By Dr. Amanda Randles, Ph.D., Alfred Winborne and Victoria Stover Mordecai Assistant Professor of Biomedical Sciences — Duke University

The healthcare industry could use a booster shot. We now spend a record $12,000 per person, per year, on healthcare in the United States. That is far more than Switzerland — the second largest spender — and nearly three times as much as Japan, which stands at number 10 on the list. 

And yet, Americans’ life expectancies have trended downward for years. Covid accelerated innovations like telemedicine but, in doing so, it has brought us to an inflection point. We need new technologies to bring down the cost of medical care while producing superior results. A very promising one is machine learning (ML).

Simulation

ML and its cousin, artificial intelligence (AI), have long been used in radiology. We are just beginning to discover its potential in other fields, such as cardiology. In this space, we are not only using ML to analyze images, but combining it with sophisticated physics-based models to assess a patient’s blood flow. ML and AI are allowing us to simulate a wider variety of outcomes before treating the patient.

For example, we can now take data from medical images, such as an MRI or CT scan, and create a three-dimensional representation of a patient’s coronary arteries. Instead of putting an invasive guide wire in a patient to measure the pressure gradient, the ML can model their blood flow virtually. This simulation can measure what the pressure should be and what it would be after a corrective procedure. The result is a reduction in the invasive procedures placing guide wires into a patient to determine if a stent is needed.

Computational models can also be useful in treating complex lesions. In the case of a Y bifurcation, for example, we need to know which branches of the Y should be treated. Using ML-assisted flow models, we are developing methods to run a simulation and predict which patients need the main branch stented, versus which patients need both the main branch and the side branch stented — all entirely virtually. 

In both of these cases, however, we not only need to simulate the patient at both a resting and exercise state, but during real-life scenarios. What if it's cold and they're shoveling in the snow? What if they decide to go jogging at altitude in Denver? ML is helping us predict the entire state space for that patient so we can see the effect of a situational change and adjust treatment accordingly. As the ML “learns” and computing power becomes more available, we will continue to look at many different patient states and real-life scenarios.

Non-Invasive

One of the most exciting things about using virtual models of the patient is that simulations are non-invasive. The big challenge is the leveling-up in terms of sophistication — i.e., having the computing power, and experience to simulate a sufficient sampling of activities a patient might do in their lifetime. 

This effort is indeed a daunting task, but consider the current standard: if you have chest pain, you go to the hospital or clinic. If you have a blockage, you will likely go to a cath lab where a doctor puts you under and inserts a pressure wire into your coronary arteries to measure the pressure gradient. Whether it is above or below a certain cutoff determines if you get a stent. 

This process is painful, time-consuming, and expensive. Today, we already use ML to perform image segmentation to not only isolate the location of lesions, but to also reconstruct a high fidelity 3D model of a patient’s arteries. In other words, ML is already playing an essential role in identifying and building the right geometry. Machine learning can also speed up segmentation for particularly complicated diseases, such as aortic dissection, which can easily take eight or nine hours for a trained person. Using ML, however, segmentation might take just a few minutes. We’ve also seen significant strides in terms of both ML and physics-based modeling and their ability to accurately predict personalized blood flow metrics. In fact, there are FDA-approved tools that have shown they can match the pressure wire in terms of accuracy here. The next step is making modifications in order to do the treatment planning. Using simulation, doctors can try the different stents and/or modify the geometry before the stent is inserted. By combining virtual surgery with prediction of flow under different conditions, the models could reveal that, while the patient is perfectly fine in the clinic, they're going to have a problem when they go running, for instance. The virtual models allow medical providers to assess a wider range of hypothetical scenarios than otherwise possible.

Personalization

Personalized medicine is not just being able to identify the optimal stent to give a patient today, but optimizing treatment when looking ahead months and years in the future. Thankfully, we have a readily available data source because millions of people are already wearing monitoring devices on their wrists. When people use a wearable health monitoring device, like the Fitbit or an Apple Watch, ML can draw from personalized data as well as population-level data to improve guidance. We envision future ML models that might suggest that rather than exercise for an hour every day, a patient should exercise for 35 minutes in the morning in order to lower their chance of adverse effects by 40%. Or it might tell them that, if they're really tired today, it would be better for them to just go slow and do a steady run instead of an endurance routine. We’re not there yet, but we are gathering the data and ML is learning from all of our experiences.

Personalization can also be a powerful tool in the clinic. Virtual models provide essentially a digital twin of the patient’s arteries and allow doctors to test out many different procedures before they ever go into the operating room. We have a pediatric cardiology project where we are building tools to allow surgeons to test two key available interventions for infants with a particular congenital heart disease. Armed with that knowledge, doctors can choose the best procedure and even fine-tune if for that patient before the intervention. That is very powerful, because even a slight angle adjustment may have a signficant impact on the blood flow, which will change how that artery is going to grow over time.

Transparency

The final benefit is that ML can bring patients into the decision loop, though it now all too often seems like a black box. Going forward, patients (and doctors) will want to know how the ML made its decisions, and they should know. We need to design technology that “opens the hood” and lets patients see not only the result, but the steps that the ML took to get there. No one wants a black box making clinical decisions. In the era of simulated medicine, transparency will be the new “second opinion.”

ML will also give us easier ways of identifying patients who need intervention in real time. How we find them, track them, and ultimately treat them will be revolutionized by ML-driven wearables, but we will need large-scale adoption before that takes place. Massive and affordable processing power is another necessity. Interpretability and data may not be everything, but physics-based simulations are an exciting way of getting to more effective, affordable, and informed treatments.