Human vs Machine: Computer Algorithm Matches Experts for PCI Appropriateness and Planning
More development is needed, but artificial intelligence, or AI, could offer unique solutions for cath labs lacking time or resources, one researcher says.
SAN DIEGO, CA— An algorithm can offer a decision on the appropriateness of coronary revascularization during pressure-wire pullback at least as well as expert consensus, according to a new machine-learning analysis.
“Physicians want to make the right decisions for their patients, and all of our lives are too busy to be experts in every single area,” Justin Davies, MBBS, PhD (Imperial College NHS Trust, London, England), who presented the findings last month at TCT 2018 in San Diego, CA, told TCTMD. “How wonderful is it to essentially have every single time you make a measurement—whether you look at IVUS images, OCT, physiology—to essentially say, well, if you had a group of world experts doing it, how would they do it?”
Explaining the CEREBRIA-1 study, Davies said his team included 1,008 instantaneous wave-free ratio (iFR) pullback traces, including 317 duplicates, and had both the computer algorithm and a multinational team of expert interventionalists analyze them for appropriateness of PCI and strategy. Median distal and proximal iFR values were 0.87 and 0.99, respectively, and most vessels were in the left anterior descending artery (79.5%). Almost one-third (31.4%) of vessels had significant pressure-wire drift.
They found that the computer was noninferior to the expert consensus decision for both appropriateness for PCI and for determining PCI strategy. While humans tend to change their minds about both PCI appropriateness and strategy in roughly on in 10 cases, the computers never waver, Davies highlighted. In their analysis, the researchers found that 3.8% of cases where the experts recommended PCI were “actually negative when accounting for wire drift,” he said. Also, “27% of cases determined as nonsignificant for PCI were actually positive.”
Physicians are “very open” to this type of technological aid in the cath lab, Davies said. “Generally speaking, we should always be very open to different opinions of how to treat patients, and actually most of the problems happen I think when people are not up to date on the latest technology, they're not using the latest technology in the right way. When you're very open and receptive to the latest research and technology, I think you probably treat patients better.”
Speaking with TCTMD, however, Partha Sardar, MD (Brown University, Providence, RI), who was not involved in the study, said he’s not so sure that physicians will welcome this kind of technology with open arms, at least at first. “In medicine, we are kind of resistant to change, so it's going to happen slowly, it's not going to happen suddenly,” he observed. “But when the outside world is changing, you have to change.”
Autopilot in the Cath Lab?
The biggest challenge for these kind of algorithms moving forward is going to be development, Sardar argued. “The limitation is that right now we don't have that many specialists in medicine or cardiology who know what these things are and can do studies or trials on these things,” he said. “We need to develop that field. We're pretty good at devices. We're pretty good at medicines, trials, and studies. We have so many experts. But with the new field [of artificial intelligence (AI)], we don't have that many experts in medicine and cardiology, so we need that expert who's going to do studies and trials to show that these things work and that sometimes it can be better than the cardiologist.”
Specifically, algorithms will need to be shown superior to human expert consensus before many in the field will want to use them, Sardar said. But low-volume operators or those who use imaging on a less-frequent basis might want to “use this kind of technology to kind of consult because [they] don't have the consult at [their] institution,” he suggested. “So in that case, this kind of technology is going to be really useful.”
Davies said his institution is not yet applying the algorithm, which he likened to the autopilot feature many pilots use today, but that they will start feasibility studies soon. “These alert you like a car would alert you that you're going to hit something or an airplane alerting you you're going too low, and you just pull the guide catheter back and the problem is solved,” he said. “And the computer sees it very easily and our eyes sometimes don't when we're distracted.
“This works exceedingly well already, and the level of quality is ready for patient decision-making,” Davies continued. “We'll continue to integrate this more with other modalities, so with angiography and with other imaging modalities, and also just going on to use these AI-based tools to make decisions more broadly in the cath lab.”
This would entail adding additional patient data and risk models to further refine the analysis, Davies added.
Davies J. CEREBRIA-1: machine learning vs expert human opinion to determine physiologically optimized coronary revascularization strategies. Presented at: TCT 2018. September 24, 2018. San Diego, CA.
- Davies reports receiving grant/research support, consulting fees/honoraria, royalty income, and intellectual property rights from Philips Volcano. Sardar reports no relevant conflicts of interest.