← Selected papers

Machine learning: helpful or hype?

Bae et al., Transplant International, 2020.

A head-to-head evaluation of machine learning and conventional regression for transplant outcome prediction using the same clinical information.

C-statistics of three prediction methods across five transplant outcomes For each outcome, regression, gradient boosting and random forests fall within a narrow range: 0.717 to 0.723 for delayed graft function, 0.705 to 0.708 for death, 0.637 to 0.642 for death-censored graft failure, 0.633 to 0.635 for all-cause graft failure, and 0.579 to 0.611 for acute rejection, where regression was highest. Delayed graft function Death Death-censored graft failure All-cause graft failure Acute rejection 0.50.60.70.8 C-statistic (validation set)

Comparisons using the same registry variables found little predictive improvement from the machine learning methods evaluated. This work reflects an emphasis on testing what a method adds in a particular setting. It is a prediction study, rather than a causal inference study.