This abstract is published in the Journal of the American Society of Echocardiography (volume 33, number 6, year 2020), as part of the proceedings of the American Society Echocardiography Conference 2020.
Title: Coronary Artery Disease Prediction from Resting Echocardiograms using Novel Imaging Biomarkers
Authors: Porumb, Mihaela and Mumith, Angela and Gao, Shan and Parker, Andrew and Beqiri, Arian and Upton, Ross and Woodward, William and Dockerill, Cameron and McCourt, Annabelle and Woodward, Gary and others
BACKGROUND: Stress echocardiography is usually considered necessary to identify patients with prognostically significant coronary disease (CAD) and relies on operator expertise to visualise wall motion abnormalities. We hypothesised that patients with CAD may already have underlying changes in myocardial function at rest and therefore be identiable using AI trained models built from novel functional biomarkers, extracted from the rest images.
METHODS: Different machine learning methods were employed to train a binary CAD risk prediction classifier, using features derived from the left ventricle quantification of resting echocardiograms. The machine learning models were trained using two and four chamber apical views, as well as the short-axis view using data obtained from patients referred for investigation of chest pain within a large multisite trial. The classifier was trained to predict the risk of a severe CAD defined as evidence of obstructive CAD (> 70% stenosis or <70% with abnormal FFR) or acute coronary event, revascularisation or death due to cardiac events.
RESULTS: The training dataset consisted of 579 studies of whom 86 had clinically proven prognostically significant CAD. The clinical characteristics of the patients are presented in Table 1. Four machine learning models were trained for the CAD prediction, as well as ensembles of these classifiers, using a soft-voting for the final outcome prediction. To test the performance of the binary risk prediction models repeated nested cross-validation was carried out with hyperparameter optimisation and feature selection, considering an initial 3318 features. The best performance on the training dataset was obtained when training the ensemble of the four classifiers, on a reduced number of features (i.e. 43), that were selected during cross-validation. The training sensitivity and specificity was 0.912 and 0.640, respectively.
CONCLUSION: This study shows promising results for automated identification of patients with prognostically significant CAD using only resting echocardiograms. Further work is required on validation to determine optimal thresholds for diagnostic use but the widespread availability of echocardiography at point of care means this approach may have broad clinical application.
Table 1 – Clinical and echocardiographic characteristics, stratified by the prediction outcome
|Training dataset||Training dataset|
|Positive CAD||Negative CAD|
|Age, years||67.2 (±10.3)||64.4 (±11.4)|
|Male, n (%)||59 (68.6)||255 (51.7)|
|BMI, kg/m2||27.9 (±3.7)||29.9 (±6.1)|
|Height, m||1.7 (±0.08)||1.6 (±0.1)|
|Previous CAD, n (%)||48 (55.8)||184 (37.3)|
|Resting regional wall motion abnormalities, n (%)||39 (44.1)||58 (11.7)|
Data are presented as mean ± SD for the continuous variables