Intravascular ultrasound-based machine learning for predicting fractional flow reserve in intermediate coronary artery lesions

Journal Pre-proof Intravascular ultrasound-based machine learning for predicting fractional flow reserve in intermediate coronary artery lesions June-Goo Lee, Jiyuon Ko, Hyeonyong Hae, Soo-Jin Kang, Do-Yoon Kang, Pil Hyung Lee, Jung-Min Ahn, Duk-Woo Park, Seung-Whan Lee, Young-Hak Kim, Cheol Whan Lee, Seong-Wook Park, Seung-Jung Park PII:

S0021-9150(19)31552-7

DOI:

https://doi.org/10.1016/j.atherosclerosis.2019.10.022

Reference:

ATH 16104

To appear in:

Atherosclerosis

Received Date: 2 September 2019 Revised Date:

25 October 2019

Accepted Date: 31 October 2019

Please cite this article as: Lee J-G, Ko J, Hae H, Kang S-J, Kang D-Y, Lee PH, Ahn J-M, Park D-W, Lee S-W, Kim Y-H, Lee CW, Park S-W, Park S-J, Intravascular ultrasound-based machine learning for predicting fractional flow reserve in intermediate coronary artery lesions, Atherosclerosis (2019), doi: https://doi.org/10.1016/j.atherosclerosis.2019.10.022. This is a PDF file of an article that has undergone enhancements after acceptance, such as the addition of a cover page and metadata, and formatting for readability, but it is not yet the definitive version of record. This version will undergo additional copyediting, typesetting and review before it is published in its final form, but we are providing this version to give early visibility of the article. Please note that, during the production process, errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. © 2019 Published by Elsevier B.V.

Intravascular ultrasound-based machine learning for predicting fractional flow reserve in intermediate coronary artery lesions

June-Goo Leea, Jiyuon Koa, Hyeonyong Haeb, Soo-Jin Kangb, Do-Yoon Kangb, Pil Hyung Leeb, Jung-Min Ahnb, Duk-Woo Parkb, Seung-Whan Leeb, Young-Hak Kimb, Cheol Whan Leeb, Seong-Wook Parkb, Seung-Jung Parkb

a

Biomedical Engineering Research Center, Asan Institute for Life Sciences, Seoul,

Korea, bDepartment of Cardiology, University of Ulsan College of Medicine, Asan Medical Center, Seoul, Korea

*Correspondence: Soo-Jin Kang, Dept. of Cardiology, Asan Medical Center, University of Ulsan College of Medicine, 88, Olympic-ro 43-gil, Songpa-gu, Seoul 05505, Korea E-mail: [email protected]

1

ABSTRACT Background and aims. Intravascular ultrasound (IVUS)-derived morphological criteria are poor predictors of the functional significance of intermediate coronary stenosis. IVUS-based supervised machine learning (ML) algorithms were developed to identify lesions with a fractional flow reserve (FFR) ≤0.80 (vs. >0.80). Methods. A total of 1328 patients with 1328 non-left main coronary lesions were randomized into training and test sets in a 4:1 ratio. Masked IVUS images were generated by an automatic segmentation model, and 99 computed IVUS features and six clinical variables (age, gender, body surface area, vessel type, involved segment, and involvement of the proximal left anterior descending artery) were used for ML training with 5-fold cross-validation. Diagnostic performances of the binary classifiers (L2 penalized logistic regression, artificial neural network, random forest, AdaBoost, CatBoost, and support vector machine) for detecting ischemia-producing lesions were evaluated using the non-overlapping test samples. Results. In the classification of test set lesions into those with an FFR ≤0.80 vs. >0.80, the overall diagnostic accuracies for predicting an FFR ≤ 0.80 were 82% with L2 penalized logistic regression, 80% with artificial neural network, 83% with random forest, 83% with AdaBoost, 81% with CatBoost, and 81% with support vector machine (AUCs: 0.84–0.87). With exclusion of the 28 lesions with borderline FFR of 0.75–0.80, the overall accuracies for the test set were 86% with L2 penalized logistic regression, 85% with an artificial neural network, 87% with random forest, 87% with AdaBoost, 85% with CatBoost, and 85% with support vector machine. Conclusions. The IVUS-based ML algorithms showed good diagnostic performance for identifying ischemia-producing lesions, and may reduce the need for pressure wires.

2

Key words: intravascular ultrasound, machine learning, artificial intelligence, fractional flow reserve

