How machine learning can produce better health outcomes

Medical technology helps

In the past two decades, health-care spending in Canada has more than tripled.

Among the 38 countries in the OECD, Canada’s per capita spending on health care is among the highest, at $7,507 – less than in Germany at $8,938 but more than in Australia at $7,248. Per capita spending remained highest in the United States ($15,275), where life expectancies have actually been declining since 2014. (All currencies in Canadian dollars).

COVID is partially to blame for the poor outcomes, but the pandemic has accelerated technological innovations like telemedicine and mRNA vaccines. Nevertheless, we will still need more such tools to bring down the cost of medical care while producing superior results. One promising technology is machine learning, or ML.

ML and artificial intelligence (AI) have long been used in radiology and other diagnostic fields. We are just beginning to discover its potential in other areas that may prove more valuable. Rather than teaching a computer how to read an MRI or x-ray, for instance, we can use ML in cardiology to speed up the simulation of a variety of outcomes before treating the patient.

For example, we can now take data from medical images and create a three-dimensional representation of a patient’s coronary arteries. Rather than insert an invasive guide wire in an anesthetized patient to measure the pressure gradient – as is the common practice now – physics-based modelling augmented with ML can model their blood flow virtually.

This simulation can tell us what the pressure should be and what it would be after a corrective procedure. In other words, for many patients, we no longer need to perform an invasive procedure to determine if a patient needs a stent.

Non-invasive simulations are of huge benefit to both patients and health-care budgets. However, they create a new challenge for prospective use—developing the bandwidth and computing power to simulate everything a patient might do in their lifetime.

Today, we have U.S. FDA-approved tools that have shown they can match the pressure wire in regard to accuracy. The logical next step is using ML to complement flow simulations in order to provide intuitive, interactive methods for treatment planning. For example, these virtual models would allow doctors to “test drive” different stents and/or modify the geometry before the stent is inserted.

Ultimately, we want to train an ML model that can adapt for any individual patient and predict blood flow response under any real-life scenario they will experience in their lifetime, beyond capturing metrics measured in the clinic. We see integration with data from wearable devices as offering unprecedented potential for accessing continuous data as people go about their normal activities.

When people use a cloud-based wearable, like the Apple Watch, ML can leverage lessons learned at the population level to make personalized predictions. The field has already seen strong success using these methods to detect cardiac arrhythmia or response to viral infections.

In our work, we are combining the continuous data from wearables with personalized simulations with the goal of improving personalized guidance. For example, the ML might suggest that a patient exercise less to lower their chance of an adverse event, or it might tell them to do a steady run instead of an endurance routine. We’re not quite there yet, but we are gathering the data we need and continuously advancing the ML and computational methods.

In the clinic, simulations are also getting more personal. Much of our recent work has focused on developing digital twins of a patient’s vasculature that clinicians could leverage to virtually test a range of different surgical procedures before they go into the operating room.

For example, we are working with a team of pediatric cardiologists to establish a framework for virtually testing different interventions to assist infants with congenital heart disease.

Our goal is to arm doctors with the knowledge of what works best so that they can choose the right procedure for that patient, and even optimize the treatment planning before operating. This is extremely valuable because even a slight adjustment to a shunt angle may impact blood flow, which will change how that artery grows over its lifetime.ML can also bring patients into the decision loop in a meaningful way. Patients (and their doctors) will want to – and should – know how ML made its decisions. We are designing technology that “opens the hood” and lets patients see not only the result but the steps that the ML took in interpreting the data because no one wants a black box making clinical decisions.

We have a unique opportunity to leverage continuous data from wearables that track a patient’s condition under real-world conditions alongside high-fidelity, virtual models of future actions. In the era of simulated, ML-assisted medicine, both personalization and transparency are critical.

Amanda Randles is the Alfred Winborne and Victoria Stover Mordecai Assistant Professor of Biomedical Sciences at Duke University in Durham, North Carolina.

This article is written by or on behalf of an outsourced columnist and does not necessarily reflect the views of Castanet.

More Writer's Bloc articles

About the Author

Welcome to Writer’s Bloc, an opinion column for guest writers to share their experiences and viewpoints with our readers.

Do you have something to say that is timely? of local interest? controversial? inspiring? foodie? entertaining? educational?

Drop a line. [email protected]

Opinions expressed in this column are those of the writer and do not necessarily represent those of Castanet. They are not news stories reported by our staff.

The views expressed are strictly those of the author and not necessarily those of Castanet. Castanet does not warrant the contents.

Previous Stories