Machine learning from quantitative coronary computed tomography angiography predicts ischemia and impaired myocardial blood flow

Background

Atherosclerotic plaque characteristics influence the hemodynamic consequences of coronary lesions. This study sought to assess the performance of a machine learning (ML) score integrating coronary computed tomography angiography (CCTA)-based quantitative plaque features for the prediction of ischemia by invasive fractional flow reserve (FFR) and impaired myocardial blood flow (MBF) by [15O]H2O positron emission tomography (PET).

Methods

This post-hoc analysis of the PACIFIC (Prospective Comparison of Cardiac PET/CT, SPECT/CT Perfusion Imaging and CT Coronary Angiography With Invasive Coronary Angiography) trial included 208 patients with suspected coronary artery disease who underwent CCTA, [15O]H2O PET, and 3-vessel invasive FFR. Plaque quantification from CCTA was performed using semiautomated software. A boosted ensemble ML algorithm (XGBoost) trained on data from the NXT (Analysis of Coronary Blood Flow using CT Angiography: Next Steps) trial was used to develop a ML score for the prediction of per-vessel ischemia (invasive FFR ≤0.80). The performance of the ML score was evaluated in 551 vessels from the PACIFIC trial for external validation. Thereafter, we assessed the discriminative ability of the ML score for per-vessel impaired hyperemic MBF (≤2.30 mL/min/g).

Results

In total, 138 (25.0%) vessels had ischemia and 195 (35.4%) vessels had impaired hyperemic MBF. CCTA-derived quantitative percent diameter stenosis and low-density noncalcified plaque (LDNCP) volume were higher in ischemic vessels compared with non-ischemic vessels (60.8% vs. 19.9%; and 42.3 mm3 vs. 9.1 mm3; both p<0.001). The ML score demonstrated a significantly higher area under the receiver-operating characteristic curve (AUC) for predicting ischemia (0.92, 95% confidence interval [CI] 0.89–0.94) compared with visual stenosis grade (0.84, 95% CI 0.80–0.87; p<0.001). Overall, quantitative percent diameter stenosis and LDNCP volume had greatest feature importance for ML, followed by percent area stenosis, minimum luminal diameter, and contrast density drop (Figure 1). An individualized explanation of ML ischemia prediction is shown in Figure 2. When applied for impaired MBF discrimination, the ML score exhibited an AUC of 0.82 (95% CI 0.78–0.85) and was superior to visual stenosis grade (AUC 0.76, 95% CI 0.72–0.80; p=0.03).

Conclusions

An externally validated ML score integrating CCTA-based quantitative plaque features accurately predicts FFR-defined ischemia and abnormal MBF by PET, outperforming standard visual CCTA interpretation.

Funding Acknowledgement

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): National Heart, Lung, and Blood Institute, United States Performance of the ML scoreIndividual explanation of ML prediction

Pericoronary adipose tissue attenuation, low-attenuation plaque burden and 5-year risk of myocardial infarction

Introduction

Pericoronary adipose tissue (PCAT) attenuation has emerged as a surrogate marker of pericoronary inflammation. To date, no studies have compared the impact of pericoronary adipose tissue (PCAT) attenuation and quantitative plaque burden on cardiac outcomes.

Purpose

We aimed to establish the relative merits of these approaches to risk prediction and hypothesised that the combination of PCAT attenuation and quantitative plaque burden measures could provide additive and improved prediction of myocardial infarction in patients with stable chest pain.

Methods

In a post-hoc analysis of a randomized controlled trial, we investigated the association between the future risk of fatal or non-fatal myocardial infarction and PCAT attenuation measured from CT coronary angiography using multivariable Cox regression models including plaque burden, obstructive coronary disease and cardiac risk score (incorporating age, sex, diabetes, smoking, hypertension, hyperlipidaemia and family history of cardiovascular disease).

Results

In 1697 evaluable participants (mean age 58±10 years), there were 37 myocardial infarctions after a median follow-up of 4.7 [interquartile interval, 4.0–5.7] years. Median low-attenuation plaque burden was 4.20 [0–6.86] % and mean PCAT −76±8 Hounsfield units (HU).

