Machine Learning and Artificial Intelligence

1355640 - Explainable Deep Learning-Based Prediction of Future Cardiovascular Events Using Multi-Domain Phenotypes Captured by Standardized Cardiovascular MRI Workflow

Thursday, January 26, 2023

1:00 PM – 2:00 PM

Location: Pacific Jewel

Presenting Author(s)

Dina Labib, MD, PhD, FSCMR

PhD student
Libin Cardiovascular Institute of Alberta, University of Calgary
Calgary, Alberta, Canada

Primary Author(s)

Dina Labib, MD, PhD, FSCMR

PhD student
Libin Cardiovascular Institute of Alberta, University of Calgary
Calgary, Alberta, Canada

Co-Author(s)

SD

Steven Dykstra, MSc, BSc

PhD Student
Libin Cardiovascular Institute of Alberta, University of Calgary
Calgary, Alberta, Canada
LL

Lucy Y. Lei, MSc

Medical student
Libin Cardiovascular Institute of Alberta, University of Calgary; Cumming School of Medicine
Calgary, Alberta, Canada
AC

Aidan Cornhill, BMSc

Master's student
Libin Cardiovascular Institute of Alberta, University of Calgary, Alberta, Canada
JF

Jacqueline Flewitt, MSc

Research Collaborations Coordinator
Libin Cardiovascular Institute of Alberta, University of Calgary
Calgary, Alberta, Canada
SR

Sandra Rivest, RN

Research nurse
Libin Cardiovascular Institute of Alberta, University of Calgary
Calgary, Alberta, Canada
PF

Patricia Feuchter, MSc, MRT

MRI Research Technologist
Libin Cardiovascular Institute of Alberta, University of Calgary, Canada
AH

Andrew G. Howarth, MD, PhD

Clinical Co-director
Libin Cardiovascular Institute of Alberta, University of Calgary
Calgary, Alberta, Canada
CL

Carmen P. Lydell, MD

Clinical Co-director
Libin Cardiovascular Institute of Alberta, University of Calgary
Calgary, Alberta, Canada
MB

Michael S. Bristow, MD

Clinical Assistant Professor
Libin Cardiovascular Institute of Alberta, University of Calgary, Alberta, Canada
EP

Edith Pituskin, PhD

Associate Professor
University of Alberta
Edmonton, Alberta, Canada
WC

Winson Y Y. Cheung, MD, MPH

Professor; Director Health Services Research, Cancer Control Alberta
University of Calgary, Canada
JL

Joon Lee, PhD

Associate Professor of Health Data Science; Director of the Data Intelligence for Health Lab
Libin Cardiovascular Institute of Alberta, Cumming School of Medicine, University of Calgary, Canada
James White, MD

Professor of Cardiology
Stephenson Cardiac Imaging Centre
Calgary, Alberta, Canada

Background: Despite prognostic value recognized for patient-reported health (PRH), electronic health record (EHR), and cardiac imaging phenotypic markers, the respective contribution of these features to deliver deep learning predictions of cardiovascular events has not been explored. Such approaches have shown early promise in other fields to deliver accurate patient-specific predictions; however, are considered black-box models that limit their acceptance for clinical decision making. We sought to explore the unique and combined value of data resources secured by standardized CMR service workflow for the delivery of explainable deep learning-based predictions of major adverse cardiac events (MACE).

Methods: A total of 9,858 subjects enrolled in the CIROC Registry with ≥6 months follow-up were used. Subjects had relevant data captured by workflow automation software (cardioDI, Cohesic Inc) using tablet-based health questionnaires followed by standardized reporting. Linkage to administrative databases captured laboratory, pharmacy, and ICD-10 coded clinical data. Patients were followed for the composite MACE endpoint of all-cause mortality, survived cardiac arrest, heart failure admission, acute coronary syndrome, stroke, or revascularization. A total of 90 variables available across all referral populations were entered into 7 deep learning survival models (DeepSurv): one for each isolated data domain and 4 for all possible combinations. The dataset was split into training (80%) and test (20%) sets, with 20% of the training set used to validate models during training. Performance of trained models was compared using C-index and integrated Brier score (IBS) in the test set. Shapley values, a measure of additive feature contribution to model prediction, were calculated to provide explainability.

Results: Over a median of 2.9 years, a total of 2,767 patients (28%) experienced the primary outcome. Patient characteristics and event rates are shown in Table 1 and were similar between the training and test sets. Performance metrics for the 7 models are shown in Table 2. The full model, exposed to all data domains, provided best performance (highest C-index of 0.83 and lowest IBS of 0.13). Strong performance was achievable using deep learning of CMR plus EHR data domains (C-index 0.82 and IBS 0.13). Figure 1 shows top performing predictors in the full model with each variable’s observed value plotted according to its influence (magnitude and direction) on final model prediction by estimated Shapley values.

Conclusion: This is the largest study to date examining the unique and combined predictive value of PRH, EHR and CMR data for the prediction of MACE by deep learning-based techniques, and highlights capacity for integrated model explainability. Each data domain was capable of delivering strong value toward MACE prediction. However, best performance was achieved through their combined access.