Using Medical Named Entity Recognition in Automatic ICD Prediction.

Area of Science:

• Artificial Intelligence in Healthcare
• Natural Language Processing (NLP) for Medical Coding
• Machine Learning Applications in Clinical Data Analysis

Background:

• The International Classification of Diseases (ICD) is a critical standard for medical coding and health statistics.
• Existing automatic ICD encoding systems often overlook crucial medical entities within raw text inputs.
• A need exists for more sophisticated algorithms that incorporate medical entity recognition for accurate ICD prediction.

Purpose of the Study:

• To propose and evaluate a novel algorithm for automated ICD code prediction directly from patient claims.
• To enhance the accuracy of ICD encoding by integrating medical named entity recognition (NER) and advanced embedding techniques.
• To develop an efficient ICD prediction system that requires minimal training data.

Main Methods:

• Utilized a BERT-based Medical Named Entity Recognition (NER) model to identify key medical entities in patient claims.
• Employed the ClinicalBERT model to generate embeddings for the extracted medical entities.
• Created embeddings for the ICD catalog, focusing on long descriptions, and calculated cosine similarity to predict the most relevant ICD codes.

Main Results:

• The proposed algorithm successfully identifies critical medical entities and predicts relevant ICD codes.
• The system demonstrates high efficiency and does not necessitate large datasets for training.
• Experimental results on a medical dataset show a precision rate of approximately 90%.

Conclusions:

• The developed algorithm offers a significant improvement in automatic ICD encoding by incorporating medical entity recognition.
• This approach provides a more accurate and efficient method for predicting ICD codes from patient claims.
• The algorithm's ability to perform well with limited training data makes it a practical solution for healthcare applications.