3

Introduction As revascularization therapy based on the presence of objective ischemia improves clinical outcomes, fractional flow reserve (FFR), a standard tool for lesionspecific hemodynamic assessment, has been used to decide whether to treat intermediate coronary artery lesions [1-4]. Despite the abundant evidence showing the clinical impact of FFR-guided percutaneous coronary intervention (PCI), the need for drug-induced hyperemia, a prolonged procedure time, and the short-term cost may restrict the widespread use of FFR in clinical practice [5,6] Intravascular ultrasound (IVUS) is a valuable tool for planning PCI, because it can provide information on lesion characteristics, vessel size, lesion length, the extent of calcification, and possible procedure-related complications. With validated criteria for optimizing stent deployment, IVUS-guided PCI has been recommended as a method of reducing major adverse cardiac events [7-9]. However, the diagnostic accuracy of IVUS-measured minimal lumen area (IVUS-MLA), one of the many clinical and local determinants of functional significance, is less than 65–70% for the identification of ischemia-producing lesions with an FFR ≤ 0.80 [10-12]. Because of an inability to integrate the anatomical and physiological parameters, both FFR and IVUS are necessary for making treatment decisions and optimizing PCI, thus increasing medical expenses. Even with an abundance of clinical and IVUS information, the development of a prediction model using traditional statistical methods is limited by the non-linearities between factors and outcomes, interactions among variables, and too many predictors [13]. Recently, machine learning (ML) techniques have emerged as highly effective computer algorithms for the identification of patterns in large datasets with a multitude

4

of variables, thus facilitating the building of models for data-driven prediction [14-16]. The aim of this study was to develop an IVUS-based supervised ML algorithm for classifying intermediate coronary lesions into those with an FFR ≤ 0.80 and those with an FFR > 0.80.

5

Materials and methods Study population. Between November 2009 and July 2015, we initially evaluated a consecutive series of 1657 stable and unstable angina patients who underwent invasive coronary angiography, pre-procedural IVUS and FFR to assess at least one intermediate lesion (defined as an angiographic DS of 40–80% on visual estimation) at Asan Medical Center, Seoul, Korea. In cases where both IVUS and FFR were measured for multiple lesions, the native coronary lesion with the lowest FFR value was chosen. A total of 329 patients were excluded (77 with tandem lesions, 95 with a stent within the target vessel, four with a side-branch evaluation, 49 with significant left main coronary artery stenosis [angiographic DS > 30%], 59 with incomplete IVUS evaluation, 16 with poor imaging quality, 12 with chronic total occlusion, eight with a scarred myocardium and regional wall motion abnormality, and nine with technical errors in the imaging files), leaving a final cohort of 1328 patients with 1328 non-left main coronary lesions for enrollment in this retrospective analysis. The patients were randomly assigned into training and test sets in a 4:1 ratio. Thus, 1063 patients were used for model training (training samples), and a non-overlapping group of 265 patients were used for evaluating the diagnostic performance of the model (test samples, Table 1). The protocol for the retrospective data analysis was approved by the institutional review board of the Asan Medical Center, and a waiver for informed consent was granted. Acquisition of angiography. Coronary angiography was performed with 5–7 F catheters through a radial or femoral access after the administration of 250µg of intracoronary nitroglycerine. FFR measurement. FFR is defined as the ratio of maximal coronary blood flow in a diseased artery to the maximal coronary blood flow in the same artery without stenosis

6

[1-3]. “Equalizing” was performed with a guidewire sensor positioned at the tip of the guiding catheter. A 0.014 inch FFR pressure guidewire (Radi, St. Jude Medical, Uppsala, Sweden) was then advanced distally to the stenosis. The FFR was measured at the maximum hyperemia induced by an intravenous infusion of adenosine. This was administered through a central vein at 140 µg/kg/min increasing to 200 µg/kg/min, to enhance the detection of hemodynamically relevant stenoses. Hyperemic pressure pullback recordings were performed, and FFR was then obtained as the ratio of the distal coronary artery pressure to the normal perfusion pressure (≈aortic pressure) during maximal hyperemia [1-3]. An FFR of 0.80 indicates that the stenotic coronary artery supplies 80% of the normal maximal flow. A stenosis was considered functionally significant when the FFR was 0.80 or less [3,4]. Acquisition of IVUS. After intracoronary administration of 0.2 mg of nitroglycerin, grayscale IVUS imaging was performed using motorized transducer pullback (0.5 mm/s) and a commercial scanner (Boston Scientific/SCIMED, Minneapolis, MN) consisting of a rotating 40 MHz transducer within a 3.2 F imaging sheath. IVUS segmentation. Lumen segmentation was performed using the interface between the lumen and the leading edge of the intima. A discrete interface at the border between the media and the adventitia corresponded almost to the location of the external elastic membrane (EEM). The initial segmentation mask was generated automatically by applying the FCN-all-at-once-VGG16 model to the IVUS images in the independent dataset (Method in the Supplementary appendix, Supplementary Figure 1). These initial segmentation results were reviewed by experts, and wrong segmentations were manually corrected. Cross-sectional images were segmented into three compartments: (i) adventitia including the pixels outside the EEM (coded as “0”), (ii) the lumen