PCAT attenuation of the right coronary artery (RCA) was predictive of myocardial infarction (hazard ratio [HR] 1.55, 95% CI 1.08–2.22; p=0.017, per 1 standard deviation increment) with an optimum threshold of −70.5 HU [Hazards ratio (HR) 2.45, 95% CI 1.2–4.9; p=0.01]. Univariable analysis also identified the burden of non-calcified, low-attenuation and calcified plaque as well as Agatston coronary calcium score, presence of obstructive coronary artery disease and cardiovascular risk score were predictors of myocardial infarction (Figure 1). In multivariable analysis, only the low-attenuation plaque burden (HR 1.80, 95% CI 1.16 to 2.81, p=0.011, per doubling) and PCAT-RCA (HR 1.47 95%1.02 to 2.13, p=0.040, per standard deviation increment) remained predictors of myocardial infarction (Figure 1).

In multivariable analysis, adding PCAT-RCA ≥-70.5 HU to low-attenuation plaque burden >4% (optimum threshold for future myocardial infarction; HR = 4.87, 95% CI 2.03–11.78; p<0.0001) led to improved prediction of future myocardial infarction (HR 11.7, 95% CI 3.3–40.9; p<0.0001); Figure 2. In ROC analysis, integration of PCAT-RCA attenuation and LAP burden, increased the prediction for myocardial infarction compared to LAP alone (ΔAUC=0.04; p=0.01).

Conclusion

CT coronary angiography defined PCAT attenuation and low-attenuation plaque have marked and additive predictive value for the risk of fatal or non-fatal myocardial infarction.

Funding Acknowledgement

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): The Chief Scientist Office of the Scottish Government Health and Social Care Directorates, British Heart Foundation, National Institute of Health/National Heart, Lung, and Blood Institute grant

Deep learning-based plaque quantification from coronary computed tomography angiography: external validation and comparison with intravascular ultrasound

Background

Atherosclerotic plaque quantification from coronary computed tomography angiography (CTA) enables accurate assessment of coronary artery disease burden, progression, and prognosis. However, quantitative plaque analysis is time-consuming and requires high expertise. We sought to develop and externally validate an artificial intelligence (AI)-based deep learning (DL) approach for CTA-derived measures of plaque volume and stenosis severity. We compared the performance of DL to expert readers and the gold standard of intravascular ultrasound (IVUS).

Methods

This was a multicenter study of patients undergoing coronary CTA at 11 sites, with software-based quantitative plaque measurements performed at a per-lesion level by expert readers. AI-based plaque analysis was performed by a DL novel convolutional neural network which automatically segmented the coronary artery wall, lumen, and plaque for the computation of plaque volume and stenosis severity. Using expert measurements as ground truth, the DL algorithm was trained on 887 patients (4,686 lesions). Thereafter, the algorithm was applied to an independent test set of 221 patients (1,234 lesions), which included an external validation cohort of 171 patients from the SCOT-HEART (Scottish Computed Tomography of the Heart) trial as well as 50 patients who underwent IVUS within one month of CTA. We report the performance of AI-based plaque analysis in the independent test set.

Results

Within the external validation cohort, there was excellent agreement between DL and expert reader measurements of total plaque volume (intraclass correlation coefficient [ICC] 0.876), noncalcified plaque volume (ICC 0.869), and percent diameter stenosis (ICC 0.850; all p<0.001). When compared with IVUS, there was excellent agreement for DL total plaque volume (ICC 0.945), total plaque burden (ICC 0.853), minimal luminal area (ICC 0.864), and percent area stenosis (ICC 0.805; all p<0.001); with strong correlation between DL and IVUS for total plaque volume (r=0.915; p<0.001; Figure). The average DL plaque analysis time was 20 seconds per patient, compared with 25–30 minutes taken by experts.

Conclusions

AI-based plaque quantification from coronary CTA using an externally validated DL approach enables rapid measurements of plaque volume and stenosis severity in close agreement with expert readers and IVUS.

