VALID: Verified AI for Limiting Inefficient Diagnostics

Abstract

The overuse of diagnostic imaging, particularly computed tomography (CT) scans, is a significant issue in Canadian healthcare, leading to increased costs, prolonged wait times, and unnecessary radiation exposure. Despite guidelines, up to 30% of scans are considered low value, highlighting a gap in clinical decision support. This study introduces VALID (Verified AI for Limiting Inefficient Diagnostics), a multimodal machine learning framework designed to prospectively predict CT scan necessity at emergency department triage using data (vitals and unstructured clinical notes) at time of admission to the ED. To address data leakage, we utilize a two-phase architecture, first extracting ground-truth labels retrospectively from radiology reports using ClinicalBERT, and then training an ensemble of classifiers (MLP, Random Forest, XGBoost) on the pre-scan data. Our ensemble model achieved a macro-average recall of 88% and precision of 91%, effectively balancing the identification of necessary scans with a high overall precision. VALID acts as an early-warning clinical decision support tool to safely reduce low-value imaging while maintaining diagnostic accuracy.