Machine learning from quantitative coronary CT angiography predicts FFR-defined ischemia and impaired myocardial blood flow

Abstract

Background:

A pathophysiological interplay exists between plaque morphology and coronary physiology. Machine learning (ML) is increasingly being applied to coronary computed tomography angiography (CCTA) for cardiovascular risk stratification. We sought to assess the performance of a ML score integrating CCTA-based quantitative plaque features for predicting vessel-specific ischemia by invasive fractional flow reserve (FFR) and impaired myocardial blood flow (MBF) by positron emission tomography (PET).

Methods:

This post-hoc analysis of the PACIFIC trial included 208 patients with suspected coronary artery disease who prospectively underwent CCTA, [¹⁵O]H₂O PET, and invasive FFR. Plaque quantification from CCTA was performed using semiautomated software. A ML algorithm trained on the prospective NXT trial (484 vessels) was used to develop a ML score for the prediction of ischemia (FFR≤0.80), which was then evaluated in 581 vessels from the PACIFIC trial. Thereafter, the ML score was applied for predicting impaired hyperemic MBF (≤2.30 mL/min/g) from corresponding PET scans. The performance of the ML score was compared with CCTA reads and non-invasive FFR derived from CCTA (FFR_CT).

Results:

139 (23.9%) vessels had FFR-defined ischemia and 195 (33.6%) vessels had impaired hyperemic MBF. For the prediction of FFR-defined ischemia, the ML score yielded an area under the receiver-operating characteristic curve (AUC) of 0.92, which was significantly higher than that of visual stenosis grade (0.84; p<0.001) and comparable with that of FFR_CT (0.93; p=0.34). Quantitative percent diameter stenosis and low-density noncalcified plaque volume had greatest ML feature importance for predicting FFR-defined ischemia. When applied for impaired MBF prediction, the ML score exhibited an AUC of 0.80; significantly higher than visual stenosis grade (AUC 0.74; p=0.02) and comparable with FFR_CT (AUC 0.77; p=0.16).

Conclusions:

An externally validated ML score integrating CCTA-based quantitative plaque features accurately predicts FFR-defined ischemia and impaired MBF by PET, performing superiorly to standard CCTA stenosis evaluation and comparably to FFR_CT.

INTRODUCTION

Invasive fractional flow reserve (FFR) has emerged as the clinical reference standard for determining lesion-specific ischemia and guiding revascularization decisions¹. Meanwhile, cardiac positron emission tomography (PET) is the established reference standard for noninvasive assessment of myocardial perfusion and quantification of myocardial blood flow (MBF)². Increasing evidence suggests that in addition to coronary stenosis severity, atherosclerotic plaque morphology and burden are important determinants of FFR and MBF^3–9. Coronary computed tomography angiography (CCTA), a noninvasive first-line test for the evaluation of luminal stenosis¹⁰, also enables atherosclerotic plaque characterization and quantification with high accuracy when compared to invasive intracoronary imaging¹¹. Recent studies have shown CCTA-derived qualitative adverse plaque characteristics such as positive remodeling and presence of low attenuation, along with quantitative measurements of noncalcified and low-density noncalcified plaque volume, to associate with ischemia by invasive FFR and abnormal MBF by PET^3–9. In prior work, we used data from the NXT (Analysis of Coronary Blood Flow using CT Angiography: Next Steps) trial¹² to develop a machine learning (ML) score integrating CCTA-based quantitative plaque features for the prediction of vessel-specific FFR-defined ischemia¹³. The present study aimed to refine the ML score using a state-of-the-art algorithm, then externally validate its performance for ischemia prediction in the PACIFIC (Prospective Comparison of Cardiac PET/CT, SPECT/CT Perfusion Imaging and CT Coronary Angiography with Invasive Coronary Angiography) trial¹⁴ using both FFR and PET as reference standards. We sought to compare the performance of the ML score with CCTA reads and non-invasive FFR derived from CCTA (FFR_CT), which represent current clinical practice.

