Deciphering Abbreviations in Malaysian Clinical Notes Using Machine Learning.

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning (ML)
• Health Informatics

Background:

• Clinical notes contain numerous abbreviations, posing challenges for automated information extraction.
• Existing NLP systems often struggle with the nuances of clinical abbreviations, especially in low-resource settings.
• Developing specialized models for local contexts is crucial for accurate healthcare data management.

Purpose of the Study:

• To develop and evaluate the first Malaysian machine learning model for detecting and disambiguating clinical abbreviations.
• To integrate this model into the MyHarmony NLP system for enhanced clinical information extraction.
• To assess the model's feasibility in low-resource settings using word embedding techniques.

Main Methods:

• A Malaysian clinical word embedding was developed using the Word2Vec model on electronic discharge summaries.
• Performance was evaluated on abbreviation detection and disambiguation tasks.
• The local embedding was compared against conventional rule-based and FastText embeddings using machine learning classifiers.

Main Results:

• The Malaysian clinical word embedding achieved high performance in both abbreviation detection (F-score 0.9519) and disambiguation (F-score 0.9903).
• The Decision Tree classifier with the local embedding outperformed other methods.
• The model demonstrated effectiveness despite a smaller vocabulary and dimension compared to non-clinical embeddings.

Conclusions:

• Local clinical word embeddings, even with simpler ML algorithms, can effectively decipher abbreviations.
• The developed model requires lower computational resources, making it suitable for low-resource settings like Malaysia.
• Integration into MyHarmony is expected to improve clinical term recognition, enhancing healthcare monitoring and policy-making.