TriAID: Chest X-Ray Follow-Up Imaging Recommendation

Abstract

This paper presents TriAID, a clinical decision-support framework for identifying follow-up imaging recommendations from chest radiography studies. The proposed system combines structured report-derived features with pathology probability outputs generated by the MedRAX computer vision model to predict follow-up imaging categories. A One-vs-Rest (OvR) multi-label classification approach is employed to classify radiology report information into recommendation categories including X-ray, CT, or no follow-up. The model is trained on 26,797 chest radiography studies paired with corresponding clinical reports extracted from the publicly available ReXGradient-160K dataset. A pre-processing pipeline was developed to identify follow-up recommendations from radiologist findings and impressions, enabling construction of a structured training dataset. Multiple feature extraction and classification strategies were evaluated to determine an effective modelling approach. The resulting model achieves a micro-averaged precision of 78.56%, recall of 82.29%, and F1-score of 80.38%. The proposed framework is intended to assist clinical workflows by providing structured follow-up recommendation predictions while maintaining interpretability. The system is designed as a supportive tool for clinical environments and is not intended to replace professional medical judgment.