A Deep Learning-Driven Disease Classification Method Using MIMIC-III Database and Natural Language Processing

The automated encoding of medical diagnoses from clinical records is an indispensable task in healthcare management. This study proposes a deep learning-driven method for International Classification of Diseases (ICD) coding, derived from clinical narratives using the MIMIC-III database. The proposed approach leverages a Label Attention Model for ICD coding from clinical text (LAAT), and explores the effectiveness of various loss functions, notably focal loss. ICD codes are classified according to their frequency, and model hyperparameters are optimized separately for each category. Empirical results demonstrate that the integration of focal loss significantly bolsters the performance of LAAT in ICD coding, achieving an Area Under the Curve (AUC) of 0.92131 and a micro-F1 score of 0.98658 for the entire MIMIC-III dataset. Additionally, the data suggests that fine-tuning hyperparameters for specific label groups, based on their frequency, fosters further performance enhancements. The developed approach can facilitate precise and efficient ICD coding from clinical notes, providing invaluable insights for healthcare practitioners and investigators.