Enhancing Disease Diagnosis Coding: A Deep Learning Approach with Bidirectional GRU For ICD-10 Classification

Abstract

The health insurance claim in hospitals involves selecting specific ICD-10 codes for primary diagnosis texts. With rising claim volumes, the need for faster, more accurate coding is critical. This study develops a deep learning model to classify diagnosis texts into relevant ICD-10 codes using 9,982 original medical records from a national referral hospital under the Indonesian Ministry of Health. The classification method employs a BiGRU layer architecture, known for its effectiveness in handling sequential data, such as diagnosis texts. BiGRU operates bidirectionally, enhancing the model’s ability to capture the context from both past and future sequences. In this architecture, the BiGRU layer serves as the classification layer, stacked above the BERT layer, which functions as the vector embedding layer, converting text into numerical representations for the model. The results of the study demonstrate a promising solution for codifying primary diagnosis texts, achieving a precision of 82.18% and a recall of 81.59%. Despite the strong performance of the model, further improvements are possible. Interestingly, the study also observed that the size of the class volume per ICD-10 code is not the only factor affecting classification performance, as some classes with smaller volumes exhibited better classification results. However, merging rare classes did not improve performance and even worsened it, suggesting that better ways to handle underrepresented classes are needed. Experiments with different embedding layers, such as IndoBERT and BioClinicalBERT, and hyperparameter tuning yielded minimal performance gains, suggesting the need for alternative optimization strategies.