Silent Model Degradation in Clinical AI Detecting and Quantifying Undocumented Data Drift in Live EHR Systems
Year: 2024
Type: article
Abstract: Clinical prediction models implemented within clinical healthcare are getting high stakes clinical decision support, however, they are seldom subjected to post deployment performance other than simple validation. This paper draws attention to and defines silent model degradation a failure mode where predictive validity deteriorates over time as a result of undocumented data drift failing to initiate standard system warnings. We will use a large longitudinal electronic health record (EHR) dataset in a tertiary care hospital system to show that both covariate drift and concept drift increase gradually over time following deployment and result in more gradual changes in discrimination, calibration, and prediction stability. To overcome this risk, we design CLIOPS, a combined framework of post-deployment monitoring that combines temporal drift, longitudinal performance deterioration modelling, and unsupervised early-warnings without depending on instant outcome classification. On comparative analysis, CLIOPS is less prone to operational load and detects degradation sooner and more steadily than current feature-based and label-dependent drift detection procedures. These results demonstrate that accurate performance loss leading to clinical consequences may be hidden and that safe and reliable implementation of clinical AI must be accompanied with label-free monitoring.
Institutions: Chandigarh University, Universitat Politècnica de València
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7218173337