1. Introduction
In the last decade, coronary computed tomography angiography (cCTA) has been primarily used as a non-invasive method to obtain morphological information about the status of atherosclerosis and plaque in coronary arteries [
1]. Studies including PROMISE [
2] and SCOT-HEART [
3] examined the diagnostic yield of an initial cCTA scan compared to standard non-invasive functional testing such as exercise electrocardiography, stress echocardiography, cardiac Magnetic Resonance Imaging (MRI) or nuclear stress testing in symptomatic patients suspected of suffering from obstructive coronary artery disease (CAD). The high negative predictive value of cCTA, specifically in patients with a lower pre-test probability for CAD, allows cardiologists to rule out obstructive CAD with high certainty and thus reduces the number of concomitant unnecessary invasive procedures [
4,
5]. One major drawback of cCTA for the detection of obstructive CAD and usage as gatekeeper for ICA is the low specificity. cCTA frequently overrates the severity of coronary stenosis in heavily calcified coronary arteries, resulting in false positive results that trigger a referral to an interventional cardiologist [
2,
3,
6,
7].
The current gold standards for the assessment of the hemodynamic significance of a coronary lesion are pressure-derived indices including the invasively measured fractional flow reserve (FFR) [
6] and, just recently, the instantaneous wave-free ratio (iwFR) [
8]. In patients with chronic coronary syndrome (CCS), FFR-guided percutaneous coronary intervention (PCI) is superior regarding morbidity and prognostic outcome compared to an optimal medical therapy (OMT) alone [
9,
10]. FFR measures the pressure proximal and distal to a lesion under maximal hyperemia induced by the vasodilator adenosine and calculates the ratio between these two values. In the last two decades, FFR guided revascularization has been established in clinical practice. Prominently, the DEFER study demonstrated that patients do not benefit from the treatment of coronary lesions that do not cause myocardial ischemia as determined by FFR compared to OMT [
11]. Furthermore, the FAME study reported a decrease in major adverse cardiac events (MACE) when treating intermediate coronary lesions guided by an FFR Index-value of ≤ 0.80 only [
12]. SWEDEHEART and DEFINE FLAIR finally demonstrated that revascularization guided by iwFR yielded comparable results to FFR [
13,
14]. Different from FFR, iwFR is obtained under resting conditions in a wave-free period during the diastole of the cardiac cycle. However, the results of the large multicenter ISCHEMIA trial, which were recently presented, indicate that stable patients with moderate ischemia did not profit from a routine invasive therapy [
15,
16]. Due to inconclusive outcome data, an extended long-term follow-up is expected. Although it is not yet published, the study shows even more the importance of a precise pre-invasive selection of patients [
15,
16].
CT-derived fractional flow reserve (CT-FFR; FFR computed from resting cCTA images) appears to have a high correlation with invasive FFR [
7]. CT-FFR is a promising novel approach that allows both anatomical and functional assessment of a coronary stenoses and has an improved specificity compared to a cCTA-only approach. The prospective multicenter PLATFORM trial already demonstrated a lower rate of patients showing no obstructive coronary lesions with ICA, when CT-FFR was obtained previously as a gatekeeper [
17,
18]. As an additional advantage, CT-FFR does not require additional radiation or pharmaceutical stress agents [
19]. A recent improvement in the development of CT-FFR is the introduction of an on-site prototype software based on a machine-learning algorithm (CT-FFR
ML). CT-FFR
ML provides a significant reduction in the required computation time compared to CT-FFR based on computational fluid dynamics [
20]. CT-FFR
ML has already been successfully validated against both invasive FFR and conventional off-site CT-FFR [
7,
20]. The overall goal of coronary diagnostics is the accurate characterization of lesion location, morphology and hemodynamic impact in light of patient characteristics and pre-test probability, with the ultimate goal to derive treatment options and risk stratification. Therefore, several scores have been developed condensing patients’ data. The CAD consortium basic pretest probability predicts the chance of having an obstructive CAD and was recommended by the ESC guidelines in 2013 [
6] and 2019 [
4]. The Agatston score [
21] quantifies the amount of deposited calcium to predict coronary events. A newly developed comprehensive cCTA score, which is based on the location and the severity of the lesion as well as the plaque composition may also be used to characterize CAD severity [
22].
The purpose of the present study was to examine the potential impact of the new machine-learning-based CT-FFRML prototype on clinical practice, particularly the diagnostic and therapeutic pathway of patients with suspected CAD. We were further interested in examining whether application of CT-FFRML could reduce the number of patients who undergo ICA following cCTA with the diagnosis of non-obstructive CAD.
