How ‘double dipping’ (with AI) can keep our patients healthier

What’s the best use case for AI in radiology? This has to be a strong contender.

First, a bit of background. Radiologic examinations, such as CT of the chest, abdomen and pelvis, are typically performed for targeted indications.?In the US alone, we conduct more than 80 million CT scans, every year. They are so common because they’re used to diagnose, monitor, and screen for everything from trauma to cancer and pneumonia.

And when you have a chest CT scan, for example, the equipment takes multiple ‘slices’ as individual images through the anatomic area – meaning those 80 million CT scans a year equate to billions of images.

This staggering figure has been rising for decades and isn’t set to plateau anytime soon. Just last year, USPFTS recommended annual chest CT screening for lung cancer, driving even more scans, and even more images.

This means that the millions of CTs, after addressing the primary clinical question (lung diseases for example), are then stored in ‘archive’ in healthcare systems – in fact medical imaging as a whole currently occupies more than 95% of all healthcare data!?

In the past decade, there have been researchers looking into how to use the parts of the imaging data that may give important information about other diseases (problems that were ‘discovered’ on the CT images but may not have been related to the reason the exam was performed).

How would this work??Well, take the chest CT: every chest CT captures images of the patient’s heart. But – despite routinely imaging the heart and other structures in the chest, most chest CTs don’t provide enough information about the heart for a specialist to determine the patient’s risk of heart disease. That’s because, in routine chest CTs, the motion of the heart during scanning prevents a detailed evaluation of coronary arteries. (Yes, this is a rare case in which a beating heart is a barrier to improving patient outcomes!) So while the routinely acquired chest CT examination can provide experts lots of information about certain structures (especially the lungs and airways) it cannot provide the detailed information about the heart required to quantify coronary artery disease.

Instead, to quantify coronary artery disease, typically by measuring the amount of calcification in the arteries, an expert would need to see images from a specialized type of chest CT – one that often requires ECG gating, beta-blockers, 3D labs – and crucially, more time and money. ?And these types of specialized chest CT scans aren’t usually covered by a patient’s health insurance.

So, to recap: every year, millions of Americans receive routinely acquired chest CT scans that generate images of their hearts. But those images don’t provide the information specialists need to quantify their risk of heart disease. (Which, let’s not forget, is our nation’s leading cause of death.)

Where does the ‘double dipping’ come in?

Well, imagine these chest CT images are an ever-growing pile of half-eaten chips. They’ve been used once, and now they’re sitting on the side of healthcare’s plate, with so much more to give.*

We know human specialists can’t currently ‘double dip’ these routine chest CTs by reviewing the blurred images of the beating heart to quantify coronary artery disease. So, a group of researchers asked the question, “What about AI? Could AI do it?”

It’s a great question, not least because it’s an uncommon one. Most AI for chest CTs is designed to mimic what humans can do – for example, diagnose pulmonary nodules, quantify emphysema, or determine if the patient has COVID-19.

Instead, this AI was designed to do something superhuman: to provide quantitative heart disease diagnosis based on perfectly routine, absolutely not fancy chest CTs with the heart beating away in all its blurred glory.

I know this, because I was lucky enough to be involved in this very cool, very exciting work. Under the incredible leadership of Bhavik Patel and Nishith Khandwala we developed an end-to-end AI model with the potential to impact the fight against heart disease in multiple ways – and best of all it can be done on data that we already acquire in the routine care of patients.

The main use case is easy to see. Millions of asymptomatic people, unaware that their risk of having a heart attack is high, could be notified using data that healthcare systems already have. Let’s say you have a routine chest CT scan as part of annual screening for lung cancer or because you had a bad car accident. AI could ‘double dip’ these images – images that would be created, regardless – and tell you something that ultimately saves your life.

This kind of superhuman, opportunistic screening can be applied to CT scans of other parts of the body too, like the abdomen.

For example, quantification of muscle and fat in and around organs – something that cannot be done by human experts in routine care without AI – can give prognostic data about nutritional status, dangerous liver disease, and metabolic syndrome. The bones of the spine can give information about future fracture risk. And, like the chest CT example, quantification of the calcifications in the large arteries in the abdomen and pelvis can also provide insight into heart disease and stroke risks.

There is lots of ongoing research to quantify and standardize these measures to connect them to overall health outcomes but the idea is that AI + human experts analyzing routine imaging data in healthcare systems hold the potential to unlock new insights that could save lives.

How would this work? Well, for example, these specialized AI models could be applied retrospectively to all the CT imaging studies stored by large healthcare systems, allowing them to identify high-risk patients, and intervene with optimized medical management at the population health level, without requiring any additional testing. Amazing, right?

I call this ‘superhuman opportunistic screening’ but it’s really just double dipping, levering routinely collected and stored medical imaging data to help take better care of our patients. And this is just one area in which the work being done by AI researchers today will shape the healthcare outcomes of tomorrow.

*Apologies to anyone who clicked through hoping to find scientific justification for their dinner party practices. Fast-tracking your next bite of salsa or guacamole might momentarily improve your mood, but the evidence suggests literal double dipping is far from being a net positive for public health: https://www.scientificamerican.com/article/is-double-dipping-a-food-safety-problem-or-just-a-nasty-habit/)