A Novel Machine Learning Strategy For Accurate ICD Code Identification: Towards Reliable Automation Of Clinical Coding

Abstract

Manual coding of the International Classification of Diseases (ICD) from free-text clinical notes is a labor-intensive process directly tied to billing, epidemiological reporting, and downstream clinical analytics. Building on recent advances in this field, this study introduces a novel hybrid machine learning approach, the Clinical-BERT Label-Graph Attention model (CB-LGA), to accurately identify ICD codes and automate the process. The proposed architecture includes two main components: (1) embeddings adapted to the various domains in which ICD code’s function; (2) a convolutional feature extractor that learns a feature map through iterative, hierarchical processing of the respective 1000 sub-domain embeddings. The architecture combines domain-adapted transformer embeddings, label-wise attention over these embeddings, and a label co-occurrence graph module to capture hierarchical and co-occurrence relationships among ICD codes. The strategy was evaluated conceptually and empirically against four representative baseline systems: a TF-IDF logistic regression system, a convolutional neural network, a ConvNet with label attention for the test (CAML), a bidirectional gated recurrent unit (Bi-GRU) with hierarchical attention, and a pretrained language model for ICD coding (PLM-ICD), using the MIMIC-III clinical notes corpus. The results show that, at the micro level, the proposed CB-LGA improves micro-averaged F1 by 3.5 pp; at the macro level, it improves macro-averaged F1 by 2.5 pp and precision at 5 (P@5) by 3.2 pp over the best-performing baseline. In addition, the proposed CB-LGA model achieved better results for rare ICD codes associated with long-tail diseases. The results indicate that integrating contextual representations of language with explicit modeling of relationships between labels is a new and promising generalizable approach to automatic ICD coding in real-world health information systems.