Algorithmic Versus Expert Human Interpretation of Instantaneous Wave-Free Ratio Coronary Pressure-Wire Pull Back Data

Objectives The aim of this study was to investigate whether algorithmic interpretation (AI) of instantaneous wave-free ratio (iFR) pressure-wire pull back data would be noninferior to expert human interpretation. Background Interpretation of iFR pressure-wire pull back data can be complex and is subjective. Methods Fifteen human experts interpreted 1,008 iFR pull back traces (691 unique, 317 duplicate). For each trace, experts determined the hemodynamic appropriateness for percutaneous coronary intervention (PCI) and, in such cases, the optimal physiological strategy for PCI. The heart team (HT) interpretation was determined by consensus of the individual expert opinions. The same 1,008 pull back traces were also interpreted algorithmically. The coprimary hypotheses of this study were that AI would be noninferior to the interpretation of the median expert human in determining: 1) the hemodynamic appropriateness for PCI; and 2) the physiological strategy for PCI. Results Regarding the hemodynamic appropriateness for PCI, the median expert human demonstrated 89.3% agreement with the HT in comparison with 89.4% for AI (p < 0.01 for noninferiority). Across the 372 cases judged as hemodynamically appropriate for PCI according to the HT, the median expert human demonstrated 88.8% agreement with the HT in comparison with 89.7% for AI (p < 0.0001 for noninferiority). On reproducibility testing, the HT opinion itself changed 1 in 10 times for both the appropriateness for PCI and the physiological PCI strategy. In contrast, AI showed no change. Conclusions AI of iFR pressure-wire pull back data was noninferior to expert human interpretation in determining both the hemodynamic appropriateness for PCI and the optimal physiological strategy for PCI.

R evascularization in stable coronary artery disease should be performed only for ischemia-producing coronary lesions (1)(2)(3)(4). Physiological measurements obtained using a coronary pressure-wire permit the identification of myocardial ischemia on a per vessel basis (5). Consequently, coronary physiology is recommended in international treatment guidelines (6)(7)(8) to guide revascularization decision making.
In addition to vessel-level ischemia detection, under resting conditions, a coronary pressure wire can also be used to produce an instantaneous wave-free ratio (iFR) pressure-wire pull back trace: a longitudinal assessment of coronary pressure loss along the length of a coronary artery. Such a trace permits the identification of lesion-level ischemia, as well as the ability to predict the physiological outcome following a proposed percutaneous coronary intervention (PCI) revascularization strategy (9). However, in the absence of clinical outcome data, a definitive interpretation of iFR coronary pressure-wire pull back data is lacking. Individual interpretation of coronary pressure-wire pull back data is complex, subjective, and dependent on the physiological expertise of the operator.
Algorithmic interpretation (AI) of coronary pressure-wire pull back data may help circumvent these limitations. Within this study, we aimed to determine if AI of iFR coronary pressure-wire pull back data could provide a standardized alternative to expert-level human interpretation. The coprimary hypotheses of this study were that AI would be noninferior to the interpretation of the median expert human in determining: 1) the hemodynamic appropriateness for PCI; and 2) the physiological PCI strategy, compared with the expert heart team (HT) opinion.   Cohen's kappa ¼ 0.70) ( Figure 3A).

RESULTS
Across the 372 cases that the HT determined were hemodynamically appropriate for PCI, 14 cases (3.8%) had hemodynamically nonsignificant physiology due to physiologically significant pressure-wire drift not corrected for by the consensus ( Figure 4A). In contrast, using AI, there were no cases in which PCI was determined appropriate for hemodynamically nonsignificant physiology, as pressure-wire drift was always identified by the computer.
There were 319 cases that were determined as not hemodynamically appropriate for PCI according to the HT. Of these, 86 cases (27.0%) had hemodynamically significant physiology ( Figure 4B) that was not identified by the HT. In contrast, there were 296 cases that were determined as not appropriate for PCI according to AI. Of these, 49 cases (16.6%) had hemodynamically significant but diffuse nonfocal physiology ( Figure 4C).

DISCUSSION
In this study we demonstrated that AI of iFR pressure-wire pull back data was noninferior to the interpretation of the median expert human in determining both the hemodynamic appropriateness for PCI and the physiological PCI strategy when judged against an expert HT opinion. AI correctly interpreted physiologically significant coronary pressure gradients and modified treatment accordingly in the presence of pressure-wire drift (Central Illustration).
HT DECISION MAKING. Group decision making has become commonplace in cardiology, with the role of the HT well established in the management of complex clinical decision making (11)(12)(13). Additionally, group decision making can be valuable in areas of medicine in which the optimal treatment approach remains uncertain because of a lack of clinical outcome data. In that regard, the interpretation of coronary pressure-wire pull back data is often complex, and a treatment plan must usually be decided upon instantaneously. However, practically speaking, HT opinion for this task is rarely available. As such,   Not captured within our study methodology are the additional barriers that exist to iFR pull back data interpretation in real-world physical environments.
Such barriers include, for example, limitations of the human visual system to discern small hemodynamic gradients on a physiology screen that may be positioned at a distance away from the operator. Additionally, the very nature of performing an invasive physiological assessment necessitates multitasking of the operator, whose ability to focus solely on interpretation of hemodynamic data is limited. However, within our study, these types of barriers were not replicated, and thus their influence on human decision making remains unmeasured. Although speculative, it is possible that true real-world interpretation of coronary pressure-wire data may be more variable and heterogenous than that recorded within our study.
Last, because of the retrospective nature of our dataset, the actual clinical decisions made at the time of iFR pull back measurement were neither recorded nor informed by knowledge of the AI.

CONCLUSIONS
AI of iFR pressure-wire pull back data provided a standardized interpretation that was automatically corrected for the presence of pressure-wire drift.
When judged against an expert HT opinion, AI was noninferior to that of the median expert human in determining both the hemodynamic appropriateness for PCI and the PCI strategy.