Tadpole: A Framework for Characterizing Temporal Model Data Drifts at the Point-of-Care

Temporal models that use timestamped data from electronic health record (EHR) systems could enhance patient outcomes and reduce costs in point-of-care (POC) settings, especially for complex prediction tasks. However, a major obstacle to their widespread use is data drifts, which are changes in how input data relates to target outcomes. Currently, no studies have systematically categorized these data drifts for temporal EHR models or identified the key questions needed to address their impact. This study aims to create a conceptual framework, TADPole, for characterizing data drifts specifically for temporal models in POC environments.