Funding Acknowledgement

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): National Heart, Lung, and Blood Institute, United States

Association of coronary artery calcium score groups with qualitative and quantitatively assessed adverse plaque

Introduction

Coronary artery calcification is a marker of cardiovascular risk, but its association with qualitatively and quantitatively assessed plaque subtypes on coronary computed tomography (CT) angiography (CCTA) is unknown.

Methods

In this post-hoc analysis, CT images and clinical outcomes were assessed in SCOT-HEART trial participants. Agatston coronary artery calcium score (CACS) was measured on non-contrast CT and was stratified as zero (0 Agatston units, AU), minimal (1 to 9AU), low (10 to 99AU), moderate (100 to 399AU), high (400 to 999AU) and very high (≥1000AU). Adverse plaques were investigated with qualitative (visual categorisation of positive remodelling, low-attenuation plaque, spotty calcification, napkin ring sign) and quantitative (calcified, non-calcified, low-attenuation and total plaque burden) methods.

Results

Images of 1769 patients were assessed (mean age 58±9 years, 56% male, median Agatston score 21 [interquartile range 0 to 230] AU). Of these 36% had a zero, 9% minimal, 20% low, 17% moderate, 10% high and 8% very high CACS. Amongst patients with a zero CACS, 14% had nonobstructive disease, 2% had obstructive disease, 2% had visually assessed adverse plaques and 13% had quantitative low-attenuation plaque (LAP) burden >4% (Figure 1). Non-calcified and low-attenuation plaque burden increased between patients with zero, minimal and low CACS (p<0.001), but there was no difference between those with medium, high and very high CACS. Over a median follow-up of 4.8 [4.1 to 5.7] years, fatal or non-fatal myocardial infarction occurred in 41 patients, 10% of whom had zero CACS. CACS ≥1000AU (Hazard ratio (HR) 4.55 [1.20 to 17.3], p=0.026) and low-attenuation plaque burden (HR 1.74 [1.19 to 2.54], p=0.004) were the only predictors of myocardial infarction, independent of obstructive disease and cardiovascular risk score. Figure 2 shows example CCTA images in a patient with zero CACS, non-calcified plaque (red), low attenuation plaque (orange) burden >4% and obstructive disease in the left anterior descending coronary artery.

Conclusions

In patients with stable chest pain, a zero CACS is associated with a good prognosis, but 1 in 6 have coronary artery disease, including the presence of adverse plaques.

Funding Acknowledgement

Type of funding sources: Private grant(s) and/or Sponsorship. Main funding source(s): British Heart Foundation, National Institute of Health/National Heart, Lung, and Blood Institute

18F-sodium fluoride coronary uptake in patients with coronary artery bypass grafts

 

18F-Sodium fluoride (18F-NaF) positron emission tomography (PET) provides an assessment of active calcification (microcalcification) across a wide range of cardiovascular conditions including coronary artery disease, carotid and penile atherosclerosis, aortic and mitral valve disease, and abdominal aortic aneurysms. To date the significance of 18F-NaF uptake in patients with coronary artery bypass grafts (CABG) is unknown.

We aimed to characterize 18F-NaF activity in CABG patients.

We performed 18F-NaF PET (30-min long single bed position acquisition 1h after a 250mB injection of 18F-NaF) and coronary CT angiography in patients with multivessel coronary artery disease and followed them for fatal or non-fatal myocardial infarction over 42 [31,49] months. On motion-corrected datasets we quantified the whole-vessel coronary 18F-NaF PET uptake (the coronary microcalcification activity (CMA)) by measuring the activity of voxels above the background (right atrium activity) + 2 * standard deviations threshold. All study subjects underwent a comprehensive baseline clinical assessment including evaluation of their cardiovascular risk factor profile with the SMART [Secondary Manifestations of Arterial Disease] risk score calculated, and the coronary calcium burden assessed with calcium scoring (CCS).