4. Discussion
In the present study, we demonstrated the potential effect of an on-site CT-FFRML machine learning algorithm in terms of clinical practicality and its diagnostic performance to determine the hemodynamic significance of coronary lesions. With the prototype software used in this study, 93% of our enrolled patients with suspected CAD or suspected progression of CAD who initially received CTA, were adequately classified. Using CT-FFR analysis, 55% of the patients enrolled could have potentially omitted further invasive testing and associated radiation exposure. Only 3% of patients had a CT-FFRML measurement of ≥0.8, despite the necessity of coronary revascularization.
In earlier studies, less than 60% of patients with suspected CAD and pathological non-invasive functional test results demonstrate a significant coronary obstruction during ICA [
2,
29]. However, other results might be possible if another stress test was used initially. In a more current study (MR-INFORM), it was shown that decisions in the diagnostic pathway in patients with a higher risk of a cardiovascular event can be guided as safely by cardiac magnetic resonance imaging (CMR) as they are by invasive FFR. Regarding the ISCHEMIA trial as well as the MR-INFORM study, always in light of the specific patient collective, it is clear that patients should be selected precisely to be referred to ICA [
15,
16,
30]. This requires the integration of patient information by generating risk-stratification scores estimating the presence of obstructive CAD as well as non-invasive new imaging technics using self-improving machine learning software. With a higher sensitivity (95–99% [
4,
6]) compared to the current non-invasive functional testing methods and a high negative predictive value, cCTA helps reduce the number of patients without obstructive CAD during ICA [
2,
3]. The NXT-Trial study was able to reveal an improved specificity of cCTA when extended by CT-FFR [
7]. Evaluating its additive value in clinical routine, the PLATFORM compared any form of non-invasive testing, including cCTA without CT-FFR, to cCTA plus CT-FFR and revealed 61% less referrals to ICA in the cCTA/CT-FFR group, resulting in significantly lower rates of unnecessary, non-obstructive ICAs. The real-world utility and safety of CT-FFR was consequently evaluated in the multicentre ADVANCE registry study [
31]. Here, the disease management plan of two-thirds of patients in the cCTA plus CT-FFR group was changed compared to the patients in the core lab cCTA-only approach group. No death or myocardial infarction (MI) occurred in patients with a CT-FFR value of >0.80 within a 90 day follow-up.
The results of our retrospective study integrate into the outcomes of the leading international multi-centre ADVANCE registry and PLATFORM study, meeting the expectations of the “gatekeeper” function of CT-FFR. Similar to these studies, 48 (55%) out of 88 of our patients did not show any obstructive findings during ICA. By retrospectively adding the CT-FFRML diagnostics into the diagnostic pathway, we were able to show the additional value of combining morphological and functional information that could lead to a reduction in invasive procedures by 55%, as well as substantially lower incidence of finding no obstructive coronary disease during ICA.
Among a subgroup in the PLATFORM-study [
18] who had invasive testing planned, 24 (12.4%) of 193 in the CT-FFR arm did not have hemodynamically significant lesions when assessed by ICA, compared to 137 (73.3%) in the usual care arm, a difference of 60.9%. In the CT-FFR arm, 17 (8.8%) had to be excluded due to poor image quality or inadequate acquisition. After the sole anatomical evaluation of the cCTA data in our study, 48 (55%) of 88 would have received ICA without the necessity of revascularization treatment, versus three (8%) of 40 patients after the analysis of the cCTA data by CT-FFR
ML, a difference of 47%. PLATFORM compares the cCTA/CT-FFR group to the usual care group within the preselected “planned ICA” arm; therefore, all patients in the usual care group receive ICA and the effect of the integration of CT-FFR is even larger than in our study, since our hypothetical control group is already selected by cCTA. Similarly, we had to exclude 11 (8%) patients due to poor image quality or inadequate acquisition which might be a limitation of CT-FFR
ML or cCTA itself. Despite the retrospective study design, our study population was carefully selected. The number of patients with coronary anomalies (7, 5%) and subtotal or chronic total occlusions (13 and 10%) indicate the real-world profile of the patient collective. Other than in the original CT-FFR study NXT-Trial, patients with total occlusions were not included, with a predefined CT-FFR value of 0.50. This would have increased the number of true positives, and therefore improved the diagnostic performance of this method. Since CT-FFR analysis was virtually available for every patient in the ADVANCE study, how the downstream treatment of each patient would have been without CT-FFR could not be determined [
31]. Since our study has a retrospective design, we were able to gather the respective information required to simulate the hypothetical CT-FFR
ML pathway and compare it to the actual diagnostic pathway that was chosen by the physician.
