Automated Diagnosis Code Assignment of Thai Free-text Clinical Notes

The International Classification of Diseases (ICD) provides a common coding system that enables health profes-sionals to identify and standardize clinical concepts for reim-bursement’ healthcare planning, and research. Manual coding is time-consuming, costly, and error-prone. A tool to assist medical coding could save time and reduce coding errors. Although many previous studies have addressed ICD classification, there have been few studies on Thai patient records. In this study, we examined techniques that predict relevant ICD-10 blocks from a discharge summary that concludes a patient's episode of care during a hospital stay. We investigated both classical supervised learning and transformers models: bag-of-n-grams, TF-IDF (Term Frequency-Inverse Document Frequency), and Multilingual BERT (Bidirectional Encoder Representations from Transformers). Our experiments were conducted on an institutional dataset of more than one hundred thousand discharge summaries. With Multilingual BERT, we achieved a macro-averaged Fl-score of 0.8824, 0.8149, and 0.7615 for the top 100, 200, and 500 most frequent ICD blocks, respectively. The 500 most frequent ICD blocks covered more than 88% of all ICD occurrences. An explainability tool can correctly highlight the words that are relevant to the predicted ICD-10 blocks. This work contributed to the lack of information on ICD classification in Thai and demonstrated the possibility of implementing a tool to assist ICD coding using institutional data.