CORAL: a machine learning algorithm to identify coronary microvascular disease from doppler coronary blood flow patterns

The direct diagnosis of coronary microvascular disease (CMD) in patients has remained elusive, relying on combinatorial criteria for diagnosis. Recent data in mice suggested that coronary blood flow (CBF) patterns harbor “fingerprints” that may be novel non-invasive indicators of CMD. The goal of this study was to test the hypothesis that machine learning (ML) of CBF patterns can directly and accurately identify CMD. Training and test data for ML were obtained from control and diabetic db/db mice at early (12 wks), established (16 wks), and senescent (36 wks) ages with CMD (n = 460). Independent test data were obtained in a western-diet-fed (WD) mouse model of CMD (n = 40). All mice were evaluated for CBF and cardiac function by echocardiography and pressure-volume (PV) loop catheterization. CBF images were processed and input into CORAL (Coronary Risk Assessment with Learning), a ML classifier with optional input of cardiac function features via a multilayer perceptron. CBF pattern analysis using CORAL yielded remarkable accuracy of 100% at 12 weeks of age – early in CMD. CORAL correctly identified established CMD with 81% and 83% accuracy in the test and independent test data sets, respectively. The inclusion of cardiac function increased accuracy up to 89% and 90% in the test and independent test data, respectively. This study demonstrates a novel ability of CORAL to unlock previously unutilized features of the CBF pattern that can accurately identify CMD early in disease progression. We envision further refinements to CORAL could lead to direct accurate CMD diagnoses in patients. The direct diagnosis of coronary microvascular disease remains elusive. It’s currently done using a combination of invasive and non-invasive methods. Can machine learning of Doppler coronary blood flow patterns directly and accurately identify coronary microvascular disease? Using comprehensive in vivo data sets in mice (n ~ 500 total), we developed a novel machine learning neural network called CORAL that can identify coronary microvascular disease with remarkable accuracy (81-100%, depending on timing of CMD) directly from non-invasive coronary blood flow patterns obtained by transthoracic Doppler echocardiography. Semi-simultaneous measures of non-invasive and invasive cardiac function improved prediction of coronary flow pattern-based coronary microvascular disease (up to 100%, depending on timing of CMD). These data suggest that coronary blood flow patterns harbor useful information that is predictive of coronary microvascular disease with a high degree of accuracy. Doppler echocardiographic evaluation of coronary flow in patients with suspected coronary microvascular disease could lead to more accurate and definitive diagnosis, particularly at early stages of disease.