Context-aware automated ICD coding: A semantic-driven approach

Identifying the exact International Classification of Diseases (ICD) codes describing a patient’ s health condition is essential in classifying patients with similar disease conditions. Numerous studies have devised automated approaches to retrieve the ICD codes from patients’ health records. However, majority of these methodologies have considered ICD codes solely as alphanumeric codes, overlooking their descriptions and thus neglecting the inherent semantics. Also, these methodologies overlook the one-to-many semantic relationships between diagnosis and assigned ICD code descriptions. Subsequently, this constrains these approaches from effectively assigning ICD codes with meaningful context. This work addresses these limitations by capturing the semantic similarity between the diagnosis and ICD code descriptions, while utilising the inherent one-to-many relationships between them, to accurately assign ICD codes. For this, we formulate the ICD coding problem as a Semantic Text Similarity task. The proposed approach uses a siamese stacked Bi-LSTM network to learn context-aware representations of diagnoses and ICD code descriptions. We transform each patient-visit data into sentence pairs by considering the one-to-many relationships between diagnosis and assigned ICD code descriptions. Further, we compute their semantic similarity and classify them as similar or dissimilar. The proposed approach was evaluated using 5-fold cross-validation on MIMIC-III dataset and achieved the highest evaluation metric scores (F1-score 0.66, precision 0.67, recall 0.84) compared with other sequential models. The per-label evaluation demonstrates the performance of the proposed approach for each ICD code. Furthermore, the proposed approach outperformed several existing attention-based models, demonstrating the potential use of semantics in automated ICD coding.