METHODS

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Study Population

This was a post-hoc analysis of the PACIFIC trial (NCT01521468)¹⁴, a single-center study in which 208 consecutive patients with suspected stable coronary artery disease (CAD) prospectively underwent CCTA, invasive coronary angiography (ICA) with 3-vessel FFR, and [¹⁵O]H₂O PET within a 2-week interval. The trial was approved by the Medical Ethics Committee at VU University Medical Center and all participants provided written informed consent.

Coronary Computed Tomography Angiography

All patients underwent CCTA using a 256-slice CT scanner (Philips Brilliance iCT, Philips Healthcare, Best, the Netherlands) as previously described¹⁴. The CCTA protocol is detailed in the Supplemental Methods. CCTA datasets were transmitted to an independent and blinded core laboratory (St. Paul’s Hospital, Vancouver, British Columbia, Canada) for the assessment of diameter stenosis severity. All coronary segments ≥2 mm in diameter¹⁵ were visually graded and classified as: 0, 1–24, 25–49, 50–69, or 70–100%. Coronary lesions were also assessed for qualitative adverse plaque characteristics including positive remodeling (defined as remodeling index ≥1.1), low-attenuation plaque (containing any voxel <30 Hounsfield units [HU]), spotty calcification (<90° of vessel circumference and <3 mm in length)¹⁶, and the napkin-ring sign¹⁷.

Quantitative Plaque Analysis

Standardized plaque quantification^11,18 was performed by an independent and blinded core laboratory (Cedars-Sinai Medical Center, Los Angeles, CA, USA) using semiautomated software (Autoplaque v2.5) (Supplemental Methods). Per-vessel plaque volumes (mm³) were calculated for the following components: total plaque, calcified plaque (CP), noncalcified plaque (NCP), and low-density NCP (LDNCP; <30 HU. The respective plaque burdens (%) were calculated as: plaque volume / analyzed vessel volume × 100¹⁸. Plaque composition (%) by CP, NCP, and LDNCP components was calculated as: plaque component volume / total plaque volume × 100¹¹. Excellent intraobserver and interobserver agreement for quantitative plaque measurements using this software have been previously reported^5,19.

For each vessel, the maximally stenotic cross-section was automatically measured for minimal luminal area, minimal luminal diameter, percent diameter stenosis, and percent area stenosis. Additionally, vessel remodeling index, and contrast density difference were calculated by the software (Autoplaque v2.5) (Supplemental Methods). Plaque analysis yielded a total of 19 quantitative measures: 12 related to plaque morphology and burden, 5 related to luminal stenosis, and 2 related to vessel dimensions. The average analysis time was approximately 25 minutes per patient.

Non-invasive FFR Derived from CCTA

FFR_CT was performed using dedicated software (HeartFlow FFR_CT version 2.7, Redwood City, California), as previously described²⁰ (Supplemental Methods).

Invasive Coronary Angiography and FFR Measurement

ICA was performed according to a standard protocol¹⁴. Using a 0.014-inch sensor-tipped guidewire introduced through a 5- or 6-F catheter, all major coronary arteries were routinely interrogated by FFR; except for in tight lesions >90% to avoid the risk of coronary dissection by the pressure wire. FFR was calculated as the ratio of mean distal intracoronary pressure to mean arterial pressure during maximal hyperemia induced via intracoronary (150 μg) or intravenous (140 μg/kg/min) adenosine infusion. The primary study endpoint was vessel-specific ischemia as defined by FFR ≤0.80. In cases where FFR was not performed, vessels with stenosis severity >90% obtained by quantitative coronary angiography were assigned to the FFR-significant group¹⁴.

[¹⁵O]H₂O Positron Emission Tomography

Patients were scanned on a hybrid PET-CT device (Philips Gemini TF64, Philips Healthcare), as previously described¹⁴. A dynamic PET perfusion scan was performed during rest and adenosine (140μg/kg/min)-induced hyperemia, using 370 MBq of [¹⁵O]H₂O as the radioactive tracer. PET images were sent to a blinded core laboratory (Turku University Hospital, Turku, Finland), where quantitative hyperemic MBF was calculated for all 3 major vascular territories derived from standard segmentation. The secondary study endpoint was impaired per-vessel hyperemic MBF (≤2.30 mL/min/g of perfusable myocardial tissue)².

