Monitoring strategies for continuous evaluation of deployed clinical prediction models

Abstract

Objective:

As machine learning adoption in clinical practice continues to grow, deployed classifiers must be continuously monitored and updated (retrained) to protect against data drift that stems from inevitable changes, including evolving medical practices and shifting patient populations. However, successful clinical machine learning classifiers will lead to a change in care which may change the distribution of features, labels, and their relationship. For example, “high risk” cases that were correctly identified by the model may ultimately get labeled as “low risk” thanks to an intervention prompted by the model’s alert. Classifier surveillance systems naive to such deployment-induced feedback loops will estimate lower model performance and lead to degraded future classifier retrains. The objective of this study is to simulate the impact of these feedback loops, propose feedback aware monitoring strategies as a solution, and assess the performance of these alternative monitoring strategies through simulations.

Methods:

We propose Adherence Weighted and Sampling Weighted Monitoring as two feedback loop-aware surveillance strategies. Through simulation we evaluate their ability to accurately appraise post deployment model performance and to initiate safe and accurate classifier retraining.

Results:

Measured across accuracy, area under the receiver operating characteristic curve, average precision, brier score, expected calibration error, F1, precision, sensitivity, and specificity, in the presence of feedback loops, Adherence Weighted and Sampling Weighted strategies have the highest fidelity to the ground truth classifier performance while standard approaches yield the most inaccurate estimations. Furthermore, in simulations with true data drift, retraining using standard unweighted approaches results in a AUROC score of 0.52 (drop from 0.72). In contrast, retraining based on Adherence Weighted and Sampling Weighted strategies recover performance to 0.67 which is comparable to what a new model trained from scratch on the existing and shifted data would obtain.

Conclusion:

Compared to standard approaches, Adherence Weighted and Sampling Weighted strategies yield more accurate classifier performance estimates, measured according to the no-treatment potential outcome. Retraining based on these strategies bring stronger performance recovery when tested against data drift and feedback loops than do standard approaches.