Selecting information in electronic health records for knowledge acquisition

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Clinical Data Mining

Background:

• Automated acquisition of biomedical relations is crucial for pharmacovigilance and decision support.
• Extracting specific, clinically meaningful relations from noisy clinical text remains challenging.
• Co-occurrence of entities does not guarantee a specific clinical relation.

Purpose of the Study:

• To improve the automated acquisition of disease-manifestation related symptom (MRS) and drug-adverse drug event (ADE) relations from clinical reports.
• To evaluate the impact of filtering by report sections on the performance of relation extraction.
• To enhance the detection of specific biomedical associations in narrative electronic health records.

Main Methods:

• Focused on extracting two relation types: disease-MRS and drug-ADE.
• Employed filtering by sections within clinical reports to refine information.
• Evaluated the performance using recall and precision metrics.

Main Results:

• Filtering by sections significantly improved recall for disease-MRS (0.85 to 0.90) and drug-ADE (0.43 to 0.75).
• Precision also saw substantial gains: disease-MRS (0.82 to 0.92) and drug-ADE (0.16 to 0.31).
• The section-filtering approach demonstrated effectiveness in detecting specific disease-MRS and drug-ADE relations.

Conclusions:

• Selecting information based on report sections enhances the detection of disease-MRS and drug-ADE relations.
• This preliminary study highlights the utility of structured information retrieval within clinical narratives.
• Further research incorporating advanced statistical and temporal models, along with external knowledge sources, is recommended.