Machine Learning Score

Model Building

Figure 1 is a flow chart of the ML methodology. All 19 CCTA-based quantitative plaque measures were used to create a ML model for the prediction of FFR-defined ischemia. The model was built using the XGBoost algorithm, which implements an ensemble of gradient-boosted decision trees, combining multiple weak classifiers (one-level decision trees) to produce a single strong classifier for outcome prediction²¹. In a prior study, we applied a boosted ensemble approach using the LogitBoost algorithm to data from the NXT (Analysis of Coronary Blood Flow using CT Angiography: Next Steps) trial for predicting invasive FFR ≤0.80¹³. For the present study, we leveraged XGBoost as a state-of-the-art algorithm which has shown high performance for cardiac CT-based risk stratification^22–24. ML analysis was performed using R 3.5 with XGBoost 1.3.3.

Figure 1. Flow chart of machine learning workflow.

A machine learning (ML) model for the prediction of vessel-specific ischemia was built, trained, and tuned in the NXT trial. The model’s predictive performance was evaluated in the PACIFIC trial as an unseen independent test set. The same ML model was then applied for the discrimination of impaired hyperemic MBF by PET in the PACIFIC dataset. CCTA, coronary computed tomography angiography; FFR, fractional flow reserve; MBF, myocardial blood flow.

Model Training

The ML model was trained in 254 patients and 484 vessels from the NXT trial. Firstly, XGBoost hyperparameters such as maximum depth of trees, minimum child weight, gamma, and number of estimators were tuned using Grid Search and 10-fold stratified cross validation. Grid Search considers a range of parameter combinations to find a potential combination of tuned hyperparameters which yields the best area under the receiver-operating characteristic curve (AUC). Ten-fold cross-validation was used within Grid Search to further increase the confidence of hyperparameter selection²⁵. Secondly, after tuning the hyperparameters from cross validation, the model was refitted on the entire NXT training set for the trained model.

Model Testing

Using the PACIFIC trial as an unseen independent test set, we evaluated the performance of the trained model for predicting FFR-defined ischemia. For each vessel, the model generated a continuous ML score (ranging from 0–1) as the predicted probability of ischemia. The XGBoost algorithm also assigned an importance value to each feature in the ML model, based on the feature’s “gain”, or relative contribution to model prediction when used in the decision trees. In a secondary analysis, the ML score was employed for the prediction of impaired hyperemic MBF.

Explainable ML

A summary chart of feature importance ranking based on gain was used to explain the output of the ML model for the PACIFIC cohort. In addition, the SHAP (SHapley Additive exPlanations) method was used to explain individualized ML predictions for patients in the cohort. The SHAP value measures each feature’s marginal impact or weight in the model for an individual, in comparison to the average model prediction²⁶. For clinical interpretability, we constructed waterfall plots of SHAP values for individuals with and without vessel-specific ischemia, using a Python 3.7 module with the XGBoost library.

Statistical Analysis

Continuous variables are expressed as mean ± standard deviation or median (interquartile range), as appropriate. Categorical variables are expressed as frequencies (percentages). Differences between continuous variables were compared using the independent samples t-test or Wilcoxon rank sum test, as appropriate. Categorical variables were compared with the Chi-square test or Fisher’s exact test. Correlations between continuous variables were assessed using the Spearman’s rank correlation coefficient. For clinical intuition, we performed traditional statistical analysis using a mixed effects logistic regression model with quantitative plaque features (percent diameter stenosis and volumes of total, calcified, non-calcified, and low-density non-calcified plaque) to predict FFR-defined ischemia. For the endpoints of FFR-defined ischemia and impaired hyperemic MBF, the AUC of the ML score was calculated and compared with that of visual stenosis grade, FFR_CT, the logistic regression model, and the prior Logitboost-based model using the method of Delong et al.²⁷. An optimum ML score cut-off for predicting FFR-defined ischemia was derived from the NXT trial using Youden’s J statistic²⁸ (sensitivity + specificity – 1) and then applied to the PACIFIC cohort for performance validation. Calibration of the ML score was visualized with a plot of the predicted risk versus observed risk of FFR-defined ischemia. Agreement between the predictions and observed outcomes was measured using the Brier score (range 0–1), with smaller values indicating better calibration²⁹. The improvement in risk classification provided by the ML score over and above visual stenosis grade was assessed using the continuous net reclassification index (NRI) ³⁰. Statistical analyses were performed using R 3.5, R Studio 1.2.1335. Stata 14.0 (StataCorp, College Station, TX, USA) was used for logistic regression. A 2-sided p-value <0.05 indicated statistical significance.