Among 293 study participants (65±9 years; 84% male), 48 (16%) had a history of CABG. Although the majority 124/128 (97%) of coronary bypass grafts showed no uptake, 4 saphenous vein grafts presented with a CMA>0 (range: 2.5–11.5, Figure). While a similar proportion of patients with and without prior CABG showed increased coronary 18F-NaF uptake (CMA>0) (58.3% versus 71.4%, p=0.11) overall prior-CABG subjects had higher CMA (2.0 [0.3, 6.6] versus 0.6 [0, 2.7], p=0.001) and CCS (1135 [631, 2120] versus 225 [59, 542], p<0.001), respectively. In line with the differences in the calcification activity and the coronary calcium burden, the SMART risk scores were higher in CABG patients (23 [17, 28] versus 17 [12, 24], p=0.01), and these patients were also older (68±8 versus 64±8, p=0.01). Despite the aforementioned differences the incidence of myocardial infarction 5/48 (9%) versus 15/245 (6%) and MACE 6/48 (12%) versus 34/245 (14%) during follow-up between subjects with and without prior CABG was similar (p=0.44 and p=0.80, respectively).

CABG patients have a higher coronary microcalcification activity on 18F-NaF PET than multivessel coronary artery disease patients without prior CABG. Despite evidence of higher 18F-NaF uptake there is no difference in outcome between these two groups.

Figure 1. 18F-NaF uptake in CABG patients. (A) 63-year old male with prominent uptake in stented saphenous vein bypass grafts and native coronary arteries who experienced a non-fatal non ST elevation myocardial infarction during follow-up. (B) 70-year old male with evident uptake in native coronary arteries and only little 18F-NaF activity within coronary bypasses.

Funding Acknowledgement

Type of funding source: Other. Main funding source(s): National Heart, Lung, and Blood Institute/National Institute of Health (NHLBI/NIH), British Heart Foundation

Non-invasive quantitative characterization of aortic valve tissue composition from computed tomography angiography improves patient risk stratification in transcatheter aortic valve implantation

Background

Computed tomography angiography (CTA) performed for procedural planning of transcatheter aortic valve implantation (TAVI) can be used for a more complete characterization of aortic valve tissue beyond calcium assessment. Combining quantitative data on both noncalcified and calcified tissues may improve differentiation of aortic stenosis (AS) subtypes and prognostication post-TAVI.

Purpose

We sought to noninvasively assess aortic valve tissue composition with quantitative cardiac CTA in patients with AS and its prognostic vaalue in those who underwent TAVI.

Methods

In 185 consecutive AS patients in a prospective registry who underwent cardiac CTA before TAVR and 90 matched controls with normal aortic valves, non-luminal aortic valve tissue were identified using semi-automated software as non-calcified (low-attenuation [−30 to 30 Hounsfield Units (HU)], fibro-fatty (31 to 130 HU), fibrous (131 to 350 HU) and calcified (>650 HU) tissue; with total tissue as (non-calcified + calcified components). Volumes of each component and composition [(tissue component volume/total tissue volume) ×100%] were quantified. The association of aortic valve composition and clinical outcomes post-TAVI including all-cause mortality was evaluated using Valve Academic Research Consortium (VARC)-2 definitions.

Results

AS patients had greater aortic valve tissue volume (median 2000.2, vs 527.8 mm3, p<0.001) with a higher calcified tissue composition (41.8% vs 3.4%, p<0.001) compared to controls. Total aortic valve tissue (noncalcified and calcified) volume yielded the highest area under the operating curve (AUC) for diagnosing severe AS (0.93,95% CI:0.93–0.99) as compared to calcified tissue volume alone (0.87,95% CI:0.81–0.94, p=0.002). Low-flow low-gradient AS was associated with increase in total tissue volume compared to controls (1515.3 vs 527.8 mm3, p<0.001), with lower volumes of calcified tissue than high-gradient AS (412.5 vs 829.6 mm3, p<0.001). Device success was achieved in 88% (164 of 185) patients and prevalence of moderate or severe paravalvular leak was 3.8%, however no differences between in aortic valve composition were observed in patients with and without device success. Early safety endpoints occurred in 16.1% (29 of 180) patients and 30-day all-cause mortality was 4.4%. Whereas only calcified tissue volume was related to VARC-2 early safety, AUC for prediction of 30-day mortality post-TAVI was 0.793 (95% CI:0.685–0.901) for total tissue volume and 0.776 (95% CI:0.676–0.876) for calcified tissue volume.

