Machine Learning Model Using Pre-Cancer Therapy Cardiac Magnetic Resonance Images to Predict Cancer Therapy-Related Cardiac Dysfunction.
Authors
Affiliations (10)
Affiliations (10)
- Department of Medicine, Division of Cardiology, Ted Rogers Program in Cardiotoxicity Prevention, Peter Munk Cardiac Centre, Toronto General Hospital, University Health Network, University of Toronto, Toronto, Ontario, Canada.
- Ted Rogers Computational Program, Ted Rogers Centre for Heart Research, Peter Munk Cardiac Centre, University Health Network, University of Toronto, Toronto, Ontario, Canada.
- Stephenson Cardiac Imaging Centre, Libin Cardiovascular Institute of Alberta, University of Calgary, Alberta, Canada.
- Sunnybrook Health Science Centre, University of Toronto, Toronto, Ontario, Canada.
- Princess Margaret Cancer Centre, University of Toronto, Toronto, Ontario, Canada.
- Joint Department of Medical Imaging, Toronto General Hospital, University Health Network, University of Toronto, Toronto, Ontario, Canada.
- Joint Department of Medical Imaging, Toronto General Hospital, University Health Network, University of Toronto, Toronto, Ontario, Canada; Department of Radiology, LMU University Hospital, LMU Munich, Munich, Germany.
- Department of Medicine, Division of Cardiology, Ted Rogers Program in Cardiotoxicity Prevention, Peter Munk Cardiac Centre, Toronto General Hospital, University Health Network, University of Toronto, Toronto, Ontario, Canada; Women's College Hospital, Toronto, Ontario, Canada.
- Ted Rogers Computational Program, Ted Rogers Centre for Heart Research, Peter Munk Cardiac Centre, University Health Network, University of Toronto, Toronto, Ontario, Canada; Joint Department of Medical Imaging, Toronto General Hospital, University Health Network, University of Toronto, Toronto, Ontario, Canada; Princess Margaret Cancer Centre, Toronto, Ontario, Canada.
- Department of Medicine, Division of Cardiology, Ted Rogers Program in Cardiotoxicity Prevention, Peter Munk Cardiac Centre, Toronto General Hospital, University Health Network, University of Toronto, Toronto, Ontario, Canada; Joint Department of Medical Imaging, Toronto General Hospital, University Health Network, University of Toronto, Toronto, Ontario, Canada. Electronic address: [email protected].
Abstract
Predicting risk of cancer therapy-related cardiac dysfunction (CTRCD) remains challenging. The purpose of this study was to assess if deep learning (DL) approaches using cardiac magnetic resonance (CMR) images before cancer therapy can predict subsequent CTRCD and compare them with clinical and conventional imaging models. Women with HER2+ breast cancer receiving anthracyclines and trastuzumab from 3 prospective studies (Toronto, Canada: EMBRACE-MRI, SPARE-HF; and Calgary, Canada: CIROC) were included. Patients were assessed before cancer therapy, after anthracycline, and trimonthly with repeated echocardiography and CMR. CTRCD was defined according to CMR. We calculated the HFA-ICOS risk score and measured CMR and echocardiography volumetric and functional parameters. Deep convolutional neural network architectures were used with pre-cancer therapy CMR short-axis cine images to develop image-based DL models to predict CTRCD. The Toronto patients were used for model derivation and internal validation; the Calgary patients were used for external validation. A total of 229 patients (mean age 50.4 ± 9.7 years) were included: 176 in the internal (52 CTRCD events) development cohort and 53 in the external data set (14 CTRCD events). Our pre-cancer therapy CMR DL model demonstrated an AUC of 0.85 (95% CI: 0.69-0.97) and F1 score 0.69 (95% CI: 0.47-0.86) to predict CTRCD. On external validation, the DL model had an AUC of 0.80 (95% CI: 0.58-0.86) and F1 score of 0.55 (95% CI: 0.32-0.69). In comparison, the best-performing models using HFA-ICOS risk score, CMR, and echocardiographic parameters demonstrated, respectively, AUCs of 0.66 (95% CI: 0.53-0.75), 0.59 (95% CI: 0.41-0.73), 0.62 (95% CI: 0.44-0.78) and F1 scores of 0.56 (95% CI: 0.47-0.61), 0.20 (95% CI: 0.00-0.42) and 0.36 (95% CI: 0.12-0.62) to predict CTRCD. A DL model using pre-cancer therapy CMR short-axis cines can predict future CTRCD risk better than clinical or manually quantified imaging models.