Automated Diagnosis Coding with Combined Text Representations.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Automated diagnosis coding from clinical text is crucial for healthcare.
• Accurately assigning a large number of variable diagnosis codes presents significant challenges.

Purpose of the Study:

• To explore various text representations and classification models for automated ICD-9 code assignment.
• To identify optimal methods for coding discharge summaries in the MIMIC-III database.

Main Methods:

• Utilized diverse text representations (e.g., TF-IDF, word embeddings) for discharge summaries.
• Applied and compared various classification models for ICD-9 code prediction.
• Evaluated model performance based on code frequency and representation effectiveness.

Main Results:

• The effectiveness of text representations varied depending on the diagnosis code's frequency.
• Combining models trained on different representations yielded the best performance.
• Achieved improved accuracy in automated diagnosis coding.

Conclusions:

• A hybrid approach combining multiple models and text representations is superior for automated diagnosis coding.
• Model performance is influenced by diagnosis code frequency, necessitating adaptive strategies.
• This research advances automated clinical coding accuracy and efficiency.