Abstract: TH-PO259
Dynamic Predictions of All-Cause Mortality in Incident Hemodialysis Patients with Extended Joint Models Adjusting for Albumin Trajectory and Competing Risks
Session Information
- Hemodialysis: Volume, Metabolic Complications, Clinical Outcomes
November 02, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 801 Dialysis: Hemodialysis and Frequent Dialysis
Authors
- Damgov, Ivan, Center for Pediatric and Adolescent Medicine, University of Heidelberg, Heidelberg, Germany
- Kieser, Meinhard, Institute of Medical Biometry, University of Heidelberg, Heidelberg, Germany
- Rutherford, Peter, Baxter Healthcare Corporation, Zurich, Switzerland
- Wong, Muh Geot, Department of Renal Medicine, Royal North Shore Hospital, St Leonards, New South Wales, Australia
- Pollock, Carol A., Kolling Institute, Sydney Medical School, University of Sydney, Royal North Shore Hospital, St Leonards, New South Wales, Australia
- Johnson, David W., Australasian Kidney Trials Network, University of Queensland, Brisbane, Queensland, Australia
- Schmitt, Claus Peter, Center for Pediatric and Adolescent Medicine, University of Heidelberg, Heidelberg, Germany
Background
Joint models (JM) provide individual patient dynamic risk predictions based on the association between longitudinal biomarkers and mortality. We explore extensions to JM methodology performed for the first time with hemodialysis (HD) clinical trial data.
Methods
We evaluated the association of albumin with all-cause mortality in incident HD patients followed over 8 years in the Initiating Dialysis Early And Late trial. To account for skewness/outliers in albumin, novel JM with mixed effects sub-model with skew-normal, t and skew-t errors were compared with conventional JM under normal distribution. Furthermore, we adjusted for the competing event of transplantation using Weibull cause-specific hazards. The dataset (N=315 patients) was randomly split into training dataset (3/4, N=236) for model fitting and validation dataset (1/4, N=78) to infer predictive performance. We obtained 12-month individual patient survival predictions from all JM at cut-offs 1, 1.5 and 2 years utilizing albumin trajectory and baseline risk factors.
Results
A median of 12 albumin records per patient with 24% and 19% event rates for death and transplantation ensured convergence of all 12 joint models. Albumin hazard ratio of death remained robust (0.89 to 0.90), suggesting significant inverse relationship between albumin and death adjusted for baseline confounders and the competing event of transplantation.
JM presented excellent survival prediction accuracy in the validation dataset which improved with longer follow-up. JM with competing event outperformed JM with survival only (area under the curve (AUC) ranges at 1, 1.5 and 2 years: 0.80-0.84 vs 0.79-0.83, 0.86-0.90 vs 0.84-0.88, 0.96-0.97 vs 0.96-0.97), across all distributional models for albumin. Prediction performance of all JM surpassed classical Cox model with baseline albumin (AUC = 0.71, 0.77, 0.93). Results were confirmed by a simulation study.
Conclusion
Our first comprehensive study of JM extensions in incident HD patients indicates its great potential for dynamic personalized survival prediction. Robustness of estimates to deviations from normality and higher predictive accuracy when adjusting competing events contribute to JM superiority over the classical Cox approach.
Funding
- Commercial Support – Baxter Healthcare Corporation