UMLS mapping and Word embeddings for ICD code assignment using the MIMIC-III intensive care database.

Area of Science:

• Medical Informatics
• Computational Linguistics
• Machine Learning

Background:

• Electronic Health Records (EHRs) utilize diagnosis codes for billing.
• Automated ICD code assignment is crucial for decision support systems.
• The task presents a significant multi-label classification challenge due to vast label sets and lengthy patient records.

Purpose of the Study:

• To establish a baseline for automatic ICD code assignment.
• To evaluate the effectiveness of Support Vector Machines (SVM) and FastText models.
• To explore the impact of Unified Medical Language System (UMLS) metathesaurus mappings on word embedding models.

Main Methods:

• Utilized the Medical Information Mart for Intensive Care (MIMIC-III) database for training data.
• Extended training data with 'is-a' relationships from the ICD-9 hierarchy.
• Implemented Support Vector Machines (SVM) and FastText with UMLS metathesaurus mappings into word embedding models.

Main Results:

• FastText, utilizing label cardinality estimation, achieved a F1-Score of 62.2%.
• The FastText model demonstrated high recall and notable performance improvements over prior models.
• Results were validated using 10-fold cross-validation.

Conclusions:

• FastText provides a strong baseline for automatic ICD code assignment.
• Integrating UMLS metathesaurus mappings enhances word embedding models for this task.
• The proposed methods show promise in improving the efficiency and accuracy of clinical coding.