Abstract 19631: Application of Machine Learning Methods for Prediction of Outcomes After Cardiac Transplantation: Insights From the UNOS Database
Objectives: It has been hypothesized that applying advanced analytic methodologies to large patient datasets can revolutionize patient care but the relative performance and comparative effectiveness of various statistical approaches remains unclear.
Methods: The United Network for Organ Sharing (UNOS) database was queried to identify initial adult heart transplants performed in the United States from 1987-2014 (N=50,453). We assessed prediction of 1 year survival using all donor and recipient variables except those with >20% missingness using traditional logistic regression and a comprehensive number of machine learning methodologies (ridge regression, regression with LASSO, support vector machines, naïve Bayesian, tree-augmented Bayesian, neural network, random forest, and stochastic gradient boosting). We tested the predictive accuracy of these methods across UNOS regions for the following time periods [Period 1 (1987-1996); Period 2 (1996-2005); Period 3 (2006-2014)].
Results: The most powerful univariate predictors were recipient age, BMI, renal function, liver function, albumin, presence of LVAD, and hemodynamics. Machine learning methodologies yielded AUCs ranging from 0.61-0.66. Neural networks yielded the highest predictive capability, with an AUC of 0.66 (Figure 1a). Predictive accuracy was highest in Period 3 (AUC 0.66) and differed significantly across UNOS regions (AUC 0.60-0.69; Figure1b).
Conclusions: The application of neural networks to the UNOS database of cardiac transplant patients yielded the most superior predictive capabilities. Use of advanced analytics could support clinical decision making for patients in need of cardiac transplantation and assist with more equitable allocation of donor organs.
Author Disclosures: B. Vaccaro: None. N.R. Desai: None. P. Rao: None. R. Ghosh: None. P. Warier: None. D.L. Jacoby: None. L. Bellumkonda: None. J.M. Testani: None. M.E. Chen: None. T. Ahmad: Speakers Bureau; Modest; Novartis. Consultant/Advisory Board; Modest; Relypsa. Research Grant; Significant; St Judes.
- © 2016 by American Heart Association, Inc.