Fully automated, deep learning, cardiac CT-based multimodal network for cardiovascular risk stratification in high-risk perioperative patients.
Authors
Affiliations (20)
Affiliations (20)
- Medical School, The University of Western Australia, Perth, Australia.
- Harry Perkins Institute of Medical Research, Perth, Australia.
- Department of Computer Science and Software Engineering, The University of Western Australia, Perth, Australia.
- Department of Cardiology, Fiona Stanley Hospital, Perth, Australia.
- Artrya Ltd, Perth, Australia.
- School of Engineering, Computer and Mathematical Sciences, Auckland University of Technology, Auckland, New Zealand.
- Department of Cardiology, Kolling Institute and Charles Perkins Centre, University of Sydney, Sydney, Australia.
- Department of Cardiology, Royal North Shore Hospital, Sydney, Australia.
- Department of Anesthesia and Intensive Care, The Chinese University of Hong Kong, Hong Kong SAR, China.
- Division of Cardiology, University of Alberta Hospital, Edmonton, Alberta, Canada.
- Department of Medicine, McMaster University, Hamilton, Canada.
- Department of Anaesthesia & Critical Care, The University of Chicago, Chicago, USA.
- Department of Radiology, Washington University School of Medicine, St. Louis, MO 63110, USA.
- Department of Medicine, Schulich School of Medicine & Dentistry, Western University, London, Ontario, Canada.
- Department of Intensive Care and Perioperative Medicine, Jagiellonian University Medical College, KrakÏŒw, Poland.
- Departments of Radiology and Medicine, Faculty of Medicine, University of Malaya, Kuala Lumpur, Malaysia.
- Department of Anaesthesia and Perioperative Medicine, Groote Schuur Hospital, Faculty of Health Sciences, University of Cape Town, Cape Town, South Africa.
- Department of Medicine, University of Edinburgh, Edinburgh, UK.
- Department of Medicine (Cardiology), University of Ottawa Heart Institute, Ottawa, Canada.
- Department of Cardiology, Victor Chang Cardiac Research Institute, Crawley, WA 6009, Australia.
Abstract
Major adverse cardiac events (MACE) significantly impact perioperative morbidity and mortality. We aimed to develop a fully automated multimodal deep learning (DL) system integrating patient demographics, comorbidities, and coronary computed tomography angiography (CCTA) findings to optimize risk prediction. We included 639 patients undergoing CCTA as part of perioperative risk assessment for elective non-cardiac surgery. Convolutional neural networks automatically identified coronary artery disease reporting and data system (CAD-RADS) scores and segmented the left ventricle, aorta, and heart. These imaging features were combined with patient demographics and comorbidities to predict MACE risk. We evaluated the performance of our multimodal model against the revised cardiac risk index (RCRI) using gradient boosting decision tree modelling and area under the receiver operating characteristic (AUROC) curves. Among 639 patients (mean age 70 ± 9 years, 56% males, median RCRI 1), 61% underwent orthopaedic surgery, 27% vascular surgery and the rest abdominal/pelvic or spine surgery. 45 patients experienced MACE within 30 days. Automated CAD-RADS (AUROC = 0.69) demonstrated comparable performance to human analysis (AUROC = 0.67, <i>P</i> = 0.77). The multimodal DL system (AUROC = 0.82) outperformed CAD-RADS (delta-AUROC = 0.13, CI: 0.02, 0.26, <i>P</i> = 0.02), and RCRI (delta-AUROC =0.22, CI: 0.05, 0.46; <i>P</i> = 0.001) in predicting MACE and demonstrated robust sensitivity (83%) and specificity (79%). Our multimodal system built using automated CAD-RADS, anatomical segmentations and patient demographics outperforms both human expert and automated CAD-RADS for MACE prediction. This approach has the potential to enhance patient outcomes by leveraging the synergy between automated imaging and clinical data.