The use of natural language processing of infusion notes to identify outpatient infusions

Abstract

Purpose: Outpatient infusions are commonly missing in Veterans Health Affairs (VHA) pharmacy dispensing data sets. Currently, Healthcare Common Procedure Coding System (HCPCS) codes are used to identify outpatient infusions, but concerns exist if they correctly capture all infusions and infusion-related data such as dose and date of administration. We developed natural language processing (NLP) software to extract infusion information from medical text infusion notes. The objective was to compare the sensitivity of three approaches to identify infliximab administration dates and infusion doses against a reference standard established from the Veterans Affairs rheumatoid arthritis (VARA) registry.

Methods: We compared the sensitivity and positive predictive value (PPV) of NLP to that of HCPCS codes in identifying the correct date and dose of infliximab infusions against a human extracted reference standard.

Results: The sensitivity was 0.606 (0.585-0.627) for HCPCS alone, 0.858 (0.842-0.873) for NLP alone, and 0.923 (0.911-0.934) for the two methods combined, with a PPV of 0.735 (0.716-0.754), 0.976 (0.969-0.983), and 0.957 (0.948-0.965) for each method, respectively. The mean dose of infliximab was 433 mg in the reference standard, 337 mg from HCPCS, 434 mg from NLP, and 426 mg from the combined method.

Conclusions: HCPCS codes alone are not sufficient to accurately identify infliximab infusion dates and doses in the VHA system. The use of NLP significantly improved the sensitivity and PPV for estimating infusion dates and doses, especially when combined with HCPCS codes.

Keywords: Healthcare Common Procedure Coding System; computerized medical records systems; natural language processing; pharmacoepidemiology.