
There are many processes across the vast continuum of healthcare that could benefit from simulation technologies. In order to realize these gains, a partnership is needed between technology organizations large and small, who have access to the enabling technical expertise and healthcare organizations and the patients they serve.
Whilst the management challenge here is not trivial, it is important to bear in mind the significant potential dividends of this approach. We believe that a healthcare organization that makes use of simulation in the ways we have described will be more effective and efficient in its daily operations and also more responsive to the many challenges in healthcare; ultimately making it safer for patients.
Cloud BlogRead More