Conclusions

Quantitative CTA assessment of aortic valve tissue volume and composition can improve identification of high-gradient AS and low-flow low-gradient AS patients referred for TAVI and predict 30-day mortality post-TAVI.

Funding Acknowledgement

Type of funding source: Public grant(s) – National budget only. Main funding source(s): National Heart, Lung, and Blood Institute (NHLBI)

5963Automated quantification of epicardial adipose tissue from non-contrast CT on multi-center and multi-vendor data using deep learning

Background

Epicardial adipose tissue (EAT), a metabolically active visceral fat depot surrounding the coronary arteries, has been shown to promote the development of atherosclerosis in underlying coronary vasculature.

Purpose

We evaluate the performance of deep learning (DL), a sub-group of machine learning algorithms, for robust and fully automated quantification of EAT on multi-center cardiac CT data.

Methods

In this study, 850 non-contrast calcium scoring CT scans, from multiple cohorts, scanners and protocols, with manual measurements of EAT from 3 different readers were considered. The DL method was based on a convolutional neural network trained to reproduce the expert measurement. DL global performance was first assessed using all the scans, and then compared to inter-observer variability on a subset of 141 scans. Finally, automated EAT progression was compared to manual measurement using baseline and follow-up serial scans available for 70 subjects. The proposed model was validated using 10-fold cross validation.

Results

Automated quantification was performed in 1.57±0.49 seconds compared to 15 minutes for manual measurement. DL provided high agreement with expert manual quantification for all scans (R=0.974, p<0.001) with no significant bias (0.53 cm3, p=0.13). EAT volume was higher in patients with hypertension (+18.02 cm3, p<0.001, N=442), with diabetes (+18.33 cm3, p<0.001, N=75) and with hypercholesterolemia (+7.33 cm3, p=0.039, N=508). Manual EAT volumes measured by two experienced readers on 141 scans were highly correlated (R=0.984, p<0.001) but presented a significant difference of 4.35 cm3 (p<0.001). On these 141 scans, DL quantifications were highly correlated to both experts' measurements (R=0.973, p<0.001; R=0.979, p<0.001) with significant and non-significant bias for readers 1 and 2 (5.19 cm3, p<0.001; 0.84 cm3, p=0.26), respectively. In 70 subjects, EAT progression quantified by DL correlated strongly with EAT progression measured by the expert reader (R=0.905, p<0.001) with no significant bias (0.64 cm3, p=0.43), and was related to increased non-calcified plaque burden quantified from coronary CT angiography (5.7% vs 1.8%, p=0.026).

Automated vs. manual EAT volume

Conclusion

Deep learning allows rapid, robust and fully automated quantification of EAT from calcium scoring CT. It performs as an expert reader and can be implemented for routine cardiovascular risk assessment.

Acknowledgement/Funding

1R01HL133616/01EX1012B/Adelson Medical Research Foundation

P6151Fully automated epicardial adipose tissue volume and density measured from non-contrast CT predict major adverse cardiovascular events in asymptomatic subjects

Background

Epicardial adipose tissue (EAT) volume and density has shown to correlate with standard markers of coronary artery disease (CAD) and may predict major adverse cardiovascular events (MACE).

Purpose

We aimed to evaluate the prognostic value of EAT volume and density measured by fully automated deep-learning software from non-contrast cardiac computed tomography (CT).

Methods

We assessed 2071 consecutive asymptomatic subjects (age 56±9 years, 59% male) from the EISNER (Early Identification of Subclinical Atherosclerosis by Noninvasive Imaging Research) trial with long-term follow-up after coronary artery calcium (CAC) measurement. EAT volume and mean density were quantified using automated deep-learning software from non-contrast cardiac CT. MACE was defined as myocardial infarction (MI), cardiac death, late (>90 days) revascularization and acute coronary syndrome (ACS). EAT volume and density were systematically compared to CAC score and atherosclerotic cardiovascular disease (ASCVD) risk score using Cox proportional hazards regression for MACE prediction.