7

including the pixels within the lumen border (coded as “1”), and (iii) plaque including the pixels between the lumen border and EEM (coded as “2”). To calibrate the pixel dimensions, grid lines were automatically found in the IVUS images and the pixel spacing was calculated. Plaque burden (PB) was calculated as (EEM area – lumen area) divided by EEM area × 100 (%), [17]. Every 12th image (inter-slice distance of 0.2 mm) was selected and used for the extraction of computed IVUS features. The inference time of the algorithm-based IVUS segmentation (vs. manual segmentation by experts) was shown in the Supplementary appendix. Computed segments. The IVUS-MLA was defined by selecting the frame with the smallest lumen area and a PB > 40%. A lesion, which necessarily contained the MLA site, was defined as a segment with a PB > 40% where the number of consecutive frames with a PB < 40% was less than 25 (< 5 mm segment). A region of interest (ROI) was defined as the segment from the ostium to a point 10 mm distal to the lesion. The proximal reference was the segment between the beginning of the ROI and the proximal edge of the lesion, and the distal reference was the segment between the distal edge of the lesion and the end of the ROI. The proximal and distal 5 mm references were within the 5 mm segments proximal or distal to the lesion. The worst segment was defined as the 4 mm segment 2 mm proximal to and 2 mm distal to the MLA site. Computational feature extraction. The 99 computed IVUS features are defined and summarized in Table 2 and Supplementary Table 1. With the addition of the six clinical features of age, gender, body surface area, involvement of proximal left anterior descending artery (LAD), vessel type, and involved segment (classified as proximal LAD, mid LAD, distal LAD, proximal to distal right coronary artery, proximal left circumflex artery, and distal left circumflex artery), a total of 105 features were used for

8

the ML. Machine learning. Six algorithms were evaluated as binary classifiers for separating lesions with an FFR ≤ 0.80 from those with an FFR > 0.80: binary class L2 penalized logistic regression, an artificial neural network, random forest, AdaBoost, CatBoost, and a support vector machine (Supplementary appendix). Each model was trained by a 5fold cross-validation scheme (Supplementary Figure 2) with data normalization. The hyperparameters were selected by using GridSearchCV based on the maximal AUC (Supplementary Table 2). Receiver operating characteristic curves for the relative performances considering the whole range of possible probability thresholds (from 0 to 1) would have an area in the range from 0.5 for classifiers without any prediction capability to 1 for algorithms performing perfect classification. The model performances were assessed by using the thresholds that were based on the maximal accuracies, the precision (positive predictive value) and recall (sensitivity) trade-off, and the sensitivity and specificity trade off. To measure the quality of binary classification, Matthews correlation coefficients (MCC) were calculated. For a non-biased assessment of the performance, the classifiers that had been previously built on the training samples were applied to the non-overlapping 265 test samples. In the 1063 training samples, train-validation random data split (with a 4:1 ratio) was repetitively performed by bootstrapping. The algorithms were independently trained on the 200 train-validation data splits and evaluated in the test set. The average performances shown as the mean [95% confidence intervals] of the 200 bootstrap replicates. Feature importance was evaluated by each ML algorithm (Supplementary appendix). The overall flow of the development of the supervised ML model is shown

9

in the Figure 1. Statistical analysis. The statistical analyses used for evaluating patient and lesion characteristics at baseline were performed using SPSS (version 10.0, SPSS Inc., Chicago, IL, USA). All values are expressed as means ± 1 standard deviation (continuous variables) or as counts and percentages (categorical variables). Continuous variables were compared using unpaired t-tests; and categorical variables were compared using χ2 statistics. A p value <0.05 was considered statistically significant. receiver operating characteristic curves analysis was performed using MedCalc Software (Mariakerke, Belgium) to assess the area under the curve and the best threshold for each angiographic feature to predict an FFR≤0.80 with maximal accuracy.

10

Results Clinical and lesion characteristics. The baseline characteristics of the study cohort are summarized in Table 1. The target vessels were the LAD in 891 (67.1%) patients, the left circumflex artery in 100 (7.5%) patients, and the right coronary artery in 337 (25.4%) patients. There were no significant differences in baseline characteristics between the training and test sets. The frequency of an FFR ≤ 0.80 was 370 (34.8%) in the training set and 80 (30.2%) in the test set (p = 0.155). Clinical and computed angiographic features. In the training set, an FFR ≤ 0.80 was more frequent in men than in women (38.8% vs. 24.0%, p < 0.001). Lesions with an FFR ≤ 0.80 (vs. FFR > 0.80) were associated with a younger patient age (60.2 ± 9.8 vs. 63.4 ± 9.4 years, p < 0.001) and a larger body surface area (1.76 ± 0.16 vs. 1.71 ± 0.16 m2, p < 0.001). The involved segment was the proximal LAD in 39.5% of lesions with an FFR ≤ 0.80 and 22.9% of lesions with an FFR > 0.80 (p < 0.001). The frequencies of FFR ≤ 0.80 were 44.4% in the LAD, 14.6% in the right coronary artery, and 15.8% in the left circumflex artery. In the training set, computed IVUS features were compared in lesions with FFR≤0.80 vs. FFR>0.80 (Supplementary Table 3). The receiver operating characteristic curves-based diagnostic performances of each IVUS feature were also calculated. When the criterion of an IVUS-MLA < 2.8 mm2 was applied to the test set, the sensitivity, specificity, and overall accuracy for predicting an FFR ≤ 0.80 were 75.0%, 61.6%, and 65.6%, respectively. Prediction of an FFR≤0.80 by ML. Ranked in order of importance, the top 20 features for predicting lesions with an FFR ≤ 0.80 are summarized in Table 3. Supplementary Table 4 summarizes the features, which were frequently listed on the top 20 features of

