CMR-Analysis (including machine learning)
Katherine Kearney, MD
Cardiology Fellow
St Vincent's Hospital, Australia
Katherine Kearney, MD
Cardiology Fellow
St Vincent's Hospital, Australia
Jim Pouliopoulos, PhD
Senior Postdoctoral Scientist
St Vincent's Hospital, Australia
Nicholas Olsen, PhD
Statistician
University of New South Wales, Australia
Jickson John
Senior Radiographer
St Vincent's Hospital, Australia
Audrey Adji, PhD
Senior Postdoctoral Scientist
St Vincent's Hospital, Sydney, Australia
Sara Hungerford, MD, PhD
Cardiologist
Royal North Shore Hospital, Australia
Cassia Kessler Iglesias
Cardiology Fellow
St Vincent's Hospital, Sydney, Australia
Nicole Bart
Cardiologist
St Vincent's Hospital, Sydney, New South Wales, Australia
Ning Song
Cardiology Fellow
St Vincent's Hospital, Sydney, Australia
Peter Macdonald
Senior Cardiologist
St Vincent's Hospital, Sydney, Australia
Kavitha Muthiah, MD, PhD
Cardiologist
St Vincent's Hospital, Sydney, Australia
Anne Keogh, MD
Senior Cardiologist
St Vincent's Hospital, Sydney, Australia
Christopher Hayward, MD, PhD
Senior Cardiologist
St Vincent's Hospital, Sydney, Australia
Justin Grenier
MR Research Assistant
University of Alberta, Canada
Richard Thompson, PhD
Professor of Biomedical Engineering
University of Alberta
Edmonton, Alberta, Canada
Jerome Yerly, PhD
Postdoc
University Hospital (CHUV) and University of Lausanne (UNIL), Vaud, Switzerland
Christopher W. Roy, PhD
Post Doc
Lausanne University Hospital and University of Lausanne (UNIL)
Prilly, Vaud, Switzerland
Mariana Falcao, MSc
PhD Candidate
Lausanne University Hospital and University of Lausanne
Lausanne, Vaud, Switzerland
Edmund MT Lau, MD, PhD
Respiratory Physician
Royal Prince Alfred Hospital, University of Sydney, Australia
Rajesh Puranik, MD, PhD
Senior Cardiologist
Royal Prince Alfred Hospital, University of Sydney, Australia
Martin Ugander, MD, PhD
Professor of Cardiac Imaging
University of Sydney
Eugene Kotlyar, MD
Senior Cardiologist
St Vincent's Hospital, Sydney, Australia
Andrew Jabbour, MD, PhD
Senior Cardiologist
St Vincent's Hospital, Sydney, England, Australia
Right heart catheterization (RHC) is the gold standard for assessing cardiac physiology. It would be valuable to have access to cardiovascular magnetic resonance (CMR) methods to non-invasively quantify invasive haemodynamic measures such as pulmonary vascular resistance (PVR), mean pulmonary arterial pressure (mPAP), and mean pulmonary arterial wedge pressure (mPAWP). This study aimed to derive these variables using CMR and supervised machine learning (SML).
Methods:
Patients undergoing RHC were recruited for same-day CMR imaging involving cine, 2D flow, respiratory-gated 4D flow (5D flow)1,2, T1/T2 mapping, and lung water density mapping3,4 (Figure 2). These measures were entered into multivariable models. The dataset was split into training and test subsets. Cross validation was used to tune models. Linear regression and decision-tree analyses using XGBoost were performed in R.
Results:
Patients underwent same-day RHC and CMR (n=100, age 57±17 years, 68% female). Pulmonary hypertension (PH) classification yielded precapillary (n=35), combined pre and postcapillary (n=16), postcapillary (n=20), and normal pressures (n=29). CMR-derived PVR (CMR-PVR) correlated well with RHC values (R2=0.80 in test subset, Table; mean difference 0.5±1.7 Wood Units). CMR-mPAP correlated well (R2=0.70 in test subset, Table, mean difference 0.4±6.0 mmHg). CMR-mPAWP had a limited correlation (R2=0.48 in test subset, mean difference 0.8±4.6 mmHg). Bland-Altman and predicted versus actual plots for all 3 are shown in Figure 1. Elevated PVR ( > 3 Woods Units) was correctly identified on CMR with an accuracy (ratio of correct predictions to total predictions) of 1.0 in both the training and test cohorts, with both sensitivity and specificity of 1.0. CMR was able to classify 3 subtypes of PH or normal pressures with an accuracy of 1.0 in the training cohort and 0.6 in the test cohort.
Conclusion: CMR-derived pulmonary hemodynamics measures utilising supervised machine learning algorithms showed good agreement with invasive RHC. CMR shows promise as a non-invasive alternative to traditional RHC.