A comparative study of structured and verbalized clinical metadata for robust medical case retrieval

Clinical metadata is widely used as auxiliary information in medical image analysis and provides important diagnostic cues for similar-case retrieval. With the development of biomedical language models, structured clinical metadata can be converted into natural-language descriptions and encoded as text representations. However, it remains unclear whether such verbalized metadata is consistently more effective than compact structured vectors when retrieval accuracy, computational efficiency, and robustness are considered together. We compare structured vector retrieval, sentence-level verbalized metadata retrieval, and phrase-level verbalized metadata retrieval using multiple biomedical text encoders. Experiments are conducted on three public medical datasets spanning lung CT (LIDC-IDRI), dermatology (PAD-UFES-20), and breast mammography (CBIS-DDSM). We report five-fold mean ± standard deviation and paired statistical tests, and further evaluate robustness under missing and incorrect metadata perturbations. Across the three datasets, structured metadata provides a strong and computationally efficient baseline, while sentence- and phrase-level verbalization produce dataset- and metric-dependent changes. The key paired Recall@5 and MV@5 comparisons do not show a consistent statistically significant improvement over structured vectors. Robustness trends also vary with the dataset and perturbation type. These findings indicate that compact structured metadata remains a practical baseline for medical case retrieval, while the benefit of verbalization depends on the metadata structure and evaluation setting.