11

each algorithm. For these classifications, all clinical and computed IVUS features in the training data were used with 5-fold cross-validation (Table 4). In the test set, the overall diagnostic accuracies for predicting an FFR ≤ 0.80 were 82% with L2 penalized logistic regression, 80% with artificial neural network, 83% with random forest, 83% with AdaBoost, 81% with CatBoost, and 81% with support vector machine (AUCs: 0.84–0.87, Table 4). Figure 2 shows the results from the ROC analysis using the different ML models. When the 28 lesions with borderline FFR values (0.75–0.80) were excluded, the overall accuracies for the test set were 86% with L2 penalized logistic regression, 85% with an artificial neural network, 87% with random forest, 87% with AdaBoost, 85% with CatBoost, and 85% with support vector machine (Supplementary Table 5). In addition, the model performances based on the precision and recall trade-off (Supplementary Table 6) and the sensitivity and specificity trade-off (Supplementary Table 7) are shown. Supplementary Table 8 summarizes the average performances with 95% confidence intervals of the 200 bootstrap replicates. All algorithms consistently showed the averaged accuracies ≥ 80% for predicting FFR ≤0.80 (AUCs: 0.84 – 0.87).

12

Discussion It is important to identify high-risk populations that may benefit clinically from an approach incorporating ischemia-guided revascularization [18,19]. As a standard index for detecting ischemia-producing lesions, the routine estimation of FFR is recommended as part of the treatment decision making process for intermediate coronary stenosis [1-4]. The Fractional Flow Reserve versus Angiography for Multivessel Evaluation (FAME) trial demonstrated that FFR-guided (vs. angiographyguided) PCI in multivessel disease significantly reduces the rates of 1 year major adverse cardiac events [3]. In the FAME 2 trial evaluating lesions with an FFR ≤ 0.80, medical therapy alone (vs. FFR-guided PCI) showed a remarkably higher rate of clinical events [4]. Nonetheless, concerns over the prolonged procedural time, expense, and risk of complications limit the widespread use of FFR. The International Survey on Interventional Strategy suggested that more than 70% of operators’ decisions were based solely on angiographic appearance, which indicates a worrisome discrepancy between current guidelines and daily practice [5]. IVUS is a useful tool for planning PCI, as it provides anatomical information on the geometry of the lumen and vessel, stenosis severity, lesion length, and the extent of calcification. With validated criteria for stent underexpansion and edge problems, IVUS-guided PCI optimization improved clinical outcomes by reducing the rate of stent thrombosis and restenosis [7-9]. Although there have been attempts to identify IVUS criteria corresponding with functionally significant stenoses, and to integrate both morphological and physiological data [10-12], IVUS-MLA thresholds of 2.0–4.0 mm2 poorly predicted an FFR ≤ 0.80, with an overall diagnostic accuracy of 60–70%. In particular, the low positive predictive value may lead to a high rate of unnecessary PCI.

13

Even approaches using subgroup-specific IVUS-MLA criteria (based on vessel size and lesion location) and adjustments for body surface area or overall left ventricular mass failed to improve the accuracy of IVUS parameters for predicting an FFR ≤ 0.80 [10,11]. One of the reasons for the visual-functional mismatch is that myocardial ischemia is determined by many attributes, including the variable size of the supplied myocardium, the degree of stenosis, local geometry, and other clinical characteristics [10]. In addition, a single measurement of lumen area alone poorly reflects the complex geometry of the entire vessel. Even with a lot of clinical and morphological information, the development of prediction models using traditional statistical methods is limited by non-linearities between factors and outcomes, interactions among variables, and too many predictors [13]. There have been a number of approaches using image-based mathematical models to assess hemodynamic significance [20-23]. A virtual functional assessment index and quantitative flow ratio derived from 3D-quantitative coronary angiographic models of computational fluid dynamics achieved overall diagnostic accuracies of 80– 86%. As part of an ongoing effort to develop better methods for functional assessment, we here applied ML techniques that have emerged as highly effective computer algorithms for the identification of patterns in large datasets containing a multitude of variables, thus facilitating the building of models for data-driven prediction [14-16]. Using the 99 computed IVUS features and six clinical variables, all algorithms showed the overall accuracies greater than 80% (AUCs 0.84 – 0.87) in predicting an FFR ≤ 0.80 in non-left main coronary lesions. A maximal diagnostic accuracy of 83% was achieved by random forest and AdaBoost. Consistently, in the 200 bootstrap replicates, all algorithms showed the averaged accuracy ≥ 80% (averaged AUC: 0.84 – 0.87). The

