Validating the representation of distance between infarct diseases using word embedding

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• The pivot and cluster strategy (PCS) is a diagnostic reasoning method that groups diseases.
• Physicians intuitively understand disease similarity, but quantitative validation is lacking.
• This study evaluates inter-disease distances derived from word embeddings within PCS.

Purpose of the Study:

• To quantitatively assess if inter-disease distances from word embedding vectors in PCS accurately represent similar disease groups.
• To validate the use of word embeddings for representing disease similarity in a specific domain.

Main Methods:

• Extracted abstracts from the Ichushi Web database for morphological analysis.
• Trained Word2Vec, FastText, and GloVe models to obtain word embedding vectors.
• Calculated internal (CCC) and external (ARI, NMI, AMI) validity measures using ICD-10 codes.

Main Results:

• Word2Vec with Euclidean distance and centroid method yielded the highest internal validity (CCC=0.8690).
• Cosine distance with correlation/average/weighted methods maximized external validity (AMI=0.4109).
• FastText and GloVe showed similar trends; optimal metrics differed for internal vs. external validity.

Conclusions:

• Internal and external validity measures require different metrics and methods for optimization.
• Cosine distance is recommended for ICD-10 alignment, while Euclidean distance is suitable for word frequency.
• Word2Vec-trained distributed representations offer objective inter-disease distances for PCS in the "infarction" domain.