1355503 - Machine Learning Analysis of Oxygenation-Sensitive Cardiovascular Magnetic Resonance Imaging in Patients with Coronary Artery Stenosis

Oxygenation-Sensitive Cardiovascular Magnetic Resonance (OS-CMR) has emerged as a powerful tool to monitor dynamic changes in myocardial oxygenation as a novel biomarker, specifically in patients with suspected coronary artery disease or microvascular dysfunction.^1,2 The large number of biomarkers available, however, complicates the data analysis. For such large data sets, radiomics algorithms can provide superior diagnostic accuracy over standard analytical and reporting methods.³ The aim of this study was to develop and evaluate a machine learning (ML)-based model for OS-CMR in patients with coronary artery stenosis (CAS).

Methods:

The data used for the ML-based model was derived from two short axis slices of OS-CMR studies performed on a 3T MRI system, in patients scheduled for clinically indicated coronary angiography for suspected coronary artery disease (CAD). We performed 3 different analyses to assess diagnostic accuracy. The first two analyses evaluated the accuracy of OS-CMR to classify patients as having significant CAS (defined as >=80% stenosis in any coronary artery). The third analysis assessed the diagnostic accuracy of OS-CMR to classify patients recommended for revascularization (either surgical or percutaneous). The neural network dataset was divided into combinations of 8 biomarkers to determine the predictability of each biomarker in classifying patients into their appropriate groups. Finally, we selected biomarkers with over 90% classification accuracy in 100 runs to identify the most predictive biomarkers.

Results:

In total, 26 patients (mean age 63 +/- 10 years, 85% males) and 21 healthy subjects (mean age 39+/-16 years, 52% males) were included, with 19 (73%) having significant stenosis, and 22 (85%) recommended for revascularization. All participants completed the scans, and the average scan duration was 28.3 +/- 5.6 mins. Over 3,300 discrete data points per participant were extracted and used in our model. The model showed excellent accuracy at differentiating healthy volunteers from patients with any degree of stenosis and at distinguishing patients with significant stenosis (Acc=0.906), as well as classifying patients recommended for revascularization (Acc=0.812). The most predictive biomarker was signal intensity during end-diastole normalized to the series.

Conclusion:

In this proof-of-concept study, a fully automated ML model has demonstrated the potential to classify patients with significant coronary artery stenoses recommended for revascularization using an entirely needle- and stress-free OS-CMR approach. Larger datasets, further training, and refinement of the algorithm will further enhance the clinical application of OS-CMR.