2025-12-12

Our paper, “Using Novel Natural Language Processing Approaches to Examine Age-Friendly Communication about Nursing Nome Residents with Dementia | The Gerontologist | Oxford Academic” has been published in The Gerontologist. This work builds directly on the [4M taxonomy](Development and evaluation of a 4M taxonomy from nursing home staff text messages using a fine-tuned generative language model - PubMed) I developed last year from fine-tuning a local LLM on NH staff texts. That one ended 2024 on a high note, teaching me volumes about embeddings, clustering, and reproducible NLP pipelines. We’ve taken those extracted 4Ms (what matters, medications, mentation, mobility) and asked: do they predict avoidable NH-to-hospital transfers, especially for folks with Alzheimer’s Disease and Related Dementias (ADRD)?

Nursing homes are a critical domain for aging America’s challenges. Nearly half of residents have ADRD, and avoidable transfers to hospitals can lead to poorer outcomes. We merged 30k+ text messages from 16 Missouri NHs in the MOQI CMS demo (2016-2020) with 3,687 transfer events. Focused on 1,031 observations timed pre-transfer (0-28+ days out).

Models compared:

• Supervised ML: Our 4M-EE pipeline with possibilistic KNN on embeddings
• Fine-tuned Gemma 2 LLM: Local, secure, taxonomy builder.

Validated against expert gold standard, picked the best model via AIC.

Findings:

Caveats: Missouri-only, transfers-only sample, ADRD staging limits. This is the kind of Nightingale-Lovelace fusion I love to work on. Grateful for the team, NIA funding (R01AG078281), and Mizzou’s support.