Enhanced neurologic concept recognition using a named entity recognition model based on transformers.

Area of Science:

• Natural Language Processing (NLP)
• Artificial Intelligence (AI)
• Biomedical Informatics

Background:

• Deep learning is increasingly used for disease and drug recognition in clinical data.
• Recognition of signs and symptoms is crucial for deep phenotyping and precision medicine.
• Limited research exists on deep learning for sign and symptom recognition.

Purpose of the Study:

• To develop and evaluate deep learning models for recognizing neurological signs and symptoms.
• To map identified signs and symptoms to a neuro-ontology for clinical concept standardization.
• To compare the performance of Convolutional Neural Networks (CNNs) and Bidirectional Encoder Representations from Transformers (BERT) for this task.

Main Methods:

• Developed a named entity recognition (NER) model using deep learning.
• Compared CNN and BERT-based models for identifying text spans of neurological signs and symptoms.
• Evaluated models on three diverse text corpora: electronic health records (physician notes), neurologic textbooks, and genetic disease databases.

Main Results:

• BERT-based models outperformed CNN-based models in recognizing signs and symptoms.
• Model performance varied across corpora, with best results on clinical synopses and worst on physician notes.
• Shorter text spans for signs and symptoms improved model performance.
• Achieved recall ranging from 59.5% to 82.0% and precision from 61.7% to 80.4%.

Conclusions:

• Deep learning, particularly transformer models, shows significant promise for automated sign and symptom recognition.
• Further NLP advancements could enable fully automated identification in clinical records and literature.
• Improved sign and symptom recognition supports deep phenotyping and precision medicine initiatives.