(176) Falling in Love with Nursing Home Electronic Health Record (EHR) Data: Greater Understanding of The Nature of Falls from Harnessing Information in Nurses' Notes

Publication Number: 850

Background: Falls are the leading cause of injuries in older adults. Some medications (e.g., antipsychotics, antidepressants) have been associated with increased fall risk. Standardized coding of falls in healthcare claims and/or Minimum Data Set assessments provides little information regarding the events surrounding falls. Unstructured EHR data may offer a novel opportunity to obtain this information and improve fall prevention strategies.

Objectives: To evaluate the extent to which nurses’ EHR notes related to fall incidents provide contextual and environmental information regarding the nature of falls in nursing homes.

Methods: Data from 51 nursing homes in a large EHR system included in the National Institute on Aging Long-Term Care Data Cooperative (2018-2021) were used for the analysis. Nurses’ notes describing fall events were identified though a keyword search for “fall” or “fell”. Natural language processing (NLP) techniques were used to extract prominent latent constructs derived from the co-occurrence of words. Weighted unigram/bigram frequencies were calculated for each note (after removing “the”, “a”, etc.), then inversely weighted by the total number of notes in which they occur (TF-IDF method). This process calculates the informativeness of a word for a specific note. Latent constructs were identified through an iterative optimization process known as non-negative matrix factorization. This method assigns scores to both words and notes. Notes scoring high on one construct are likely to contain words also scoring high on that construct and vice versa. Two reviewers independently examined high-scoring words and read the corresponding full nurses’ notes. After discussing their interpretations, they were able to achieve consensus on the construct labels.

Results: A total of 9,104 unique fall events were identified. The first 10 constructs corresponded to three stages of fall response events: 1) fall context (i.e., wheelchair, bed); 2) post-fall checks (i.e., injury, neurological, well-being); and 3) post-fall actions (i.e., assisted transfers, escalation to supervisors, root-cause analysis). Medications were notably absent from the top 10 constructs identified.

Conclusions: Nurses’ notes provide valuable contextual information about fall events which may be useful in developing novel fall prevention strategies in nursing homes. Future research includes extending the use of NLP to further isolate and categorize fall response stages. Topic categories can also be compared with structured data (e.g., muscle weakness, medication) to further demonstrate the value-add of EHR nurses’ notes.