Code Synonyms Do Matter: Multiple Synonyms Matching Network for Automatic ICD Coding
The paper titled "Code Synonyms Do Matter: Multiple Synonyms Matching Network for Automatic ICD Coding" introduces a novel approach to improve automatic ICD coding by leveraging synonyms of disease codes. This research primarily addresses the limitations of extant methodologies which typically employ label attention with code representations derived from either code hierarchies or descriptions provided in the International Classification of Diseases (ICD). It is posited that these approaches might be constrained by the inherent variability in code expressions found within electronic medical records (EMRs) compared to their standard descriptions in ICD databases.
Synonyms as a Valuable Resource
The central thesis of this work is that synonyms can furnish a more holistic representation of the codes. The authors utilize the Unified Medical Language System (UMLS) to align ICD codes with their corresponding Concept Unique Identifiers (CUIs) and extract related synonyms. This comprehensive assembly of synonyms is then integrated into a Multiple Synonyms Matching Network (MSMN), which enhances the learning of code representations and ultimately refines the process of code classification.
Methodological Innovations
The MSMN approach applies a shared LSTM to encode both EMR text and code synonyms, utilizing a sophisticated multi-synonyms attention mechanism inspired by transformer-based architectures like multi-head attention. This allows for a nuanced extraction of code-related representations across varying snippets of text. In classification, a biaffine transformation is adopted to measure similarity between text and code representations, circumventing the need for code-dependent parameters that typically limit the performance due to the sparse occurrence of certain codes.
Empirical Validation
The efficacy of MSMN is rigorously tested using the MIMIC-III dataset under both full code and top-50 code settings. The proposed method demonstrates superior performance over prior state-of-the-art models in key metrics such as macro and micro AUC, F1 scores, and precision@k. For instance, in the full code setting, MSMN achieves a macro-AUC of 95.0 and a micro-F1 score of 58.4, among others, indicating an advancement over benchmark methods by as much as +2.0 in terms of macro-AUC and +0.9 in micro-F1.
Discussion and Implications
The introduction of multi-synonyms attention in MSMN leads to notable improvements, as demonstrated by an ablation study that elucidates the incremental performance gains when increasing the count of synonyms. Furthermore, the MSMN is shown to handle diverse synonym representations efficiently, clustering similar texts accurately, yet distinguishing sufficiently different texts to avert reliance on semantically disparate representations.
Conclusion and Speculation on Future Directions
This study not only reaffirms the potential of leveraging external medical knowledge bases such as UMLS but underscores the role of synonyms as pivotal contributors to enhancing ICD code representation and classification. Moving forward, the implications for improving clinical decision support, patient similarity assessments, and reducing administrative burden in ICD coding are significant. Future research may explore the integration of this approach with more sophisticated language models or extend to other domains where analogous hierarchical classification systems and synonym redundancies are prevalent.