Deep neural networks are increasingly being used for computer-aided diagnosis, but erroneous diagnoses can be extremely costly for patients. Researchers have developed a learning to defer with uncertainty (LDU) algorithm which considers the diagnostic network’s predictive uncertainty when learning which patients to diagnose automatically and which patients to defer to human experts.
The algorithm minimizes patients’ risk when machine learning (ML) models are deployed in healthcare settings, by preventing the application of computer-aided diagnosis in groups of patients for whom the expected diagnostic error is large.
LDU results in higher diagnostic accuracy and fewer deferred patients when compared with learning to defer (without uncertainty) and direct triage by uncertainty algorithms, across different types of data including electronic medical records, clinical notes, and x-ray images. LDU’s performance increases monotonically as more patients are deferred, suggesting that the desired trade-off between performance and defer ratio can be obtained for a wide variety of diagnostic tasks.