Identifying Clinical Terms in Medical Text Using Ontology-Guided Machine Learning

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Accurate automatic recognition of medical concepts in unstructured text is crucial for clinical and research applications, impacting electronic health record analysis.
• Medical concept mining is challenging due to the prevalence of synonyms and nonstandard terminology.

Purpose of the Study:

• To develop a machine learning model for concept recognition in large unstructured text.
• To optimize the use of ontological structures for identifying previously unobserved synonyms.

Main Methods:

• A neural dictionary model, the Neural Concept Recognizer (NCR), was developed using a convolutional neural network.
• The model encodes input phrases and ranks medical concepts by similarity, utilizing the hierarchical structure of biomedical ontologies.
• Trained on Human Phenotype Ontology (HPO) and Systematized Nomenclature of Medicine - Clinical Terms (SNOMED-CT).

Main Results:

• Achieved 1.7%-3% higher F1-scores than rule-based baselines on HPO-trained models using PubMed abstracts and clinical reports.
• Achieved 0.9%-1.3% higher F1-scores than baselines on SNOMED-CT-trained models using ICU discharge summaries.
• Demonstrated high accuracy and the value of using ontology taxonomy structure in concept recognition.

Conclusions:

• The Neural Concept Recognizer (NCR) generalizes efficiently to new synonyms without large-scale labeled data.
• Outperforms state-of-the-art methods, offering an advantage over rule-based systems that struggle with unseen synonyms.
• Provides a robust solution for medical concept recognition, addressing limitations of existing machine learning methods.