Machine 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

Abstract

Aims: Our aim was to evaluate the performance of machine learning (ML), integrating clinical parameters with coronary artery calcium (CAC), and automated epicardial adipose tissue (EAT) quantification, for the prediction of long-term risk of myocardial infarction (MI) and cardiac death in asymptomatic subjects.

Methods and results: Our study included 1912 asymptomatic subjects [1117 (58.4%) male, age: 55.8 ± 9.1 years] from the prospective EISNER trial with long-term follow-up after CAC scoring. EAT volume and density were quantified using a fully automated deep learning method. ML extreme gradient boosting was trained using clinical co-variates, plasma lipid panel measurements, risk factors, CAC, aortic calcium, and automated EAT measures, and validated using repeated 10-fold cross validation. During mean follow-up of 14.5 ± 2 years, 76 events of MI and/or cardiac death occurred. ML obtained a significantly higher AUC than atherosclerotic cardiovascular disease (ASCVD) risk and CAC score for predicting events (ML: 0.82; ASCVD: 0.77; CAC: 0.77, P < 0.05 for all). Subjects with a higher ML score (by Youden's index) had high hazard of suffering events (HR: 10.38, P < 0.001); the relationships persisted in multivariable analysis including ASCVD-risk and CAC measures (HR: 2.94, P = 0.005). Age, ASCVD-risk, and CAC were prognostically important for both genders. Systolic blood pressure was more important than cholesterol in women, and the opposite in men.

Conclusions: In this prospective study, machine learning used to integrate clinical and quantitative imaging-based variables significantly improves prediction of MI and cardiac death compared with standard clinical risk assessment. Following further validation, such a personalized paradigm could potentially be used to improve cardiovascular risk assessment.

Keywords: Coronary calcium scoring; Epicardial adipose tissue; Machine learning; Myocardial infarction and cardiac death.

Published on behalf of the European Society of Cardiology. All rights reserved. © The Author(s) 2019. For permissions, please email: journals.permissions@oup.com.

Figures

Figure 1

Variable importance for the classification of MI and/or cardiac death (in 1912 subjects). The gain denotes how much a feature contributes to the prediction made by the XGB algorithm. RF, risk factor.

Figure 2

Receiver-operating characteristic curves for the prediction of MI and/or cardiac death (1912 subjects). XGB performed significantly better than the ASCVD risk score (P < 0.05) and the CAC scoring (P < 0.01). The straight line illustrates the sensitivity, at the same specificity as XGB at maximum Youden’s index. At this specificity, the sensitivity was 0.816 for XGB, 0.724 for CAC, and 0.645 for ASCVD risk score.

Figure 3

Calibration plots of predicted scores for MI and/or cardiac death prediction (N = 1912). (A) ML scores in red. (B) ASCVD risk scores in blue. The lines and dots represent the average prediction scores per decile. Bars show the average percentage of observed events per decile.

Figure 4

Kaplan–Meier estimator for population (N = 1912) with low (blue) and high machine learning risk (red).

Figure 5

Variable importance in ML classification for women (A, N = 795) and men (B, N = 1117). Kaplan–Meier curves for subjects with high (red) and low ML risk (blue) in women (C) and (D).

Figure 6

Explanations of individual prediction with subject-specific feature importance for a 62-year-old man with no event (A) and a 74-year-old woman with an event at 8.79 years (B). The x axis corresponds to the ML risk score. The arrows represent the influence of each covariate on the overall prediction; blue and red arrows indicate whether the covariates decrease (blue) or increase (red) the risk of future events. The combination of all covariates’ influence provides the final ML risk score. For the male subject in (A), there is a low ML risk score (0.0103), with an ASCVD risk score of 7.25%, CAC score of 0 and 0th CAC percentile. For the female subject in (B), there is a high ML risk score (0.2231), with an ASCVD risk score of 30.36%, a CAC score of 94 and 63rd CAC percentile. The red and blue colours provide the separation between low and high ML risk by optimal Youden’s index and grey dashed lines correspond to the base risk obtained from the rate of event in the population (∼4%).