Abstract
The task of classifying clinical case reports by their specialty is challenging because of the heterogeneity of the text in the reports, overlapping terminology of specialties, class imbalance, and the use of proxy labels in the biomedical literature where the reports are located. We tested a MultiCaRe-CERGFormer multi-view classifier, which fuses word- and character-level TF–IDF features obtained via truncated SVD, precomputed biomedical document embeddings, and specialty-specific concept indicators and features learned from similar cases in the training partition. A gated fusion module is provided to fuse the encoded branches for multiclass prediction. There were 8564 case reports, assigned to 13 specialty categories created from the metadata. The models were evaluated with five-fold grouped cross-validation, where the reports in the same source article belong to the same fold. The proposed model achieved a mean accuracy of 0.8152 ± 0.0070, a macro-F1 of 0.6693 ± 0.0179, a weighted-F1 of 0.8147, a macro-AUC of 0.9706 ± 0.0049, and a macro-average precision of 0.7430 ± 0.0227. The mean macro-F1 (0.6705 ± 0.0264) of the word- and character-level TF–IDF SVM was slightly higher. Therefore, the approach proposed was competitive under the evaluated protocol but was not "leading" on all metrics. Due to the nature of the targets, which were based on metadata of the articles instead of clinician adjudication, these results provide a measure of the prediction of proxy labels and are not meant to indicate diagnostic validity, clinical utility, or external generalizability.