Clinical Judgement Study using Question Answering from Electronic Health Records

Area of Science:

• Pharmacovigilance and Artificial Intelligence
• Clinical Informatics
• Computational Medicine

Background:

• Identifying causal relationships between medications and adverse drug reactions (ADRs) is crucial for patient safety.
• Traditional methods rely on manual expert review of electronic health records (EHRs), which is time-consuming and resource-intensive.
• Automating this process can significantly enhance the efficiency and scalability of ADR detection.

Purpose of the Study:

• To develop and evaluate an end-to-end deep learning question answering model for automatically inferring causal relations between medications and ADRs.
• To leverage EHR data for automated ADR identification, reducing reliance on manual chart reviews.
• To assess the model's performance using a subset of the Naranjo questionnaire criteria.

Main Methods:

• An end-to-end deep learning model employing multi-level attention mechanisms was designed.
• The model was trained to answer specific questions from the Naranjo questionnaire using EHR data.
• Local and global context within the EHRs were incorporated to improve relation inference.
• An ablation study was conducted to validate the contribution of contextual information.

Main Results:

• The proposed deep learning model achieved a macro-weighted F-score ranging from 0.4598 to 0.5142 across selected Naranjo questionnaire questions.
• An overall F-score of 0.5011 was obtained, indicating moderate performance in identifying causal ADR relationships.
• The ablation study confirmed the significance of both local and global context for the model's performance.

Conclusions:

• Deep learning question answering models show promise for automating the detection of causal relationships between medications and ADRs.
• The developed model offers a potential alternative to traditional manual chart reviews, improving efficiency in pharmacovigilance.
• Further research and model refinement are warranted to enhance accuracy and clinical utility.