T5 Based ICD-10 Classification and Clinical Text Generation from South Bali Medical Records

Abstrak

This study aims to develop a T5-based natural language processing model for classification and text generation tasks on Indonesian medical records from the South Bali region. The research focuses on two main tasks: diagnosis classification based on ICD-10 codes and medical text generation for three clinical elements: Medication and Dosage, Management and Procedure, and Recommendation. The dataset used consists of 7,263 anonymized medical record entries, which are narrative in nature and vary in length and structure. On the ICD-10 classification task, the model achieved 91.6% accuracy after filtering out inconsistent diagnosis labels, with high and balanced precision, recall, and F1 scores. The training process showed stable convergence, with validation loss decreasing to 0.03. For text generation tasks, the model was evaluated using BLEU, ROUGE, and BERTScore metrics. The best results were obtained on the Management and Procedure task with an average BERTScore F1 of 0.90 and ROUGE of 0.53, indicating high semantic similarity between the predicted and reference texts. Meanwhile, the Medication and Dosage task yielded a BERTScore F1 of around 0.84, with a low BLEU score, attributed to the large amount of empty reference data. Overall, the T5 model performed well in understanding the local clinical context and producing semantically appropriate output, although there is still room for improvement through label enrichment and further fine-tuning. This research is an initial contribution to the use of generative models for Indonesian medical records and demonstrates the potential to develop more accurate and context-aware NLP-based clinical systems.

Identifier

• DOI: https://doi.org/10.62527/joiv.10.3.4130
• OpenAlex: https://openalex.org/W7163739744