Autonomously Driven: Artificial Intelligence in Cardiothoracic Surgery

Artificial intelligence offers myriad novel applications that will influence health care. In this issue of The Annals of Thoracic Surgery, Baxter and colleagues¹ proffer a glimpse into the broad spectrum of digital technologies relevant to the modern cardiothoracic surgeon. Through a Donabedian lens, this is likely to influence structure, process, and outcome. The coterie of options in this regard is as promising as it is boundless. Within the realms of data analysis, for example, we shall no longer restrict our variable selection to the confines of standard logistic regression, limited by requisite dependence on precedence, preselected variables, or biological plausibility. Machine learning will unleash the power and flexibility of convolutional deep neural networks. These networks closely mimic the connectivity of the human brain, with multilayered arborization and a multidirectional flow of information.² We will now evaluate thousands of variables simultaneously, proffering the option of considering the unlikely, evaluating the implausible, and offering unique, unconventional solutions to conventional hypotheses. Indeed, this heralds the ultimate mitigation of algorithm bias that has characterized human observation for decades.

The process, nevertheless, may unearth statistically significant predictors that have no clinical relevance whatsoever: a tacit reminder that human input cannot be entirely supplanted in the realms of clinical decision-making. To this end, the authors are justified in their emphasis on “accountability, liability, and culpability” that will accompany the increased use of autonomous systems.¹ The responsibility still falls squarely on the surgeon. The applications and uses of machine learning are complex and detailed and involve multiple techniques such as regularization (a means to avoid overfitting), supervised and unsupervised training, discrimination, and calibration.²

These techniques may be applied to advanced algorithms, which can then be tested on data from remote monitoring devices, video footage from robotic procedures, 3-dimensional models, and complex networks, thus integrating big data in real time. Currently, neural networks are in use and correlate various cardiac sounds with pathological conduction patterns with over 90% accuracy.² Furthermore, precision medicine will likely revolutionize genomic studies designed to predict risks from genetic diseases such as Loeys-Dietz and Marfan syndromes.³ The challenge in applying machine learning algorithms, however, is in defining clinical practice utility. Whether these algorithms will change our risk prediction, enhance our clinical acumen, or improve our outcomes remains to be seen. The expectations are high, and thus a tempered and thoughtful adoption is likely to be the best approach to safeguard against disappointment as we prepare to surf this new digital wave.