RESULTS

Study Population

The final study population comprised 203 patients and 601 evaluable coronary arteries, following exclusion of 5 patients and 23 vessels due to non-evaluable CCTA scans. Baseline patient characteristics have been previously reported and are shown in Table S1. FFR_CT analysis was feasible in 489 (81.3%) vessels. A total of 581 (96.7%) vessels were interrogated by invasive FFR (n=540; 92.9%) or showed >90% stenosis on ICA (n=41; 7.1%); FFR was ≤0.80 in 100 (17.2%) vessels. PET imaging was incomplete in 4 patients due to claustrophobia or technical reasons. In the remaining 199 patients, 195 (32.7%) vessels had impaired hyperemic MBF.

Qualitative Plaque Characteristics

Table 1 provides a summary of CCTA findings in the study population. Vessels with FFR-defined ischemia had a higher median stenosis grade compared to vessels without ischemia (3 [2–4] vs. 1 [0–2]; p<0.001). Similarly, FFR-defined ischemic vessels more frequently exhibited each of the individual adverse plaque characteristics compared with non-ischemic vessels (all p<0.001).

Table 1.

Per-vessel CCTA characteristics

	Total CCTA Evaluable Vessels (n = 601)	Vessels with FFR>0.80 (n = 442)	Vessels with FFR ≤0.80 or >90% stenosis (n = 139)	p Value for Subgroups
Visual stenosis grade	1 (0–3)	1 (0–2)	3 (2–4)	<0.001
Qualitative plaque characteristics
Positive remodeling	62 (10.3)	17 (3.8)	45 (32.4)	<0.001
Low-attenuation plaque	73 (12.1)	19 (4.3)	54 (38.8)	<0.001
Spotty calcification	33 (5.5)	10 (2.3)	23 (16.5)	<0.001
Napkin-ring sign	21 (3.5)	5 (1.1)	16 (11.5)	<0.001
Quantitative plaque measurements
Plaque length, mm	44.8±34.5	21.8±28.5	60.3±37.7	<0.001
Plaque volume, mm3
Total plaque	227.4±245.1	156.9±203.1	397.7±261.3	<0.001
Noncalcified plaque	119.8±135.0	62.4±81.6	203.6±194.2	<0.001
Low-density noncalcified plaque	20.7±34.4	9.1±19.7	42.3±54.1	<0.001
Calcified plaque	108±140.4	68.0±92.1	178.3±184.6	<0.001
Plaque burden, %
Total plaque	38.2±29.1	28.3±24.2	57.3±16.6	<0.001
Noncalcified plaque	20.0±18.5	14.4±13.2	31.7±17.1	<0.001
Low-density noncalcified plaque	3.5±3.9	1.9±3.7	6.1±4.8	<0.001
Calcified plaque	18.1±17.2	15.3±16.5	24.8±19.7	0.01
Plaque composition, %
Noncalcified plaque	48.9±20.2	40.6±21.4	52.1±23.8	<0.001
Low-density noncalcified plaque	9.1±7.8	6.1±5.9	12.6±7.1	<0.001
Calcified plaque	46.6±24.5	48.5±29.3	50.6±34.6	0.22
Diameter stenosis, %	42.2±25.2	19.9±21.0	60.8±27.7	<0.001
Area stenosis, %	60.2±24.5	31.9±30.1	75.7±24.7	<0.001
Minimum luminal diameter, mm	1.8±0.9	2.1±0.9	1.1±0.8	<0.001
Minimum luminal area, mm2	3.1±2.9	3.9±3.0	1.4±1.5	<0.001
Contrast density difference, %	20.9±13.0	10.5±12.2	27.9±13.0	<0.001
Remodeling index	1.1±0.3	0.8±0.7	1.3±0.4	<0.001
Vessel volume, mm3	598.1±478.5	502.6±468.2	734.1±512.6	<0.001
FFRCT (n = 489)
FFRCT ≤0.80	163 / 489 (33.3)	59 / 374 (15.8)	104 / 115 (90.4)	<0.001
FFRCT	0.81±0.13	0.86±0.08	0.63±0.13	<0.001

Values expressed as n (%), mean ± SD, or median (interquartile range).

