Digital twins could allow engineers to move beyond testing whether a medical device works to predicting how it will perform in a specific patient. Here’s how researchers are working towards a more personalised approach.
Imagine a 65-year-old patient who needs a heart valve replacement.
There are 10 different devices available, all with slightly different designs. The clinical data shows that all of these valves have performed safely in other patients. But that data does not necessarily tell you which one is going to perform best for this individual.
Now imagine the same patient has a detailed digital counterpart.
Engineers could look at how the valve interacts with the patient’s anatomy, how blood flows through each design and whether there are differences that might only become apparent years down the track.
This is the future that digital twins could help us move towards.
A digital twin is essentially a computational representation of a real-world system. In medicine, that could mean representing an individual patient, a medical device or the interaction between the two.
The technology is still emerging, but researchers are already developing many of the components needed to make this possible.
READ: These engineers are creating living blueprints to help understand natural systems
Stage one: Build a digital version of the patient
The first step is creating a model that captures enough information about an individual patient to make simulation useful.
Each patient’s anatomy is unique, and so is their specific defect or pathology. Where medical devices have traditionally been designed for relatively broad patient cohorts, what we are trying to do now is move towards a much more personalised approach.
This approach is being tested through various experiments. For example, in our research at The University of Sydney, we have collaborated with U.S-based Massachusetts General Hospital to develop a digital twin of a sheep to trial next-generation scaffold implants.
If we can represent different responses computationally, we can stop asking whether a device “works”, and start asking whether it’s effective for the specific patient we’re talking to.
Stage two: Test thousands of possibilities
Once we have a digital representation of the patient, the next step is to use it to test different options.
If a patient is having a hip reconstruction, for example, we could potentially simulate thousands of different designs and determine which is most likely to produce an optimal outcome for that individual.
We have already been exploring this idea in jaw reconstruction. During my PhD studies at the University of Sydney, I developed a surgical planning tool that uses CT scan data to build a virtual patient, and then rapidly simulates different designs before 3D-printing the final, optimal design.
The advantage is that simulation allows us to explore a much larger design space than we could realistically investigate through physical prototyping alone.
Stage three: Model the whole treatment journey
A medical device does not just exist at the moment it is implanted. Its performance can change as the patient’s body responds to it, which means we ultimately need to model more than the device and the anatomy at a single point in time.
In our work, we have taken a jawbone and modelled it when it is intact. We can then create a defect, reconstruct it with a particular scaffold implant and, ultimately, simulate how the bone might heal around that implant over time.
The digital twin could also be updated using information from the patient. For example, imagine an implant with a sensor attached. You could track the implant deformation and feed that data back into the model to continuously recalibrate the model predictions with the latest data.
This gives us higher confidence in the model predictions. Think 10 years post-implantation: the patient will be different. That data could be fed back into the model. If the model predicts 39 Nm of deformation but the implant measures 40, we can compare those two results and use the difference to improve the model.
Over time, that could allow the digital twin to become a more accurate representation of how the patient and device are actually behaving – rather than a static model that is created once and then left unchanged.
Stage four: Prove the model is trustworthy
If we are going to use these models to inform decisions about patients, we need to establish that their predictions are sufficiently reliable for their intended purpose.
There are a number of different ways of doing this. You might have a representative experiment and compare the results with your simulation. You might have data from a real patient. Or, eventually, you might have data coming directly from a smart implant.
A lot depends on what happens if the model is wrong. If the consequences are relatively minor, you can potentially accept a wider tolerance of error. However, if an error could lead to fatal complications, that same level of uncertainty is unacceptable.
Stage five: Turn simulations into evidence
Once we can demonstrate that these models are making trustworthy predictions, the next question is whether we can use them as evidence.
The two main hurdles here are regulatory approval and reimbursement.
The Therapeutic Goods Administration and the US’s Food and Drug Administration need to determine the evidence required to show that a medical device is safe to put into people. Separately, healthcare systems need to determine whether a device provides enough clinical benefit to justify the cost of providing it.
This is where I think computer simulated evidence could become particularly compelling. If we can generate reliable evidence from computer simulations, that evidence could eventually contribute to both regulatory approval and reimbursement.
The idea shouldn’t be that we simply replace clinical trials with computer models, but there are questions that are very difficult to answer with conventional clinical trials.
Take the heart valve example. You might have two valves where, after two years, neither has had any deaths associated with it. You could conclude that they have performed similarly. But can we really conclude that one device is superior to another?
One valve might have a better leaflet design that allows blood to flow through it more effectively. A high-resolution computer simulation could show that one valve is producing more turbulent flow or has a higher clotting potential. You could then start to ask what that difference might mean five years down the track.
That is the sort of information that could complement clinical evidence rather than replacing it. Computer models are more sensitive at predicting device performance than what can be measured in a clinical trial.
READ: This new tool could change the way surgeons plan jawbone reconstruction
What needs to happen next?
There are a lot of hurdles between where we are now and that future. The models need to be validated. Standards need to develop. Regulators need to determine how computational evidence should be treated. And ultimately, the models need to be sufficiently robust to generate trustworthy results.
The interesting question for biomedical engineering is not just whether we can build a digital twin. It is whether we can build one that is sufficiently accurate and trustworthy to help us make better decisions about a real patient.
If we can do that, then we move from using computers simply to design medical devices to using them to understand how those devices could perform across the entire life cycle of a patient.
Register now for Engineers Australia’s webinar Medical devices: from concept to clinical translation
