Bayesian Uncertainty Estimation Reduces Confident Medical Misdiagnoses to 2.7%

Applying Monte Carlo dropout to a multi-task chest-radiograph classifier improves error-detection AUROC from 0.74 to 0.77. Presenting uncertainty as a binary error-risk flag to clinical agents reduced confident misdiagnoses on unreliable findings from 8.5% to 2.7%.