The series of diagnostic tools used by us, including the CAD consortium basic score, the Agatston score, the comprehensive CTA score and the CT-FFR analysis, reflects how the integration and combination of patient data leads to a precise risk stratification guiding a corresponding sophisticated therapeutic strategy. The discriminatory value of the CAD consortium basic pretest probability score is not displayed in the results of our study (AUC = 0.578,
p = 0.3212) (
Figure 4). The clinical score was previously evaluated in view of a diameter stenosis ≥50% in cCTA, not regarding revascularizing therapy [
23]. Therefore, our population is biased, since it is a selection of patients who were referred to ICA after a positive cCTA. The means of the Agatston score showed a significant difference in the revascularized and not-revascularized group (
p = 0.0467). However, the validity of this score is restricted due to a large standard deviation, since the amount of calcium in the coronary tree is just sufficiently able to predict the necessity of revascularization therapy (AUC = 0.631,
p = 0.1448). Therefore, a high calcium score does not inevitably indicate myocardial ischemia; its prognostic accuracy is discussed, since it does not include non-calcified lesions and does not display if the amount of coronary calcium is accumulated in on spot or spread diffusely across the coronary tree [
21]. The comprehensive CTA score offered a slightly better diagnostic accuracy (AUC = 0.633,
p = 0.0158), which shows that including the typicity as well as the location of the lesion has an impact on the chosen therapy (
Figure 4). The higher the comprehensive CTA score, the higher the risk of needing revascularization. However, the discriminatory value of this score from previous trials [
22] could not be shown in this study due to the selection of our patient collective, patients who received both cCTA and ICA. Finally, the ischemia producing factor, the hemodynamic severity is quantified by the CT-FFR value and therefore provides a great discriminatory value (AUC = 0.958,
p ≤ 0.0001) (
Figure 4) to guide revascularization therapy.
Easy accessibility and the economical operation of the CT-FFR
ML-software are essential for its integration into clinical practice [
32,
33]. The off-site CT-FFR software by Heartflow Inc. (1400 Seaport Blvd, Redwood City, CA 94063) has attained approval by the Food and Drug Administration in 2015, thus indicating the growing importance and utility of CT-FFR [
7,
34]. However, the CT-FFR software used in these studies was based on computational fluid dynamics. Thus, CTA data has to be transferred to external core laboratories and has to be analysed off-site, which takes several hours and implies great costs. With this setup the assessment is not accessible instantly and does not help in the clinical decision process [
32]. In our study, we used a prototype software based on machine learning (cFFR, version 3.1, Siemens Healthineers, Forchheim, Germany) that allows an on-site computation on a regular workstation. Tesche et al. recently proved that the algorithm based on machine learning receives adequate results compared to the algorithm based on computational fluid dynamics, while significantly improving clinical practicability [
20]. Our mean calculation time was 24 ± 11 min. An earlier publication by Renker et al. in 2014 demonstrated a processing and calculation time of 37.5 ± 13.8 min using an on-site prototype CT-FFR software which was based on a computational fluid dynamic algorithm [
35]. The type of algorithm of the software, either CT-FFR based on machine learning (CT-FFR
ML or CT-FFR based on computational fluid dynamics (CT-FFR
CFD), as well as the type of algorithm for the reconstruction of cCTA-data, seems to have an impact on the efficiency as well: iterative reconstruction in CT-FFR
ML lowers the post-processing speed compared to the filtered-back projection reconstruction of CT-FFR
CFD [
36].
Using the data of the PLATFORM-study, Hlatky et al. analysed the costs and the Quality-of-Life when integrating CT-FFR
ML into the clinical practice. In the group of patients planned for ICA, the new diagnostic pathway lead to a 32% reduction (7.343 vs. 10.734 USD,
p < 0.0001) in healthcare expenditures [
37]. Standardized questionnaires as a part of this study revealed a significantly improved Quality-of-Life in the CT-FFR branch [
37].
Although CT-FFR
ML is a very promising new non-invasive method in the disease management plan of patients with suspected obstructive CAD, there are a few caveats to consider. A total of 13% in the NXT-Trial, 8.8% in the PLATFORM study, and 8% of the patients in our study selected for CT-FFR analysis, had to be excluded because of inadequate image quality [
7,
18]. Furthermore, a “cut-off” value that allows a safe decision if the stenosis is obstructive or irrelevant has just been derived from the invasive FFR-Index (0.75–0.80) and has never been prospectively examined [
38]. The result is a “grey-zone” of 0.7 to 0.9, which does not clearly indicate a normal or pathological CT-FFR
ML value [
39]. Founded in the retrospective profile of our study, we exposed three (3%) false negative cases, patients who received stents but revealed a CT-FFR
ML-value ≥ 0.80. Although single cases, two of the three false negatives revealed an iwFR value between 0.86 and 0.89 (grey-zone: 0.86–0.93 [
40]) and one of these had a very high profile of risk factors.
Our results furthermore support the importance of artificial intelligence whose rising capability and efficiency allows the development of medical software that instantly assists in clinical patient management. This therefore enables further “exploitation of the considerable richness of coronary CTA data” [
32], leading to a more time and also a cost-efficient process.