Ranking Medical Terms to Support Expansion of Lay Language Resources for Patient Comprehension of Electronic Health...

Area of Science:

• Natural Language Processing (NLP)
• Health Informatics
• Biomedical Informatics

Background:

• Medical jargon in electronic health records (EHRs) hinders patient understanding.
• Existing lay language resources for EHR terms are limited and costly to expand.
• Prioritizing key EHR terms for annotation is crucial for improving patient comprehension.

Purpose of the Study:

• To develop an NLP system, adapted distant supervision (ADS), for ranking EHR terms.
• To prioritize high-ranking EHR terms for lay language annotation, enhancing patient comprehension.
• To address the scarcity of public domain lay language resources for EHRs.

Main Methods:

• Utilized distant supervision from consumer health vocabulary and transfer learning for the ADS system.
• Investigated feature space augmentation and supervised distant supervision transfer learning algorithms.
• Incorporated distributed word representations from EHR data and expert-annotated importance for 6038 candidate terms.

Main Results:

• The ADS system with feature space augmentation achieved an average precision of 0.850 with 1000 training examples.
• The ADS system with supervised distant supervision achieved an average precision of 0.819 with only 100 training examples.
• Both ADS systems significantly outperformed baseline systems (P<.001), with rich features enhancing performance.

Conclusions:

• The adapted distant supervision (ADS) system effectively ranks EHR terms for comprehension.
• Transfer learning enhances ADS performance, even with limited training data.
• ADS-prioritized terms facilitated the expansion of lay language resources, improving patient EHR understanding.