Machine Learning in Quantitative Stress Echocardiography

New onset chest pain is a common presenting complaint and can be a marker of significant cardiovascular disease and risk of myocardial infarction and death; therefore obtaining an accurate diagnosis is critical to guide patient management. It is noteworthy that only 40-50% of patients who have invasive arteriography subsequently undergo revascularisation. This underscores the imprecision of the initial tests employed prior to arteriography.

Historically electrocardiographic stress testing during exercise has been used to detect inducible myocardial ischaemia but its diagnostic sensitivity and specificity are low (about 65%). Diagnostic accuracy can be improved by by incorporating echocardiography or single photon emission computed tomography. Current NICE guidelines recommend that patients with chest pain of recent onset should be investigated with CT coronary angiography as a first line, and if this reveals a significant stenosis then a functional imaging test should be performed.

The Myocardial Doppler in Stress Echocardiography (MYDISE) study assessed the diagnostic value of quantitative stress echocardiography during the infusion of dobutamine, a short-acting synthetic catecholamine that acts on β-1 adrenergic receptors to increase heart rate and myocardial contractility. Measuring the systolic velocities of LV long-axis function at peak stress had good reproducibility (coefficients of variation in basal segments 9-14% at rest and 11-18% at peak stress) and similar sensitivities and specificities (about 70%) to published studies in which expert observers reported wall motion scores. When adjusted for the independent effects of age, gender and heart rate, however, diagnostic accuracy increased significantly with C statistics (area under receiver-operator curves) up to 90%.

Visual analysis of stress echocardiography to detect myocardial ischaemia depends on qualitative assessment of multiple parameters. Major studies of quantitative stress echocardiography have been limited to identifying the single best echocardiographic variable, and they have used diameter stenosis as the reference criterion. Progressive subclinical reductions of regional (long-axis) myocardial function have been demonstrated in subjects with cardiovascular risk factors, affecting myocardial deformation (strain and strain rate) as well as velocities. Ischaemia changes the timing of events during the cardiac cycle - for example prolonging pre-ejection and post-ejection phases. These factors confirm the clinical need for objective measurement of regional myocardial function throughout the cardiac cycle.

It is now possible to create algorithms that are based not just on a single time point (e.g., peak velocity) but instead rely on analysis of the whole of the velocity trace. This concept can also be extended to include strain and strain rate curves. Investigators at Universitat Pompeu Fabra, Barcelona, have developed this approach to create a statistical atlas of the heart to detect dyssynchrony. A similar concept has been applied using multiple kernel learning to patients with dyspnoea who have undergone exercise stress testing to identify those with evidence of diastolic heart failure.This has enabled velocity traces taken from the whole of the cardiac cycle to be compared and discriminated between control subjects (with and without dyspnoea) and those diagnosed with heart failure with preserved ejection fraction (HFpEF); the major differences observed are in early diastolic function. This application has not previously been used to explore inducible myocardial ischaemia in stress echocardiography, but similar findings might be expected, as changes during diastole are amongst the earliest and most sensitive indicators of myocardial ischaemia. Individuals at the University of Leuven (Prof Jan D'hooge) have recently developed supervised machine-learning methods that allow for automatic classification of myocardial segments based on their local mechanical behaviour (i.e. velocity/strain/strain rate) after going through a training phase; the proposed machine-learning approach outperforms expert wall motion readings as well as expert interpretation of segmental strain (rate) traces in classifying ischemic segments.