Comparison of different feature extraction methods for applicable automated ICD coding.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Automated ICD coding using machine learning is crucial for medical record analysis.
• Traditional methods like Bag-of-Words (BoW) are common, but advanced techniques like word2vec (W2V) and BERT are underutilized.
• This research addresses the need to identify optimal feature extraction for improved coding accuracy.

Purpose of the Study:

• To compare the effectiveness of Bag-of-Words (BoW), word2vec (W2V), and BERT variants for automated ICD coding.
• To determine which feature extraction method performs best across different code frequencies and datasets.
• To provide insights into building more accurate machine learning models for medical coding.

Main Methods:

• Experiments were conducted on Chinese and Spanish medical datasets with varying numbers of records and ICD codes.
• Coding tasks were designed with different code frequency thresholds to simulate varying task complexity.
• Bag-of-Words (BoW), word2vec (W2V), and BERT variants were evaluated using traditional classifiers.

Main Results:

• BERT variants achieved superior performance (Micro-F1 up to 93.9%) for tasks with high frequency codes (e.g., Fuwai dataset >= 140).
• Bag-of-Words (BoW) outperformed other methods (Micro-F1 up to 83%) for tasks involving both frequent and infrequent codes (e.g., Fuwai dataset < 140).
• Performance variations were observed between datasets due to language and codeset differences, highlighting the importance of method selection.

Conclusions:

• Fine-tuning BERT variants is optimal for automated ICD coding tasks focused on frequent codes.
• Bag-of-Words (BoW) remains the best approach for tasks encompassing both frequent and infrequent ICD codes.
• The choice of feature extraction method significantly impacts automated ICD coding performance, influenced by data characteristics.