Results

At 14±3 years, 217 subjects suffered MACE. In age-and-gender-adjusted multivariate analysis, ASCVD risk score, CAC (two-fold increase) and EAT volume (two-fold increase) were associated with increased risk of suffering MACE [Hazard Ratio (HR) (95% CI): 1.03 (1.01–1.04); 1.25 (1.19–1.30); and 1.36 (1.08–1.70) respectively, p<0.01 for all] (Figure); the corresponding Harrell's C-statistic was 0.76. The area-under-the curve from receiver-operator characteristic analysis for MACE prediction increased significantly from 0.69 to 0.77 (p<0.0001) when EAT volume and CAC were added to the current clinical standard (ASCVD, family history and obesity measures BMI and BSA). Both in men and women, increase in EAT volume was associated with increased risk of MACE, with HR 1.14 (1.06–1.22), p<0.001 in men vs. 1.15 (1.01–1.31), p=0.03 in women, for each 20 cubic centimeter increase in volume. EAT density (HU) was independently inversely associated with MACE [HR: 0.96 (0.93–0.99), p=0.01].

MACE Prediction

Conclusions

EAT volume and density measurements improve prediction of MACE in asymptomatic populations over the current clinical standard. Fully automated EAT volume and density quantification by deep-learning from non-contrast cardiac CT can provide additional prognostic value for the asymptomatic patient.

Acknowledgement/Funding

1R01HL133616, Forschungsstiftung Medizin Universitätsklinikum Erlangen, grant from Dr Miriam and Sheldon G. Adelson Medical Research Foundation

5966Whole-vessel coronary 18F-sodium fluoride coronary microcalfication activity is associated with Low density plaque

Background

18F-sodium fluoride (18F-NaF) showed promise in imaging vulnerable coronary plaques. To date reporting of the highest per patient target to background ratio (TBR), total number of lesions with visual uptake and whole-heart tracer activity have been proposed. Unfortunately, each of these approaches has limitations which become especially prominent in patients with multiple foci of uptake, where reproducible global per-vessel measures are required. In oncology, the total metabolic active tumor volume has been found to be a significant prognostic factor for disease progression, recurrence and death. We evaluated if such methodology could be applied to coronary PET imaging.

Purpose

To quantify whole-vessel coronary 18F-NaF PET activity by utilizing automatically derived coronary vessel regions of interest (ROI) from CT angiography and assess the relationship between coronary microcalcification activity (CMA) and per vessel quantitative plaque characteristics on coronary CT angiography (CTA).

Methods

Twenty patients (68±6 years old, 70% males) with multivessel coronary artery disease underwent a 30 min single bed position PET 1h after a 250mB injection of 18F-NaF and CTA on a hybrid PET/CT scanner. We assessed coronary 18F-NaF uptake using novel whole-vessel tubular and tortuous 3D ROIs which were automatically extracted from CTA datasets. Within such ROIs we measured mean standard uptake value (SUV), maximum TBR (TBRmax) and the activity of voxels (CMA) above 1.25 the background SUV (left atrium activity). We used a previously established 1.25 TBRmax threshold to distinguish vessels positive and negative for 18F-NaF uptake. Coronary CTA datasets were analyzed by semi-automated software to quantify volumes and percentage lesion content of non-calcified plaque (NCP), low-density non-calcified plaque (LD-NCP, attenuation <30 Hounsfield units) and calcified plaque (CP).

