A simple neural vector space model for medical concept normalization using concept embeddings.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Medical Concept Normalization (MCN) unifies diverse textual references to identical concepts.
• Existing MCN methods often involve complex pipelines and hand-crafted rules.

Purpose of the Study:

• To introduce a simple, direct neural model for Medical Concept Normalization (MCN).
• To achieve competitive and state-of-the-art performance on clinical MCN tasks.

Main Methods:

• A neural MCN model utilizing an encoder (SAPBERT) and a normalized temperature-scaled softmax (NT-softmax) layer.
• Directly predicting concepts from mentions, bypassing complex preprocessing.
• Initializing NT-softmax weights with pre-computed concept embeddings.

Main Results:

• Achieved competitive performance on the ShARe/CLEF 2013 dataset.
• Established a new state-of-the-art on the 2019 n2c2/OHNLP MCN task.
• Demonstrated the superiority of NT-softmax over conventional softmax for MCN.

Conclusions:

• The proposed simple neural MCN model offers a more effective and straightforward approach.
• The NT-softmax layer and specific initialization strategies enhance MCN performance.
• Future MCN research may need to address the challenge of unseen concepts.