Question Answering for Electronic Health Records: Scoping Review of Datasets and Models.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Health Data Science

Background:

• Question answering (QA) systems for patient data aid clinicians and patients by improving decision-making and understanding of medical history.
• Electronic Health Records (EHRs) contain substantial patient data, making EHR QA a critical research area distinct from QA on medical literature.
• The unique data formats and modalities in EHRs necessitate specialized research approaches for effective QA.

Purpose of the Study:

• To conduct a methodological review of existing research on QA for EHRs.
• To identify and analyze EHR QA datasets and state-of-the-art methodologies.
• To compare evaluation metrics and identify challenges in EHR QA.

Main Methods:

• Systematic literature search from January 2005 to September 2023 across four digital databases (Google Scholar, ACL Anthology, ACM Digital Library, PubMed).
• Screening of 4111 identified papers following PRISMA guidelines, resulting in 47 selected papers.
• Classification of selected studies into 'EHR QA datasets' and 'EHR QA models' categories.

Main Results:

• The review identified 47 papers on EHR QA, with most work being recent and indicating a nascent research field.
• emrQA is the most cited and used EHR QA dataset; MIMIC-III and n2c2 datasets are popular EHR databases.
• Analysis covered various EHR QA models, their methodologies, and evaluation metrics.

Conclusions:

• EHR QA research faces challenges including limited clinical annotations and concept normalization.
• Generating realistic EHR QA datasets remains a significant hurdle.
• This review identifies research gaps and suggests future directions for EHR QA.