CCTA, coronary computed tomography angiography; FFR, fractional flow reserve.

Quantitative Plaque Measurements

The volumes and burdens of total plaque and individual plaque components (NCP, LDNCP, CP) were higher in vessels with FFR-defined ischemia compared to vessels without ischemia (all p<0.001; Table 1). Quantitative percent diameter stenosis and percent area stenosis were greater in ischemic versus non-ischemic vessels (60.8±27.7% vs. 19.9±21.0%; and 75.7±24.75 vs. 31.9±30.1%; both p<0.001), as were contrast density difference (27.9±13.0% vs. 10.5±12.2%) and vessel remodeling index (1.3±0.4 vs. 0.8±0.7; both p<0.001).

Prediction of FFR-defined Ischemia

The ML score demonstrated an AUC of 0.92 (95% confidence interval [CI] 0.89–0.94) for predicting FFR-defined ischemia, which was significantly greater than that of visual stenosis grade (0.84, 95% CI 0.80–0.87; p<0.001) and comparable with that of FFR_CT (0.93, 95% CI 0.91–0.96; p=0.34) (Figure 2). The ML score exhibited a moderate inverse correlation with invasive FFR values (r=−0.56) and a strong inverse correlation with FFR_CT values (r=−0.67).

Figure 2. Performance of the ML score for discrimination of invasive FFR-defined ischemia.

The machine learning (ML) score integrating quantitative plaque features outperformed visual stenosis grade and was comparable with FFR_CT for discriminating ischemia by invasive fractional flow reserve (FFR). AUC, area under the receiver-operating characteristic curve; FFR_CT, FFR derived from computed tomography.

At a cut-off value of 0.25 determined by Youden’s J statistic, the ML score had an accuracy, sensitivity, and specificity of 84%, 87%, and 82%, respectively, for ischemia discrimination. A comparison of the XGBoost-based model with the prior LogitBoost-based model is detailed in the Supplemental Methods.

The ranking of feature importance for ML prediction in the PACIFIC cohort is shown in Figure 3. Quantitative percent diameter stenosis and LDNCP volume had greatest feature importance, followed by percent area stenosis, minimum luminal diameter, and contrast density difference. Conversely, features relating to CP contributed least to ML prediction.

Figure 3. Ranking of feature importance for ML ischemia prediction in the PACIFIC cohort.

Features from CCTA-based quantitative plaque analysis are ranked in order of importance for ischemia prediction and color-coded as: luminal (blue), plaque-specific (red), or vessel-related (gray). CCTA, coronary computed tomography angiography; CP, calcified plaque; LDNCP, low-density noncalcified plaque; NCP, noncalcified plaque.

The calibration plot of the ML score (Figure 4) shows the predicted probability of ischemia to correspond well with the actual observed frequency of ischemic vessels. The Brier score for ML (range 0–1) was 0.12, indicating a low difference between the predicted risk and observed risk for ischemia.

Figure 4. Calibration plot of the ML score for FFR-defined ischemia.

The red dots and lines represent the average predicted risk of per-vessel ischemia per decile of the ML score (X-axis). Bars show the actual observed frequency of ischemia per decile of the ML score. FFR, fractional flow reserve; ML, machine learning.

Adding the ML score to visual stenosis grade resulted in substantial net reclassification improvement for FFR-defined ischemia (NRI 1.16, 95% CI 1.00–1.33, p<0.001), driven primarily by the reclassification of non-events (82%) rather than events (37%; both p<0.001).

Individualized Explanations of the ML Score

Figure 5 depicts SHAP waterfall plots explaining ML prediction of per-vessel ischemia in two individual patients from the PACIFIC cohort. The calculation of the ML score begins at the mean prediction of the training set, then sums SHAP values that are color coded red for increasing or blue for decreasing the probability of FFR-defined ischemia, and ends at the individual prediction. Case 1 demonstrates a high ML score in the left anterior descending artery; invasive FFR was 0.72 (ischemic). In Case 2, a low ML score was computed for the right coronary artery, with a subsequent FFR of 0.91 (non-ischemic).

