EuroIntervention PCRonline - Connecting with the community

Coronary interventions

From rest to hyperaemia: initial validation of a data-driven approach for functional assessment of coronary lesions

This study evaluated a new approach to infer the hyperaemic cycle-averaged pressure ratio (Pd/Pahyperemia)computed using a machine learning (ML) algorithm that incorporates resting state pressure and imaging measurements.

Invasive aortic (proximal) and coronary (distal) pressures were measured at rest and at adenosine-induced hyperaemia using a pressure wire in 43 patients with suspected coronary artery disease (CAD). Eighty-six lesions were included in the study, 44 left anterior descending (LAD), 21 left circumflex (LCx), 21 right coronary artery (RCA). The measured pressure traces were processed offline to determine Pd/Parest and FFR as ratio of cycle-averaged distal pressure to aortic pressure at (respectively) rest and hyperaemia. Pd/Parest has recently been validated against FFR for functional assessment of CAD with an optimal cut-off value of 0.92 for determining haemodynamically significant lesions. Routine coronary angiograms were acquired under resting conditions and two angiographic views were selected to reconstruct a three-dimensional anatomical model of each diseased vessel. To determine (Pd/Pahyperemia)computed a support vector regression algorithm was trained on invasively measured Pd/Pahyperemia (i.e., FFR). This ML algorithm uses as input a set of features composed from the invasively measured Pd/Parest value and geometric characteristics extracted automatically from the reconstructed anatomical model. Since functional coronary measures are influenced by both the upstream and the downstream circulation, the features encapsulate local, upstream and downstream information. Local features are based on the effective radius of the lumen; upstream and downstream features are related to characteristics of the most significant stenoses (minimum radius, lengths, percentage diameter reduction, etc.). The 86 lesions were randomly divided into two sets: 60% (51 lesions) were used for the training and 40% (35 lesions) for the validation of the proposed (Pd/Pahyperemia)computed against invasively measured FFR. From the 35 lesions of the validation set (LAD: 19, LCx: 10, RCA: 6), 7 were haemodynamically significant (FFR ≤0.8). The diagnostic accuracy of  (Pd/Pahyperemia)computed was 94.3% (81%-97%) (sensitivity: 85.7% [49%-97%], specificity: 96.4% [82%-99%], positive predictive value (PPV): 85.7% [49%-97%], negative predictive value (NPV): 96.4% [82%-99%]). The diagnostic accuracy of Pd/Parest was 80.0% (64%-90%) (sensitivity: 71.4% [36%-92%], specificity: 82.1% [64%-92%], PPV: 50.0% [23%-76%], NPV: 92.0% [75%-97%]). The receiver operating characteristic curves were computed and the area under the curve (AUC) was 0.97 for (Pd/Pahyperemia)computed and 0.78 for Pd/Parest. The former led to a net reclassification improvement of 0.29 when compared to the latter. Correlation between (Pd/Pahyperemia)computed (0.83±0.12) and invasive FFR (0.82±0.12) was r=0.93. The mean difference, average absolute error and the standard deviation were 0.002, 0.034 and 0.047, respectively.

(Pd/Pahyperemia)computed determined from coronary angiography and invasively measured Pd/Parest has a high diagnostic accuracy and correlates closely with invasively determined FFR, thus potentially obviating the need for hyperaemia. Future research activities will focus on the refinement of the mapping and further validation of the method. The feature (mentioned herein) is not commercially available. Due to regulatory reasons its future availability cannot be guaranteed.