Can natural language processing improve the completeness of immunization data?

Challenge

The VISION surveillance collaboration aims to support public health efforts by routinely collecting and synthesizing respiratory illness data and publishing timely statistical analyses, including vaccine effectiveness (VE).

Surveillance methods often rely on the structured data elements within the electronic health record (EHR) to identify key health characteristics, procedures, and conditions. However, in some instances, details may only be documented in the narrative, text-based sections of the EHR.

Solution

Rule-based natural language processing (NLP) methods have the potential to identify health conditions and procedures using a dictionary of key terms and pattern-matching techniques. We developed NLP rules that leveraged grammatical dependencies within the text and accounted for a range of text variations using a combination of synthetic and publicly available data. Our rules aimed to identify immunizations for respiratory infections, influenza, and respiratory syncytial virus.

We applied the rules to a large sample of individuals from the EHRs of a large health care system and measured performance by conducting a manual review of individual notes, assessing concurrence with structured data, and comparing performance with prior methods.

Results

We were able to identify additional immunizations for respiratory infections, influenza, and RSV that were not present in the structured data. This research demonstrated the feasibility and utility of NLP methods to enhance structured vaccine information from EHRs and served to validate the information contained within the structured EHR data.