Efficient CMR Report Classification Through Synthetic Data Distillation and Metadata Fusion

This study presents a ClinicalBERT model specialized for cardiovascular MR (CMR) report classification through synthetic data-driven knowledge distillation and metadata integration. To enable efficient and accurate classification of CMR reports, we generated 20,000 synthetic reports using a large language model, transferring nuanced medical knowledge to Clinical-BERT while minimizing computational costs. Metadata integration further enhanced classification specificity for target diseases, with mixed results observed for negation filtering. Our optimized model achieved a 50% performance improvement over the baseline, demonstrating synthetic data's potential for training specialized NLP models in clinical applications. This work establishes a framework for efficient, deployable CMR classification models in dataset creation, cleaning, and auditing. All relevant source code can be found at our GitHub repository: https://github.com/PRMA-PRMA/HeartScan_Classifier