“ Je pense, donc Je suis. [I think, therefore I am.]”
Using a handheld transducer placed on the epicardium, the first examples of perioperative ultrasound using M-mode were published in 1972. The authors wrote, “Considerable technical skill and practice are required to recognize the echocardiographic patterns obtained from intracardiac structures. […] To date no automatic method of pattern recognition has been devised. Consequently, the operator must be skilled in the methodology if accurate results are to be obtained.” The need for automation to overcome limits of cardiac ultrasound was self-evident to the first perioperative ultrasoundographers.
The need for automation remains clear today. I recently presented a rather routine case of functional mitral regurgitation to an international audience of senior cardiologists and asked them to grade the severity. The audience’s response ranged from 2+ to 4+. While surprising to the audience, I had anticipated this spectrum of result. Biner et al , in 2010 asked 18 echocardiologists from the United States, Japan, and Israel to interpret echocardiographic exams from 16 patients with mitral regurgitation using a web-based platform. Agreement was strikingly uncommon, even between expert observers looking at the same echo clips.
There are technical and psychological explanations for this variation. Much of the technical limitations relate to the inability of one-dimensional static metrics such as vena contracta width and the PISA radius to accurately describe the nonlinear and dynamic hydraulics of in vivo . While there are other metrics to consider besides the two mentioned here and elaborate scoring systems have been proposed, for most clinicians there is undoubtedly a psychological basis for the relative weight assigned to individual parameters. Indeed, it is difficult to disregard a jet with a large color area despite multiple reports that reliance on this metric can be misleading. On a deeper level, while we consider ourselves to be highly intelligent with years of accumulated echocardiographic experience and thus assume ourselves capable of applying guidelines in an objective manner, this conclusion is likely not true.
Looking outside echocardiography, the myth of “legal formalism” holds that judges apply legal reasoning in an almost “mechanical” manner. Justice Oliver Wendell Holmes in 1881 alternatively wrote “the life of the law has not been logic; it has been experience.” A recent sobering study cited over 500 times sought to test the caricature that in parole decisions, justice is “what the judge ate for breakfast.” The authors found experienced judges indeed made more favorable parole rulings after their food breaks. I suspect that consistency and accuracy of echocardiographic interpretations is marred too by not only technical limitations of technology but also by a myriad of cognitive biases, perhaps including what we ate for breakfast.
Gordon Moore, cofounder of Intel, in 1965 predicted an exponential growth in the number of components per integrated circuit. His prediction has proven true. The growth in computer processing, especially the recent development of the graphic processing unit, has turned out to be the major force driving advances in echocardiography. The notion of artificial intelligence was largely science fiction when the first perioperative M-mode images were being obtained. Today, intelligent algorithms have become commonplace in our lives. Most of us are familiar with advanced facial recognition software used by companies such as Facebook. Indeed, even our laws are being rewritten to allow for self-driving cars. Computer science experts predict that artificial intelligence will exceed human performance across a variety of fields including surgery by the middle of this century.
While the potential use of intelligent algorithms has been explored in medicine in the past, more recent studies have suggested computers indeed may perform better than humans in some visual recognition with obvious implications in medical imaging. The incorporation of advanced artificial intelligence systems in the field of echocardiography has lagged behind the rest of society that is witnessing its rapid promulgation. This may be related to the relatively large lag time between developing an idea, obtaining approval, and marketing the product to medical consumers. Until recently, the application of intelligent algorithms within echocardiography has been limited to early pilot projects testing important, but clinically limited, applications such as ventricular hypertrophy and pericarditis. Indeed, such early studies were the basis of Dr. Partho Sengupta’s 14 th Feigenbaum lecture at the Scientific Sessions of the American Society of Echocardiography in 2013 as he proposed that we were on the verge of entering into a new “Intelligent Era” in echocardiography.
In the new era, automated platforms will become more common and help us quantify complex anatomic structures and function, and compute in vivo cardiac hydraulics with little to no user input. Recently, commercial vendors have introduced a variety of automated platforms to clinicians. The use of artificial intelligence in the operating room helped considerably in the quantification of functional mitral regurgitation and predicting annuloplasty band length in mitral repairs. The need for this technology in the perioperative arena is becoming even more evident in catheter-based interventions. Indeed, such interventions rely on the geometric measurements provided by the echocardiographer. Platforms that provide nearly immediate, reproducible, and accurate ventricular and valvular assessments are clearly valuable.
In the early 1970’s, the pioneers of perioperative imaging commented that “considerable technical skill and practice [were] required to recognize the echocardiographic patterns,” and hinted at the need for automation. Unfortunately, even in 2017, perioperative echocardiography still remains far from perfect. Clinicians pressed for time frequently employing the time-honored “eyeball” method. The pressure for immediate answers in the time sensitive perioperative environment often precludes detailed three-dimensional analyses of cardiac structures, particularly while managing unstable patients. New intelligent platforms enable the perioperative echocardiographer to perform high quality sophisticated and detailed dynamic analyses on three-dimensional data sets with just the push of a button. One of the most important factors in the development of this technology will be a tight collaboration between our industry partners developing the software and clinicians to ensure that the data algorithms generate answers relevant to the perioperative environment and truly enhance consistency, quality and workflow. The use of intelligent algorithms in echocardiography is still in its infancy and perhaps, while we are myopically focused on immediate clinical needs, it is admittedly hard for us to see beyond the horizon. However, if the computer scientists are correct, the eventual question may not be if we will need artificial intelligence in echocardiography, but rather if the algorithms will need us.
Frederick C. Cobey, MD, MPH, FASE, an Assistant Professor of Anesthesiology and Division Chief of Cardiac Anesthesiology at Tufts Medical Center, is a member of the ASE Council on Perioperative Echocardiography Steering Committee.