In the technology industry, new ideas and innovations emerge that are more than intriguing but, in short order, get diluted by commercialization. Digital twins is one of those ideas.
Most of us were introduced to the idea of a digital twin (though, at the time, it wasn’t even digital) in the movie Apollo 13. Astronaut Ken Mattingly spent hours on the ground in a simulator trying to figure out a way to get the astronauts home safely in their crippled spacecraft by trying different sequences of steps that would not deplete the scant resources in the craft.
Today, a useful and concise definition of a digital twin from Michael Ouissi on diginomica is:
A digital twin is an up-to-date and accurate copy of the physical object’s properties and states, including position, shape, status, and motion. As a digital representation, it provides the elements and the dynamics of how a connected asset operates throughout its lifecycle. Using this technology, an organization can increase predictability and lower risks.
Digital twins are emerging as critical elements of innovative city technologies – reaching the United Nations’ Sustainable Development Goals (SDGs) by minimizing waste and boosting resource efficiency, economic growth, well-being and prosperity. Digital copies of products and infrastructure systems may help protect the environment and create a sustainable future for generations.
The key word here is “dynamics.” A digital twin is only as good as a model. Data is secondary. These days, most of the digital twin hype is about sensor data volume, speed, detail, and the high-speed networks to process it. If the digital twin is incapable of faithfully and accurately replicating the object’s behavior, it not only has no value, it can be dangerous.
Consider two alternatives:
Simulation models: Necessary elements for digital models include distributed (edge) and large-scale computing capacity, artificial intelligence, and the best domain experts available. Table stakes is a model of how the thing works in its most intricate and often unusual ways with pretty high fidelity. That requires that the developers have a deep and thorough knowledge of how the thing works.
Commercial aviation’s safety record is nearly a miracle in improvement over the past few decades, because designers and manufacturers understand the variances and minutiae of operation. Digital twins emerged as working models of engineered objects. A good example is a jet engine. Every engine has unique characteristics, even in identical models. The model has to consider the history of its use, the aircraft(s) to which it is employed, all of the external conditions in which it operates, the quality and cadence of maintenance and any out-of-range events that occurred in its operation. Aircraft engine manufacturers load the engines with redundant sensors (and logic to sort out differential readings).
Simulation versus digital twin
Part of the problem is confusion between digital twins and simulation models, the latter of which has been practiced for decades. A simulation is used to project possible outcomes, not what is happening in real time. Though they may use some of the same types of data, simulations generate data as part of the simulation. Digital twins typically do not.
Designing any kind of simulation takes skill, practice and domain knowledge, and the difficulty level is high. Inside a fully-formed simulation model, we start with a hypothesis and gather as much data as possible, and the machine looks for patterns. In today’s world, this is the reverse, such as machine learning. Inside a fully-formed model, we start with a hypothesis and gather as much data as possible, and the machine looks for patterns. In today’s world, this is the reverse, such as machine learning.
Digital twin models may begin with simulation, but the defining characteristic of a digital twin model is to start with logic, no data, a faithful representation of the object, and not attempting to derive it from existing data.
We attempt to predict what humans will do next in classical propensity models. The problem is that humans are unpredictable. When models perform well in a short timeframe, they begin to degrade, and their overall predictive fidelity is quite low. For that reason, they are applied where the stakes are likewise low. That would be unacceptable in jet engines. It would be unethical in medicine. The ethical implications for therapy or preventative care are extreme.
Digital twins in healthcare
The use of digital twins when knowledge is uncertain, incomplete, contradictory, and outcomes unethical, are a perversion of the concept. The commercialization of the digital twin concept is being promoted to bizarre lengths. For example, an immature technology that proposes to dynamically reflect the “-omics” (the addition of “omics” to a molecular term implies a comprehensive, or global, assessment of a set of molecules), such as genomics, biomics, proteomics, or metabolomics, as well as physical markers, demographic, and lifestyle data over time of an individual are years away from being practical.
A misguided approach from the EU researchers:
Created a digital twin of the brain, which may help predict a person’s health, mental states and behaviors with unprecedented accuracy. The “NeuroTwin” project will simulate the behavior of a healthy brain based on data from MRI scans of people who volunteered to participate in the study. It will include detailed 3D depictions of the brain’s activity at multiple levels, from individual neurons to the level of entire brain areas, as well as information from wearable devices that measure factors such as heart rate and body temperature.
There is nothing more mysterious in the human body than the brain. This barely raises above phrenology. In an earlier article, I wrote:
One aim of digital twins in medicine is personalized medicine – by identifying deviations from normal. It is questionable how feasible this is at our current level of knowledge. However, there are projects to use digital twin techniques for PARTS of the human body, such as the fit and placement of prosthetic devices. Buy a digital twin of the heart as more than a pump. This will not work. The heart is loaded with neurotransmitters, sentient epithelial cells and even its microbiome. In other words, it is much more complex than it appears at first blush. Nevertheless, some beneficial effects may emerge, but that’s the problem with human digital twins – researchers get far over their ski tips.
There is a reasonably feasible application of the technology in healthcare: A digital twin of a hospital with operational strategies, capacities, staffing, and care models is more aligned with the engineering model. GE Healthcare proposes uses such as assisting in bed shortages (a simple model), spreading germs, staff schedules, and operating rooms (which seems promising). The effect would be improving patient care and performance (and cost, apparently). Digital twins can be a virtual test of alternatives without actual risks.
It begs the question, even if the data in the Digital twin is pristine, are the models that determine “abnormal” and generate inferences reasonable enough? The models and interpretation of the model are often wrong, catastrophically wrong. Lower salt or not for high blood pressure? Low-fat or high-protein?
Something that gives me pause about personalized medicine: a 2013 study of a decade of medical journal articles found that of the 363 articles focused on the standard of care practices, 146, or about 40%, led to reversals of the practice. A 2019 study of over 3,000 randomized controlled trials published in three prominent general medical journals concluded that 396 of these trials constituted medical reversals. The most common disease category among the reversals identified was cardiovascular disease.
The engineering paradigm inherent to digital twins-oriented health care will raise novel ethical, legal, and social issues for therapy and enhancement. Digital twins also can impact a person’s identity, since meaning can be assigned to the patterns in the data. Digital twins, for instance, can challenge equality
The differences between persons can be sharply defined and made extremely transparent based on the differences in their compiled information, potentially leading to segmentation and discrimination. Personal digital twins are an asymptotically data-intense scenario that clarifies the importance of governance concerning the production and use of personal biological and lifestyle data.
This is the dim view. I ascribe to it. No data is secure from bad actors, who may learn what you bought from Amazon, the model of your iPhone, and the path of your locations. Likewise, healthcare providers cannot prevent access to your most confidential information.