Hybrid approach combining deep learning and a rule based expert system for concept extraction from prescriptions

Area of Science:

• Medical Informatics
• Natural Language Processing

Background:

• Concept extraction from prescriptions is vital for healthcare applications like pharmacovigilance and medication adherence.
• Rule-based systems struggle with the complexity of natural language directions in prescriptions.

Purpose of the Study:

• To develop and evaluate a hybrid approach for accurate concept extraction from prescription text.
• To improve upon existing methods for identifying key information such as dosage, frequency, and duration.

Main Methods:

• A combination of rule-based expert systems and deep learning (DL) models was employed.
• A fine-tuned BERT transformer and Gram Convolutional Neural Network (CNN) based Named Entity Recognition (NER) architecture formed the DL module.
• Domain heuristics, intelligent labeling, and bootstrapping were used to enhance DL model performance.

Main Results:

• The hybrid method achieved high evaluation scores for concept extraction from real-world prescription data.
• This approach demonstrated superior performance compared to existing methods in the literature.
• Successfully extracted concepts including frequency, dosage, and duration from complex prescription directions.

Conclusions:

• The proposed targeted approach offers a significant advancement in concept extraction from medical prescriptions.
• This method provides a robust foundation for various downstream healthcare decision-making processes.
• Achieved state-of-the-art performance in concept extraction from doctor's prescriptions.