Figure 5. Individualized prediction of FFR-defined ischemia with explainable ML: case examples.

Beginning from the mean prediction at the bottom (machine learning [ML] score 0.217, or probability of 21.7%), the waterfall plots display SHAP (SHapley Additive exPlanations) values in increasing order of magnitude, with red for increasing risk and blue for decreasing risk of ischemia. The positive or negative SHAP value of each feature is progressively added to reach the individual prediction. Case 1 shows quantitative plaque analysis and ML score calculation in the left anterior descending artery of a 63-year old male. The highest impact feature increasing the ischemia risk was percent diameter stenosis (68.5%), followed by low-density noncalcified plaque (LDNCP) volume (125.3 mm³), percent area stenosis (77.4%), and contrast density difference (43.5%). The resultant ML score was 0.704 (70.4% probability of ischemia); invasive fractional flow reserve (FFR) was 0.68 (ischemic). In such a patient, the ML score could inform referral for ICA and revascularization following CCTA. Case 2 is a right coronary artery in a 75-year old female. Percent diameter stenosis (35.1%) was the highest impact adverse feature, while the absence of LDNCP was protective. The ML score was 0.196 (19.6% probability of ischemia); subsequent FFR was 0.91 (non-ischemic). In this example, based on the ML score, downstream testing could have been deferred.

Prediction of Impaired MBF

Impaired hyperemic MBF was more frequently observed in vessels with versus without FFR-defined ischemia (77.9% vs 29.6%; p<0.001); mean MBF values were 1.98±1.02 versus 3.51±1.21, respectively (p<0.001).

The ML score exhibited an AUC of 0.80 (95% CI 0.75–0.84) for predicting impaired hyperemic MBF, which was superior to that of visual stenosis grade (0.74, 95%CI 0.69–0.79; p=0.02) and comparable with that of FFR_CT (0.77, 95% CI 0.73 to 0.82; p=0.16) (Figure S1). At an optimum threshold of 0.25, the ML score yielded an accuracy, sensitivity, and specificity of 77%, 73%, and 80%, respectively, for impaired MBF prediction.

Within the ML score, LDNCP volume was the highest-ranked predictor of impaired MBF, followed by NCP volume and total plaque volume. Minimal luminal diameter and percent diameter stenosis were also highly predictive features, whereas percent area stenosis and minimum luminal area had lower importance for ML prediction (Figure S2).

Comparison of the ML Score with a Logistic Regression Model

The ML score outperformed a mixed effects logistic regression model with quantitative plaque parameters for the prediction of both FFR-defined ischemia (AUC 0.92, 95% CI 0.89–0.94 vs. 0.88, 95% CI 0.84–0.91; p=0.002) and impaired hyperemic MBF (AUC 0.80, 95% CI 0.76–0.84 vs. 0.77, 95% CI 0.73–0.81; p=0.03).

DISCUSSION

This PACIFIC trial substudy represents an external validation of a ML score integrating CCTA-based plaque measures for the prediction of ischemia by invasive FFR. We show that: 1) the ML score accurately quantifies ischemia risk, performing superiorly to standard CCTA stenosis evaluation and comparably to FFR_CT; 2) the same score exhibits high discriminative ability for impaired hyperemic MBF by PET; 3) LDNCP volume has a large impact on ML prediction of both FFR-defined ischemia and impaired MBF; and 4) explainable ML enables precise identification of the plaque features contributing to ischemia risk in an individual vessel or patient.

