Osvald Nitski1, Amirhossein Azhie2, Fakhar Ali Qazi-Arisar3, Xueqi Wang4, Shihao Ma5, Leslie Lilly6, Kymberly D Watt7, Josh Levitsky8, Sumeet K Asrani9, Douglas S Lee10, Barry B Rubin11, Mamatha Bhat12, Bo Wang13. 1. Faculty of Applied Science and Engineering, University of Toronto, Toronto, ON, Canada; Peter Munk Cardiac Centre, University Health Network, Toronto, ON, Canada. 2. Multi Organ Transplant Program, University Health Network, Toronto, ON, Canada. 3. Division of Gastroenterology & Hepatology, Department of Medicine, University of Toronto, Toronto, ON, Canada; Multi Organ Transplant Program, University Health Network, Toronto, ON, Canada. 4. Peter Munk Cardiac Centre, University Health Network, Toronto, ON, Canada. 5. Department of Computer Science, University of Toronto, Toronto, ON, Canada. 6. Faculty of Medicine, University of Toronto, Toronto, ON, Canada; Multi Organ Transplant Program, University Health Network, Toronto, ON, Canada. 7. Division of Gastroenterology & Hepatology, Mayo Clinic, Rochester, MN, USA. 8. Northwestern University Feinberg School of Medicine, Chicago, IL, USA. 9. Baylor University Medical Center, Dallas, TX, USA. 10. Faculty of Medicine, University of Toronto, Toronto, ON, Canada; Peter Munk Cardiac Centre, University Health Network, Toronto, ON, Canada; Ted Rogers Centre for Heart Research, Toronto, ON, Canada; Institute for Clinical Evaluative Sciences, Toronto, ON, Canada. 11. Faculty of Medicine, University of Toronto, Toronto, ON, Canada; Peter Munk Cardiac Centre, University Health Network, Toronto, ON, Canada. 12. Faculty of Medicine, University of Toronto, Toronto, ON, Canada; Division of Gastroenterology & Hepatology, Department of Medicine, University of Toronto, Toronto, ON, Canada; Institute of Medical Sciences, University of Toronto, Toronto, ON, Canada; Multi Organ Transplant Program, University Health Network, Toronto, ON, Canada; Toronto General Hospital Research Institute, Toronto, ON, Canada. 13. Department of Computer Science, University of Toronto, Toronto, ON, Canada; Department of Laboratory Medicine and Pathobiology, University of Toronto, Toronto, ON, Canada; Peter Munk Cardiac Centre, University Health Network, Toronto, ON, Canada; Vector Institute, Toronto, ON, Canada. Electronic address: bo.wang@uhnresearch.ca.
Abstract
BACKGROUND: Survival of liver transplant recipients beyond 1 year since transplantation is compromised by an increased risk of cancer, cardiovascular events, infection, and graft failure. Few clinical tools are available to identify patients at risk of these complications, which would flag them for screening tests and potentially life-saving interventions. In this retrospective analysis, we aimed to assess the ability of deep learning algorithms of longitudinal data from two prospective cohorts to predict complications resulting in death after liver transplantation over multiple timeframes, compared with logistic regression models. METHODS: In this machine learning analysis, model development was done on a set of 42 146 liver transplant recipients (mean age 48·6 years [SD 17·3]; 17 196 [40·8%] women) from the Scientific Registry of Transplant Recipients (SRTR) in the USA. Transferability of the model was further evaluated by fine-tuning on a dataset from the University Health Network (UHN) in Canada (n=3269; mean age 52·5 years [11·1]; 1079 [33·0%] women). The primary outcome was cause of death, as recorded in the databases, due to cardiovascular causes, infection, graft failure, or cancer, within 1 year and 5 years of each follow-up examination after transplantation. We compared the performance of four deep learning models against logistic regression, assessing performance using the area under the receiver operating characteristic curve (AUROC). FINDINGS: In both datasets, deep learning models outperformed logistic regression, with the Transformer model achieving the highest AUROCs in both datasets (p<0·0001). The AUROC for the Transformer model across all outcomes in the SRTR dataset was 0·804 (99% CI 0·795-0·854) for 1-year predictions and 0·733 (0·729-0·769) for 5-year predictions. In the UHN dataset, the AUROC for the top-performing deep learning model was 0·807 (0·795-0·842) for 1-year predictions and 0·722 (0·705-0·764) for 5-year predictions. AUROCs ranged from 0·695 (0·680-0·713) for prediction of death from infection within 5 years to 0·859 (0·847-0·871) for prediction of death by graft failure within 1 year. INTERPRETATION: Deep learning algorithms can incorporate longitudinal information to continuously predict long-term outcomes after liver transplantation, outperforming logistic regression models. Physicians could use these algorithms at routine follow-up visits to identify liver transplant recipients at risk for adverse outcomes and prevent these complications by modifying management based on ranked features. FUNDING: Canadian Donation and Transplant Research Program, CIFAR AI Chairs Program.
BACKGROUND: Survival of liver transplant recipients beyond 1 year since transplantation is compromised by an increased risk of cancer, cardiovascular events, infection, and graft failure. Few clinical tools are available to identify patients at risk of these complications, which would flag them for screening tests and potentially life-saving interventions. In this retrospective analysis, we aimed to assess the ability of deep learning algorithms of longitudinal data from two prospective cohorts to predict complications resulting in death after liver transplantation over multiple timeframes, compared with logistic regression models. METHODS: In this machine learning analysis, model development was done on a set of 42 146 liver transplant recipients (mean age 48·6 years [SD 17·3]; 17 196 [40·8%] women) from the Scientific Registry of Transplant Recipients (SRTR) in the USA. Transferability of the model was further evaluated by fine-tuning on a dataset from the University Health Network (UHN) in Canada (n=3269; mean age 52·5 years [11·1]; 1079 [33·0%] women). The primary outcome was cause of death, as recorded in the databases, due to cardiovascular causes, infection, graft failure, or cancer, within 1 year and 5 years of each follow-up examination after transplantation. We compared the performance of four deep learning models against logistic regression, assessing performance using the area under the receiver operating characteristic curve (AUROC). FINDINGS: In both datasets, deep learning models outperformed logistic regression, with the Transformer model achieving the highest AUROCs in both datasets (p<0·0001). The AUROC for the Transformer model across all outcomes in the SRTR dataset was 0·804 (99% CI 0·795-0·854) for 1-year predictions and 0·733 (0·729-0·769) for 5-year predictions. In the UHN dataset, the AUROC for the top-performing deep learning model was 0·807 (0·795-0·842) for 1-year predictions and 0·722 (0·705-0·764) for 5-year predictions. AUROCs ranged from 0·695 (0·680-0·713) for prediction of death from infection within 5 years to 0·859 (0·847-0·871) for prediction of death by graft failure within 1 year. INTERPRETATION:Deep learning algorithms can incorporate longitudinal information to continuously predict long-term outcomes after liver transplantation, outperforming logistic regression models. Physicians could use these algorithms at routine follow-up visits to identify liver transplant recipients at risk for adverse outcomes and prevent these complications by modifying management based on ranked features. FUNDING: Canadian Donation and Transplant Research Program, CIFAR AI Chairs Program.