14

frequency of misclassification was remarkably high in the gray zone of FFR values (between 0.75 and 0.80), which would require a physician’s clinical decision. With the exclusion of lesions with an FFR of 0.75–0.80, the overall accuracy was further improved to 85%–87%. By efficiently integrating both morphological and physiological information, this approach extends the role of IVUS in decision making for the management of intermediate coronary stenosis. Moreover, the data-driven ML model would continue to learn from new data, further improving its prediction accuracy. Our current data indicate how the IVUS features of a coronary stenosis impact the FFR value. According to the various algorithms, the high-ranking features mainly involved, with a reduced FFR < 0.80 were lesion length, IVUS-MLA, PB, the mean lumen, plaque, and EEM areas within the lesion, area stenosis, lumen and plaque eccentricity, the variance in lumen or plaque area, the stenosis degree and PB in both proximal and distal reference segments, longitudinal eccentricity, the distance between two centers of lumen and EEM, and the length of lumen area < 4.0 mm2, < 3.0 mm2 and < 2.5 mm2. In addition, the high-ranking features also included the clinical variables of age, sex, body surface area, and involved vessel and segment, which were related to the subtended myocardial territories. Although the rankings of the features were specific to the fitted model, the approach suggests that the best performing features warrant consideration in future models. Limitations. As we excluded significant left main disease, side branches and diffuse and tandem lesions, the ML models cannot be generally applied to all lesions. In addition, the binary classifiers cannot provide a numerical hemodynamic index for incremental risk stratification. Because manual correction of the whole frames takes time, only every 12th image (inter-slice distance of 0.2 mm) was selected and used for

15

the extraction of the computed IVUS features. Although the performances of the independently trained models were tested on the 200 train-test random splits by bootstrap, the possibility of over-fitting cannot be completely excluded. The performance and clinical impact of this approach should be further validated by an external study with a large cohort. Finally, an image-based deep learning strategy using large datasets is warranted to achieve optimal diagnostic performance. Conclusion. The IVUS-based ML models showed good diagnostic performance for identifying ischemia-producing lesions, and may reduce the future need for pressure wires. This approach potentially promotes the utilization of physiologically guided decision making.

16

Conflict of interest The authors declared that they do not have anything to disclose regarding conflict of interest with respect to this manuscript.

17

Financial support This study was supported by grants from the Korea Healthcare Technology R&D Project, Ministry for Health & Welfare Affairs, Republic of Korea (HI17C1080); and the Ministry of Science and ICT (NRF-2017R1A2B4005886).

Author contributions: JG Lee and J Ko: data analysis and development of models, review and edit of the manuscript with revision. H Hae: review and edit of the manuscript with revision. SJ Kang: Conception and design of study, Funding, data interpretation, and writing the manuscript with revision. DY Kang, P Lee, JM Ahn, DW Park, SW Lee, YH Kim, CW Lee, SW Park, and SJ Park: data acquisition, review and edit of the manuscript.

18

References 1. Pijls NH, Van Gelder B, Van der Voort P, Peels K, Bracke FA, Bonnier HJ, el Gamal MI. Fractional flow reserve. A useful index to evaluate the influence of an epicardial coronary stenosis on myocardial blood flow. Circulation 92 (1995) 3183–3193. 2 Pijls NH, De Bruyne B, Peels K, Van Der Voort PH, Bonnier HJ, Bartunek J Koolen JJ, Koolen JJ. Measurement of fractional flow reserve to assess the functional severity of coronary-artery stenoses. N Engl J Med 334 (1996) 1703–1708. 3. Tonino PA, De Bruyne B, Pijls NH, Siebert U, Ikeno F, van' t Veer M, Klauss V, Manoharan G, Engstrøm T, Oldroyd KG, Ver Lee PN, MacCarthy PA, Fearon WF; FAME Study Investigators. Fractional Flow Reserve versus Angiography for Guiding Percutaneous Coronary Intervention FAME. N Engl J Med 360 (2009) 213-224. 4. De Bruyne B, Fearon WF, Pijls NH, Barbato E, Tonino P, Piroth Z, Jagic N, MobiusWinckler S, Rioufol G, Witt N, Kala P, MacCarthy P, Engström T, Oldroyd K, Mavromatis K, Manoharan G, Verlee P, Frobert O, Curzen N, Johnson JB, Limacher A, Nüesch E, Jüni P; FAME 2 Trial Investigators. Fractional flow reserve-guided PCI for stable coronary artery disease. N Engl J Med 371 (2014) 1208-1217. 5. Toth GG, Toth B, Johnson NP, De Vroey F, Di Serafino L, Pyxaras S, Rusinaru D, Di Gioia G, Pellicano M, Barbato E, Van Mieghem C, Heyndrickx GR, De Bruyne B, Wijns W. Revascularization decisions in patients with stable angina and intermediate lesions: results of the international survey on interventional strategy. Circ Cardiovasc Interv. 7 (2014) 751–759. 6. Dattilo PB, Prasad A, Honeycutt E, Wang TY, Messenger JC. Contemporary patterns of fractional flow reserve and intravascular ultrasound use among patients undergoing percutaneous coronary intervention in the United States: insights from the National