There is emerging evidence that CCTA-derived atherosclerotic plaque characteristics influence both FFR and myocardial perfusion³. Prior studies have examined either endpoint in isolation, focused on a select few plaque metrics, or employed qualitative CCTA evaluation^5–9. The present analysis is the first to apply ML to a comprehensive set of quantitative plaque features for the prediction of both FFR-defined ischemia and impaired hyperemic MBF by PET. ML has recently demonstrated excellent performance for cardiovascular CT-based outcome prediction^31,32, driven by its ability to iteratively select and weight individual imaging features as well as identify multidimensional associations between features. This is particularly pertinent for CCTA, as plaque characteristics display complex relationships which may be difficult to model using conventional statistical methods. The use of the SHAP tool in this study to provide individualized explanations of the ML score (explainable ML) allows users without ML expertise to interpret the direction, magnitude, and relative importance of plaque features in a given vessel or patient for outcome prediction. The current ML score represents a substantial improvement upon our prior work¹³; largely attributable to the unique advantages of the state-of-the-art XGBoost algorithm over other gradient boosting methods. XGBoost is highly flexible in setting an objective function, uses regularization to prevent model overfitting, handles missing data implicitly, and has built-in cross validation²¹. It has been widely applied in the domains of computer science and medicine, and is highly scalable due to its parallel processing and rapid computation speeds.

Within the ML score, LDNCP volume was the second-highest ranked predictor of FFR-defined ischemia behind quantitative percent diameter stenosis. This is consistent with recent reports demonstrating CCTA-determined LDNCP volume to have independent and incremental predictive value for abnormal FFR beyond stenosis severity^5,6. Mechanistically, plaques with larger necrotic cores often exhibit significant positive remodeling, and the maximally stretched vessel may restrict further local dilatation analogous to the limits of Glagov’s phenomenon³³. Furthermore, a large, lipid-rich necrotic core induces local oxidative stress and inflammation, which can lead to local endothelial dysfunction and an impaired vasodilatory response^34–37. Also highly ranked in the ML score was contrast density difference, a measure of the maximum reduction in luminal contrast attenuation across a lesion. This metric is automatically calculated following quantitative plaque analysis, and has been shown to outperform the transluminal attenuation gradient and visual stenosis grade for detecting hemodynamically significant CAD³⁸. Concordant with prior studies demonstrating qualitative evaluation of positive remodeling to predict FFR-defined ischemia^4,7, our automated quantitative measure of the remodeling index also ranked within the top 10 features of the ML score.

With respect to discrimination of impaired MBF by PET, the volumes of LDNCP, NCP, and total plaque had greatest contribution to ML prediction, ranking above quantitative metrics of luminal stenosis. While FFR may be abnormal in the presence of a significant focal epicardial stenosis, PET MBF measures perfusion across the entire coronary vascular bed, including both the epicardial arteries and myocardial microvasculature. An increased plaque volume in a major epicardial artery may be reflective of diffuse atherosclerotic or downstream small vessel disease; important determinants of myocardial perfusion³⁹. Furthermore, intracoronary imaging studies have demonstrated large necrotic core volumes to associate with both epicardial and microvascular endothelial dysfunction on coronary reactivity testing^40,41. Recently, CCTA-based per-vessel quantification of NCP has been shown to predict abnormal regional hyperemic MBF by PET, independently of stenosis severity and other plaque characteristics^4,9. The present ML analysis extends these findings by highlighting the low-density or necrotic core component of NCP as the major contributor to risk of impaired MBF.

Functional tools which complement CCTA without the need for additional image acquisition or drug administration have been developed to improve the detection of lesion-specific ischemia. The application of computational fluid dynamics (CFD) to anatomic CCTA datasets enables the noninvasive determination of FFR_CT using commercially available software with high diagnostic accuracy^12,42. However, this process is time-consuming and computationally expensive, requiring off-site processing by a core laboratory. On-site FFR_CT using a reduced-order CFD algorithm has been described, yet is heavily dependent on physician interaction⁴². More recently, a deep learning-based method for on-site FFR_CT was trained using a large database of synthetically-generated coronary trees, and applied to a real-world registry where it performed equally as well as a CFD-based approach for predicting ischemia by invasive FFR⁴³. In contrast to all these techniques, our proposed ML score relies only on anatomic information from CCTA without the addition of physiologic parameters, utilizes semiautomated plaque analysis which is highly reproducible by trained clinicians or technicians^11,19, and can be computed in less than 30 minutes on a standard workstation. We showed the performance of the ML score to be comparable to the FFR_CT analysis platform most widely used in clinical care.

