702 - Enhancing surveillance of pressure injuries using NLP and clinical notes

Pressure injuries (ulcers) are often related to immobility or poorly fitting casts or other medical equipment or devices. This is an important patient safety event that is monitored and reported as a Healthcare Acquired Condition (HAC). The primary data source for surveillance is the system-wide collection of provider self-reports.  

Objective:

Assess the feasibility of using nursing handoff notes to identify under-reported pressure injury events.

Design/Methods: We extracted a corpus of nursing handoff notes across all medical inpatient and ICUs in a tertiary care pediatric center.  We have established a workflow for data extraction using Natural Language Processing (NLP) assisted manual review process, trained by two domain experts (an RN and an MD).  We begin with the use of keywords relevant to pressure injuries and treatments, then expand with regular expressions (RegEx), distributive semantics, and N-gram document classifier, and finally a validated random forest model which combines the RegEx and N-gram models.  

Results: During the study period there were N=70,981 notes available for review.  Of these, 69,436 notes (97.5%) were nursing handoff notes, 1,091 (1.5%) were pain assessment notes, 393 (0.4%) were addenda for event nursing, and 64 (0.1%) were ICU event nursing notes. We illustrate a note with highlighted RegEx in Figure 1. 

Our combined model had a sensitivity of 95% and specificity of 71%. Metrics of our models are presented in Table 1. 

Our system identified n=618 notes addressing pressure injuries, of which 336 were classified as known events while 282 were flagged of which, we identified 61 new/first described cases while the remainder are repeated documentation of an index event. 

Conclusion(s): NLP-assisted review is a feasible method for surveillance of pressure ulcers within a hospital setting which complements and enhances other data sources.