19

Cardiovascular Data Registry. J Am Coll Cardiol. 60 (2012) 2337–2339. 7. Witzenbichler B, Maehara A, Weisz G, Neumann FJ, Rinaldi MJ, Metzger DC, Henry TD, Cox DA, Duffy PL, Brodie BR, Stuckey TD, Mazzaferri EL Jr, Xu K, Parise H, Mehran R, Mintz GS, Stone GW. Relationship between intravascular ultrasound guidance and clinical outcomes after drug-eluting stents: the assessment of dual antiplatelet therapy with drug-eluting stents (ADAPT-DES) study. Circulation. 129 (2014) 463-470 8. Hong SJ, Kim BK, Shin DH, Nam CM, Kim JS, Ko YG, Choi D, Kang TS, Kang WC, Her AY, Kim YH, Hur SH, Hong BK, Kwon H, Jang Y, Hong MK; IVUS-XPL Investigators. Effect of Intravascular Ultrasound-Guided vs Angiography-Guided Everolimus-Eluting Stent Implantation: The IVUS-XPL Randomized Clinical Trial. JAMA. 314 (2015) 2155-2163. 9. Zhang YJ, Pang S, Chen XY, Bourantas CV, Pan DR, Dong SJ, Wu W, Ren XM, Zhu H, Shi SY, Iqbal J, Gogas BD, Xu B, Chen SL. Comparison of intravascular ultrasound guided versus angiography guided drug eluting stent implantation: a systematic review and meta-analysis. BMC Cardiovasc Disord. 15 (2015) 153 10. Park SJ, Kang SJ, Ahn JM, Shim EB, Kim YT, Yun SC, Song H, Lee JY, Kim WJ, Park DW, Lee SW, Kim YH, Lee CW, Mintz GS, Park SW. Visual-functional mismatch between coronary angiography and fractional flow reserve. JACC Cardiovasc Interv 5 (2012) 1029-1036. 11. Kang SJ, Ahn JM, Song H, Kim WJ, Lee JY, Park DW, Yun SC, Lee SW, Kim YH, Lee CW, Park SW, Park SJ. Usefulness of minimal luminal coronary area determined by intravascular ultrasound to predict functional significance in stable and unstable angina pectoris. Am J Cardiol 109 (2012) 947-953.

20

12. Nascimento BR, de Sousa MR, Koo BK, Samady H, Bezerra HG, Ribeiro AL, Costa MA. Diagnostic accuracy of intravascular ultrasound-derived minimal lumen area compared with fractional flow reserve--meta-analysis: pooled accuracy of IVUS luminal area versus FFR. Catheter Cardiovasc Interv 84 (2014) 377-385. 13. Goldstein BA, Navar AM, Carter RE.Moving beyond regression techniques in cardiovascular risk prediction: applying machine learning to address analytic challenges. Eur Heart J. 23 (2017) 1805-1814. 14. Krittanawong C, Zhang H, Wang Z, Aydar M, Kitai T. Artificial Intelligence in Precision Cardiovascular Medicine. J Am Coll Cardiol 69 (2017) 2657-2664. 15. Henglin M, Stein G, Hushcha PV, Snoek J, Wiltschko AB, Cheng S. Machine Learning Approaches in Cardiovascular Imaging. Circ Cardiovasc Imaging. 10 (2017) e005614. 16. Darcy AM, Louie AK, Roberts LW. Machine learning and the profession of medicine. JAMA 315 (2016) 551–552. 17. Mintz GS, Nissen SE, Anderson WD, Bailey SR, Erbel R, Fitzgerald PJ, Pinto FJ, Rosenfield K, Siegel RJ, Tuzcu EM, Yock PG. American College of Cardiology Clinical Expert Consensus Document on Standards for Acquisition, Measurement and Reporting ofIntravascularUltrasound Studies (IVUS). A report of the American College of Cardiology Task Force on Clinical Expert Consensus Documents. J Am Coll Cardiol 37 (2001) 1478-1492. 18. Gada H, Kirtane AJ, Kereiakes DJ, Bangalore S, Moses JW, Généreux P, Mehran R, Dangas GD, Leon MB, Stone GW. Meta-analysis of trials on mortality after percutaneous coronary intervention compared with medical therapy in patients with stable coronary heart disease and objective evidence of myocardial ischemia. Am J

21

Cardiol 115 (2015) 1194-1199 19. Hachamovitch R, Hayes SW, Friedman JD, Cohen I, Berman DS. Comparison of the short-term survival benefit associated with revascularization compared with medical therapy in patients with no prior coronary artery disease undergoing stress myocardial perfusion single photon emission computed tomography. Circulation 107 (2003) 29002907. 20. Westra J, Tu S, Winther S, Nissen L, Vestergaard MB, Andersen BK, Holck EN, Fox Maule C, Johansen JK, Andreasen LN, Simonsen JK, Zhang Y, Kristensen SD, Maeng M, Kaltoft A, Terkelsen CJ, Krusell LR, Jakobsen L, Reiber JHC, Lassen JF, Bøttcher M, Bøtker HE, Christiansen EH, Holm NR. Evaluation of Coronary Artery Stenosis by Quantitative Flow Ratio During Invasive Coronary Angiography: The WIFI II Study (Wire-Free Functional Imaging II). Circ Cardiovasc Imaging. 11 (2018) e007107. 21. Papafaklis MI1, Muramatsu T, Ishibashi Y, Lakkas LS, Nakatani S, Bourantas CV, Ligthart J, Onuma Y, Echavarria-Pinto M, Tsirka G, Kotsia A, Nikas DN, Mogabgab O, van Geuns RJ, Naka KK, Fotiadis DI, Brilakis ES, Garcia-Garcia HM, Escaned J, Zijlstra F, Michalis LK, Serruys PW. Fast virtual functional assessment of intermediate coronary lesions using routine angiographic data and blood flow simulation in humans: comparison with pressure wire - fractional flow reserve. EuroIntervention 10 (2014) 574-583. 22. Tu S, Westra J, Yang J, von Birgelen C, Ferrara A, Pellicano M, Nef H, Tebaldi M, Murasato Y, Lansky A, Barbato E, van der Heijden LC, Reiber JH, Holm NR, Wijns W; FAVOR Pilot Trial Study Group. Diagnostic Accuracy of Fast Computational Approaches to Derive Fractional Flow Reserve From Diagnostic Coronary