Our study findings have several important clinical implications. Increasingly advanced analytic approaches are required to interpret the vast amount of high-dimensional data being generated by CCTA. Here, we demonstrate that ML can be used to combine a wide range of quantitative plaque features into an objective risk score for the prediction of ischemia and impaired MBF. The flexible nature of the ML model will enable the addition of new plaque features and other CCTA-derived metrics (such as subtended myocardial mass) to continuously update and optimize prediction. At the same time, rapid improvements in artificial intelligence algorithms will facilitate full automation of software-based plaque quantification. In future, our proposed workflow could be embedded in the background of routine CCTA analysis and reporting software, automatically performing plaque measurements and calculating individual risk in real-time. The ML score would function as a clinical decision support tool for physicians, informing referral for ICA in patients with high ischemia risk and deferral of downstream testing in those with low ischemia risk. This on-site tool has the potential to enhance anatomic CCTA-guided care and improve the efficiency of referral to ICA; future studies will need to assess whether it increases ICA diagnostic yield and predicts revascularization. In this era of personalized medicine, there is a growing demand for more precise and patient-tailored risk prediction. Explainable ML using SHAP enables the interrogation of individualized predictive models by clinicians or researchers who lack familiarity with ML techniques. SHAP is applicable to any ML model, comparable across different datasets, and easily understood due to its additive format. This may enable physicians to interact with the inputs for an individual prediction, and to survey the impact of specific plaque features on ischemia risk.

Limitations

There are several limitations to this study. First, CCTA-based plaque findings were not confirmed by intracoronary imaging. However, plaque quantification using the described method has previously shown excellent agreement with intravascular ultrasound¹¹. Second, the ML score was trained specifically using invasive FFR-defined ischemia as the reference standard, and hence it follows that its predictive performance in the test set was higher for this endpoint than for the secondary endpoint of impaired hyperemic MBF. Third, quantitative PET was performed with the less frequently used [¹⁵O]H₂O tracer, and our findings may not be generalizable to more commonly used PET tracers such as [¹³N]NH₃ and Rubidium-82. Finally, although the ML model was trained on the multicenter NXT trial which incorporated several different CT scanners and acquisition protocols, the PACIFIC trial test set comprised a single center and CT scanner.

Conclusions

An externally validated ML score integrating CCTA-based quantitative plaque features accurately predicts ischemia by invasive FFR and abnormal MBF by PET, performing superiorly to standard CCTA stenosis evaluation and comparably to FFR_CT. Explainable ML techniques enable precise identification of the plaque features contributing ischemia risk in an individual vessel or patient.

CLINICAL PERSPECTIVE.

There is increasing evidence of a pathophysiological interplay between plaque morphology and coronary physiology. Coronary computed tomography angiography (CCTA), a first-line test for the evaluation of luminal stenosis, enables comprehensive atherosclerotic plaque characterization and quantification. Here, we show a machine learning (ML) score integrating CCTA-derived quantitative plaque features to accurately predict ischemia by invasive fractional flow reserve (FFR) and impaired myocardial blood flow by positron emission tomography. Our ML score performs superiorly to conventional CCTA stenosis evaluation and comparably to non-invasive FFR derived from CCTA (FFR_CT), which represent current clinical practice. We use explainable ML techniques which enable precise identification of the plaque features contributing risk of ischemia in an individual vessel or patient. As artificial intelligence applications permeate into daily cardiovascular care, our proposed ML workflow could in future be embedded into routine CCTA image analysis and reporting software, automatically calculating individualized ischemia risk in real-time. The ML score would function as a clinical decision support tool for physicians, informing referral for downstream noninvasive stress testing or invasive angiography.

NON-STANDARD ABBREVIATONS AND ACRONMYS

CFD

computational fluid dynamics

CP

calcified plaque

LDNCP

low-density noncalcified plaque

NCP

noncalcified plaque

NXT

Analysis of Coronary Blood Flow using CT Angiography: Next Steps

PACIFIC

Prospective Comparison of Cardiac PET/CT, SPECT/CT Perfusion Imaging and CT Coronary Angiography with Invasive Coronary Angiography

SHAP

SHapley Additive exPlanations