Instability in clinical risk stratification models using deep learning
This addresses the problem of unreliable patient-level predictions in healthcare deployments, which is critical for clinicians and patients, though it is incremental as it builds on known instability issues in deep learning.
The study investigated the instability of deep learning models in clinical risk stratification, showing that repeated training runs on the same data lead to significantly different patient-level outcomes despite stable global performance metrics, and proposed metrics and strategies to measure and mitigate this instability.
While it has been well known in the ML community that deep learning models suffer from instability, the consequences for healthcare deployments are under characterised. We study the stability of different model architectures trained on electronic health records, using a set of outpatient prediction tasks as a case study. We show that repeated training runs of the same deep learning model on the same training data can result in significantly different outcomes at a patient level even though global performance metrics remain stable. We propose two stability metrics for measuring the effect of randomness of model training, as well as mitigation strategies for improving model stability.