22

Angiography: The International Multicenter FAVOR Pilot Study. JACC Cardiovasc Interv 9 (2016) 2024-2035. 23. Xu B, Tu S, Qiao S, Qu X, Chen Y, Yang J, Guo L, Sun Z, Li Z, Tian F, Fang W, Chen J, Li W, Guan C, Holm NR, Wijns W, Hu S. Diagnostic Accuracy of Angiography-Based Quantitative Flow Ratio Measurements for Online Assessment of Coronary Stenosis. J Am Coll Cardiol 70 (2017) 3077-3087.

23

Figure legends

Figure 1. Workflow for developing the machine learning models. ANN = artificial neural network; FFR = fractional flow reserve; SVM = support vector machine.

Figure 2. Receiver operating characteristic curves of the machine learning models for classifying the test set according to lesions with a fractional flow reserve ≤ 0.80 vs. > 0.80. ANN = artificial neural network; SVM = support vector machine.

24

Table 1. Baseline characteristics Training set

Test set*

1063/1063

265/265

62.3±9.7

62.6±9.2

Men

776 (73.0%)

182 (68.7%)

Diabetes mellitus

309 (29.1%)

72 (27.5%)

Hypertension

626 (58.9%)

166 (62.6%)

Current smoker

547 (51.5%)

120 (45.3%)

Hyperlipidemia

624 (58.7%)

166 (62.6%)

Stable (vs. unstable) angina

833 (78.4%)

200 (75.5%)

Body surface area, m2

1.72±0.16

1.71±0.17

FFR at maximal hyperemia

0.83±0.09

0.83±0.10

Proximal LAD

305 (28.7%)

72 (27.2%)

Mid LAD

396 (37.3%)

96 (36.2%)

18 (1.7%)

4 (1.5%)

268 (25.2%)

69 (26.0%)

Proximal LCX

42 (4.0%)

11 (4.2%)

Distal LCX

34 (3.2%)

13 (4.9%)

Patient/lesion number Age, years

Involved segment

Distal LAD Proximal to distal RCA

LAD= left anterior descending artery lesion, RCA=right coronary artery, LCX= left circumflex artery *All p values <0.05 vs. training set

1

Table 2. Definitions of the computed IVUS features

Feature definitions

Within the ROI

Length of the lesion, mm

Within

the lesion the worst

Within

Within

Within

Within

proximal

distal

proximal ref

distal ref

reference

reference

5 mm

5 mm

At the MLA

No 1

Length of PB >70%, mm Distance from the ostium to proximal edge of

Within

No 2 No 3

the lesion, mm MLA, mm2

No 4

Distance from the ostium to the MLA, mm2

No 5

EEM area at the MLA site, mm2

No 6

PB at the MLA site¶, %

No 7

Maximal PB, %

No 8

Total count of frames with PB>40%

No 9

Total count of frames with PB>70%

No 10

Total count of frames with lumen<4.0 mm2

No 11

No 20

No 29

1

No 38

No 45

No 52

No 59

Total count of frames with lumen<3.0 mm2

No 12

No 21

No 30

Total count of frames with lumen<2.5 mm2

No 13

No 22

No 31

Sum of the plaque areas, mm2

No 14

No 23

No 32

No 39

No 46

No 53

No 60

Sum of the EEM areas, mm2

No 15

No 24

No 33

No 40

No 47

No 54

No 61

Segmental plaque burden, %§

No 16

No 25

No 34

No 41

No 48

No 55

No 62

Averaged lumen area, mm2

No 17

No 26

No 35

No 42

No 49

No 56

No 63

Averaged plaque area, mm2

No 18

No 27

No 36

No 43

No 50

No 57

No 64

Averaged EEM area, mm2

No 19

No 28

No 37

No 44

No 51

No 58

No 65

Lumen eccentricity*

No 66†

No 67†

No 68

Plaque eccentricity#

No 69†

No 70†

No 71

Distance between 2 centers of lumen – EEM

No 72†

No 73†

No 76

Maximal distance between two centers of

No 74

No 75

Variance of lumen area, mm2

No 78

No 77

Variance of plaque area, mm2

No 80

No 79

lumen and EEM, mm

2

Calculated features No 81

Averaged reference lumen = (No 56 + No 63) / 2

No 82

Averaged reference EEM = (No 58 + No 65) / 2

No 83

Area stenosis 1 = (No 81 – MLA) / No 81 × 100(%)

No 84

Area stenosis 2 = (No 56 – MLA) / No 56 × 100(%)

No 85

Area stenosis 3 = (No 63 – MLA) / No 63 × 100(%)

