Instability in clinical risk stratification models using deep learning

TL;DR

This paper investigates the instability of deep learning models in healthcare, revealing that repeated training can lead to patient-level outcome variability despite stable overall performance, and proposes metrics and strategies to improve stability.

Contribution

It introduces new stability metrics and mitigation strategies for deep learning models trained on electronic health records in clinical risk prediction.

Findings

Repeated training yields patient-level outcome variability.

Global performance metrics remain stable despite instability.

Proposed metrics effectively measure training randomness effects.

Abstract

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.