Differentiating Sense through Semantic Interaction Data.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Semantically related biomedical terms with different representations (e.g., dementia, delirium) present challenges in text understanding.
• Accurate differentiation of such concepts is crucial for effective clinical text analysis and knowledge extraction.

Purpose of the Study:

• To explore the utility of interaction frequency data between semantic elements for differentiating semantically related concept pairs.
• To develop and evaluate machine learning models for classifying biomedical concepts based on semantic predications.

Main Methods:

• Extracted semantic predications from biomedical literature to generate feature datasets.
• Applied Expectation Maximization clustering (with and without labels) to these datasets.
• Trained and evaluated various concept classifying algorithms using the processed data.

Main Results:

• Unlabeled datasets showed expected cluster counts and proportions in 80% of cases.
• Labeled datasets demonstrated similar proportions when cluster counts were restricted to unique labels.
• The top-performing classifier achieved 89% accuracy, with F1 scores ranging from 0.69 to 1.

Conclusions:

• Interaction frequency data derived from semantic predications can effectively differentiate semantically related biomedical concepts.
• The developed classification models show high performance, suggesting applicability to natural language processing of clinical text.
• This approach offers a potential solution for improving the accuracy and efficiency of clinical information extraction.