Exploiting ICD Hierarchy for Classification of EHRs in Spanish Through Multi-Task Transformers

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Electronic Health Records (EHRs) contain critical clinical information requiring expert coding using the International Classification of Diseases (ICD).
• Automated coding of EHRs facilitates information sharing and statistical analysis.
• Existing Natural Language Understanding (NLU) tools for clinical text mining primarily focus on English, leaving a gap for other languages like Spanish.

Purpose of the Study:

• To explore computer-aided multi-label classification for Spanish EHRs.
• To address the challenge of limited annotated data for clinical text mining in non-English languages.
• To leverage the hierarchical structure of the International Classification of Diseases (ICD) for improved classification.

Main Methods:

• Utilized language-modeling aware Transformers for clinical text mining on Spanish EHRs.
• Employed a multi-task classification strategy incorporating a hierarchical head to exploit ICD coding synergies.
• Trained models using a smaller, in-domain, unannotated corpus of Spanish EHRs.

Main Results:

• Developed and released a hierarchical head for multi-label classification tailored for Spanish EHRs.
• Demonstrated the benefit of exploiting the ICD hierarchy through multi-task classification.
• Showcased a viable approach for NLU in clinical text mining for under-resourced languages.

Conclusions:

• The proposed hierarchical approach enhances multi-label classification of Spanish EHRs.
• Leveraging language models and ICD hierarchy is effective even with limited in-domain data.
• This work contributes to advancing clinical text mining for Spanish and potentially other languages.