Results

13 (65%) patients and 24 (40%) out of 60 main epicardial vessels presented with 18F-NaF uptake exceeding the 1.25 TBRmax threshold. While coronaries positive for uptake had higher CMA 0.92 [0.17, 2.03] vs 0.0, p<0.001 and TBRmax 1.42 [1.35, 1.74] vs 1.09 [1.0, 1.19], there was no difference in whole-vessel SUVmean 0.90 [0.77, 1.17] vs 0.87 [0.78, 0.96], p=0.33 compared to 18F-NaF negative arteries. Of the quantitative plaque characteristics vessels positive for uptake had higher NCP 278.4 [145.6, 576.9] vs 184.6 [63.8, 367.0]mm3, p=0.030; and LD-NCP 8.4 [0.3, 11.0] vs 2.7 [12.1, 43.5]mm3, p=0.01. CMA showed a stronger correlation with LD-NCP (r=0.70, p<0.001) than TBRmax (r=0.52, p<0.001). On regression analysis LD-NCP acted as an independent predictor of CMA after adjustments for CP and vessel SUVmean (p<0.001).

Figure 1

Conclusions

Whole-vessel 18F-NaF coronary microcalcification activity assessment with CT angiography automatically derived 3-dimensional ROIs is feasible and the measured coronary microcalcification burden correlates well with low density plaque.

Acknowledgement/Funding

This research was supported by grants R01HL135557 and R01HL133616 from the NHLBI/NIH and a grant from the Dr. Miriam & Sheldon G. Adelson MRF

30Machine learning to predict the long-term risk of myocardial infarction and cardiac death based on clinical risk, coronary calcium and epicardial adipose tissue: a prospective study

Background/Introduction

Machine learning (ML) allows objective integration of clinical and imaging data for the prediction of events. ML prediction of cardiovascular events in asymptomatic subjects over long-term follow-up, utilizing quantitative CT measures of coronary artery calcium (CAC) and epicardial adipose tissue (EAT) have not yet been evaluated.

Purpose

To analyze the ability of machine learning to integrate clinical parameters with coronary calcium and EAT quantification in order to improve prediction of myocardial infarction (MI) and cardiac death in asymptomatic subjects.

Methods

We assessed 2071 consecutive subjects [1230 (59%) male, age: 56.049.03] from the EISNER (Early Identification of Subclinical Atherosclerosis by Noninvasive Imaging Research) trial with long-term follow-up after non-enhanced cardiac CT. CAC (Agatston) score, age-and-gender-adjusted CAC percentile, and aortic calcium scores were obtained. EAT volume and density were quantified using a fully automated deep learning method. Extreme gradient boosting, a ML algorithm, was trained using demographic variables, plasma lipid panel measurements, risk factors as well as CAC, aortic calcium and EAT measures from CAC CT scans. ML was validated using 10-fold cross validation; event prediction was evaluated using area-under-receiver operating characteristic curve (AUC) analysis and Cox proportional hazards regression. Optimal ML cut-point for risk of MI and cardiac death was determined by highest Youden's index (sensitivity + specificity – 1).

Results

At 152 years' follow-up, 76 events of MI and/or cardiac death had occurred. ML obtained a significantly higher AUC than the ASCVD risk and CAC score in predicting events (ML: 0.81; ASCVD: 0.76, p<0.05; CAC: 0.75, p<0.01, Figure A). ML performance was mostly driven by age, ASCVD risk and calcium as shown by the variable importance (Figure B); however, all variables with non-zero gain contributed to the ML performance. ML achieved a sensitivity and specificity of 77.6% and 73.5%, respectively. For an equal specificity, ASCVD and CAC scores obtained a sensitivity of 61.8% and 67.1%, respectively. High ML risk was associated with a high risk of suffering an event by Cox regression (HR: 9.25 [95% CI: 5.39–15.87], p<0.001; survival curves in Figure C). The relationships persisted when adjusted for age, gender, CAC, CAC percentile, aortic calcium score, and ASCVD risk score; with a hazard ratio of 3.42 for high ML risk (HR: 3.42 [95% CI: 1.54–7.57], p=0.002).

Conclusion(s)

Machine learning used to integrate clinical and quantitative imaging-based variables significantly improves prediction of MI and cardiac death in asymptomatic subjects undergoing CAC assessment, compared to standard risk assessment methods.

Acknowledgement/Funding

NHLBI 1R01HL13361, Bundesministerium für Bildung und Forschung (01EX1012B), Dr. Miriam and Sheldon G. Adelson Medical Research Foundation