No 86

Area stenosis 4 = (No 82 – MLA) / No 82 × 100(%)

No 87

Area stenosis 5 = (No 58 – MLA) / No 58 × 100(%)

No 88

Area stenosis 6 = (No 65 – MLA) / No 65 × 100(%)

No 89

Area stenosis 7 = (No 81 – No 35) / No 81 × 100(%)

No 90

Area stenosis 8 = (No 56 – No 35) / No 56 × 100(%)

No 91

Area stenosis 9 = (No 63 – No 35) / No 63 × 100(%)

No 92

Area stenosis 10 = (No 82 – No 35) / No 82 × 100(%)

No 93

Area stenosis 11 = (No 58 – No 35) / No 58 × 100(%)

No 94

Area stenosis 12 = (No 65 – No 35) / No 65 × 100(%)

No 95

Remodeling index 1 = No 6 / No 82

No 96

Remodeling index 2 = No 6 / No 58

No 97

Remodeling index 3 = No 37 / No 82

No 98

Remodeling index 4 = No 37 / No 58

No 99

Longitudinal eccentricity = [distance from the proximal edge to the MLA / No 1]

ROI= region of interest, PB= plaque burden, MLA= minimal lumen area, EEM= external elastic membrane ¶

[EEM – MLA] / EEM × 100 (%), § [Sum of the plaque areas / Sum of the EEM areas] within the defined segment x 100%

* †

,

#

Averaged value within the defined segment

3

Table 3. Top 20 features for predicting a fractional flow reserve ≤0.80 by each algorithm. L2 penalized logistic

Artificial neural

regression

networks

1st

Sex

Vessel

No 21

Age

Vessel

No 78

2nd

BSA

Segment

No 22

Segment

Age

No 99

3rd

No 62

Sex

No 4

Vessel

Segment

Sex

4th

No 78

BSA

No 8

No 8

No 8

No 19

5th

No 87

Proximal LAD

Age

BSA

BSA

No 49

6th

No 49

No 71

BSA

No 81

No 7

No 16

7th

No 77

Age

No 7

No 4

No 4

No 79

8th

No 85

No 68

Segment

No 86

No 34

No 17

9th

No 52

No 66

No 34

No 9

Sex

No 20

10th

No 89

No 7

No 20

No 7

No 9

No 59

11th

No 7

No 59

No 16

No 71

No 21

No 92

12th

No 8

No 72

No 71

Sex

No 13

No 38

Rank

Support vector RF

AdaBoost

CatBoost machine

1

13th

No 18

No 77

No 9

No 41

No 20

No 13

14th

No 27

No 8

No 12

No 65

No 71

No 21

15th

No 31

No 78

No 68

No 16

No 68

No 75

16th

No 26

No 80

No 99

No 29

No 1

No 48

17th

No 41

No 45

No 66

No 21

No 17

No 14

18th

No 22

No 51

No 46

No 46

No 81

No 45

19th

No 11

No 49

No 97

No 74

No 65

No 30

20th

No 99

No 69

No 77

No 63

No 99

No 38

*Coefficient (by Elastic Net), #Feature importance

2

Table 4. Model performances for predicting fractional flow reserve ≤0.80 Threshold of predictive score#

AUROC

Sensitivity Specificity

PPV

NPV

Overall accuracy

MCC

5-fold cross validation in the 1063 training samples L2–logistic regression

0.51

0.83

0.65

0.83

0.67

0.82

0.77

0.49

Artificial neural networks

0.50

0.81

0.62

0.84

0.68

0.81

0.77

0.47

Random forest

0.49

0.82

0.58

0.84

0.67

0.79

0.75

0.44

AdaBoost

0.50

0.83

0.54

0.87

0.70

0.78

0.76

0.45

CatBoost

0.46

0.84

0.64

0.83

0.67

0.81

0.77

0.48

Support vector machine

0.50

0.83

0.6

0.86

0.69

0.8

0.77

0.48

L2–logistic regression

0.51

0.87

0.70

0.86

0.70

0.87

0.82

0.56

Artificial neural networks

0.50

0.84

0.65

0.87

0.68

0.85

0.80

0.53

Random forest

0.49

0.85

0.71

0.88

0.72

0.88

0.83

0.60

AdaBoost

0.50

0.86

0.62

0.92

0.77

0.85

0.83

0.58

In the 265 test samples

1

CatBoost

0.46

0.86

0.66

0.87

0.70

0.86

0.81

0.54

Support vector machine

0.50

0.87

0.66

0.88

0.70

0.86

0.81

0.55

AUROC= area under the receiver operating characteristic curve, PPV= positive predictive value, NPV= negative predictive value, MCC= Matthews correlation coefficients #

Thresholds to achieve the maximal accuracy on the ROC

2

Highlights – Machine learning provides the ability to automatically learn without being explicitly programmed, which enables improving diagnostic accuracies. – The models using computed IVUS features predicts the intermediate stenosis with an FFR ≤0.80 with an overall accuracy of 80%. – The data-driven approach may help clinicians in identifying ischemia-producing coronary lesions.

Conflict of interest The authors declared that they do not have anything to disclose regarding conflict of interest with respect to this manuscript.