Using Neural Multi-task Learning to Extract Substance Abuse Information from Clinical Notes.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Linguistics

Background:

• Substance abuse has severe health implications.
• Clinical notes contain crucial, yet unstructured, patient substance abuse history.
• Accurate extraction is essential for risk assessment and patient care.

Purpose of the Study:

• To develop and evaluate a novel neural architecture for automated substance abuse information extraction from clinical text.
• To improve the accuracy and efficiency of identifying substance abuse status and related details.

Main Methods:

• Utilized a novel neural architecture employing multi-task learning.
• Compared the model's performance against previous work and discrete model baselines.
• Evaluated the model on a withheld test set for substance abuse status and entity extraction.

Main Results:

• Achieved high F1 scores (0.88-0.95) for detecting substance abuse status (current, none, past, unknown).
• Successfully extracted other substance abuse entities like amount, frequency, and history with high performance.
• Demonstrated feasibility of extraction with limited annotated data, achieving 0.84-0.89 precision on a large, automatically annotated dataset.

Conclusions:

• The proposed neural multi-task learning model effectively extracts substance abuse information from clinical notes.
• This approach enhances the ability to assess patient risks associated with substance abuse.
• The model shows promise for large-scale, automated clinical note analysis with minimal manual annotation.