$\mathcal{LATA}{-}$ Label Attention Transformer Architectures for ICD-10 Coding of Unstructured Clinical Notes

Effective code assignment for patient clinical records in a hospital plays a significant role in the process of standardizing medical records, mainly for streamlining clinical care delivery, billing, and managing insurance claims. The current practice employed is manual coding, usually carried out by trained medical coders, making the process subjective, error-prone, inexact, and time-consuming. To alleviate this cost-intensive process, intelligent coding systems built on patients' structured electronic medical records are critical. Classification of medical diagnostic codes, like ICD-10, is widely employed to categorize patients' clinical conditions and associated diagnoses. In this work, we present a neural model
$\mathcal{LATA}$
, built on Label Attention Transformer Architectures for automatic assignment of ICD-10 codes. Our work is benchmarked on the CodiEsp dataset, a dataset for automatic clinical coding systems for multilingual medical documents, used in the eHealth CLEF 2020-Multilingual Information Extraction Shared Task. The experimental results reveal that the proposed
$\mathcal{LATA}$
variants outperform their basic BERT counterparts by 33-49% in terms of standard metrics like precision, recall, F1-score and mean average precision. The label attention mechanism also enables direct extraction of textual evidence in medical documents that map to the clinical ICD-10 diagnostic codes.