Learning to identify treatment relations in clinical text

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Data Mining

Background:

• Physicians document treatment rationales in clinical notes, but this information is often unstructured and not computationally accessible.
• Extracting treatment relations is crucial for understanding clinical decision-making and improving healthcare data analysis.

Purpose of the Study:

• To develop and evaluate a supervised learning system for predicting treatment relations between medical concepts in clinical notes.
• To assess the system's performance against established methods like SemRep.

Main Methods:

• A supervised learning model was trained on 958 manually annotated treatment relations from 6,864 discharge summaries.
• Features included lexical, semantic, MEDication Indication (MEDI) resource, and SemRep data.
• The system predicted the existence of a treatment relation between pairs of medical concepts.

Main Results:

• The developed supervised learning system achieved an F1-measure of 84.90.
• This result was significantly superior to the F1-measure of 72.34 obtained by SemRep.
• The system demonstrated enhanced accuracy in identifying treatment relations.

Conclusions:

• The supervised learning system effectively predicts treatment relations in clinical notes, offering a more computable format for this vital information.
• This advancement surpasses current methods, paving the way for improved clinical data utilization and analysis.