Comparative Analysis of BiLSTM and IndoBERT for Classification in Indonesian Radiology Reports

The process of reviewing and classifying radiology reports in Indonesian healthcare involves too much labor and carries the chance of errors which requires moving towards automation. The purpose of this study is to determine which method, BiLSTM vs IndoBERT, works best for the automatic classification of diseases in Indonesian radiology reports to lessen the workload of radiologists and improve how accurately cases are diagnosed. The study involved processing 1,719 anonymous radiology reports from many hospitals in Jakarta, Indonesia. We tried out two methods for Natural Language Processing (NLP): BiLSTM after much preprocessing, including SMOTE for ensuring equal class sizes and IndoBERT which requiring less preprocessing. The BiLSTM algorithm was designed with sequential bidirectionality, embedding layers and dropout regularization, whereas IndoBERT made use of transformer structure and multi-headed self-attention. We tested different models for accuracy, macro F1-score and weighted F1-score. IndoBERT scored 93% in accuracy, 88% in macro F1-score and 93% in weighted F1-score which is higher than BiLSTM which got 92% accuracy, 86% macro F1-score and 91 % weighted F1score. Because IndoBERT does not need a lot of preprocessing, it can be easily used in Indonesian hospitals, since it still works well even with many imbalanced kinds of diseases.