Anyone who has sat in a geriatrics clinic while a doctor scrolls knows the problem. The chart is enormous and mostly says the same thing twice. In a cross-sectional analysis of 104,456,653 routinely generated clinical notes published in JAMA Network Open on September 26, 2022, Jackson Steinkamp and colleagues found that 16,523,851,210 words, 50.1 percent of all the text, had been duplicated from prior documentation. Half the record is an echo. Somewhere inside the other half, often in a single sentence a nurse typed on a Tuesday, sits the detail that matters: the daughter who has stopped filling the prescriptions, the water glass that is always full, the appointments that quietly stopped happening.
A research group at UTHealth Houston set out to see whether a language model could find those sentences. In an abstract by Jessica Hernandez-Chilatra, Carolyn Pickering and Xiaoqian Jiang, published in the Gerontological Society of America's journal Innovation in Aging on December 31, 2025, the team analyzed 379 clinical notes from 18 people with dementia. They first used MPNet sentence embeddings and cosine similarity to identify redundancy, flagging 1,146 highly similar notes, then grouped what remained at the patient level. GPT-4 was then asked to pull structured insights out of that narrative pile, sorting neglect risk into three domains, medical, emotional and financial, and attaching caregiving recommendations. A final report summarized common patterns, risk factors and possible interventions.
What came back is familiar to anyone who works this beat, which is itself a kind of signal. On the medical side, the recurring patterns were unmanaged pain, dehydration and medication nonadherence. On the emotional side, social isolation and distress behaviors that nobody had addressed. On the financial side, healthcare that was missed because the money was not there. Running underneath all three, the authors report, were caregiver burnout, gaps in follow-up care, and an absence of support. The model was not discovering a hidden syndrome. It was reading the chart the way an unhurried clinician would, if an unhurried clinician existed.
That is the argument for bothering. Mistreatment of seniors is caught rarely and late. In a retrospective cross-sectional analysis of 2012 national emergency department data published in the Journal of the American Geriatrics Society in January 2017, Christopher Evans and colleagues found a diagnosis of abuse recorded in an estimated 3,846 visits out of roughly 29 million by people aged 60 and older, a weighted prevalence of 0.013 percent, with neglect the most commonly diagnosed type at 32.9 percent of cases. Compare that with what families themselves report. A systematic review and meta-analysis abstract by Fei Sun and colleagues, published in Innovation in Aging on December 21, 2023, put pooled prevalence among people with dementia at 43 percent for emotional abuse and 16 percent for neglect. In a 21-day diary study reported by Pickering, Maria Yefimova and Christopher Maxwell in November 2019, 74 percent of dementia family caregivers self-reported at least one abuse or neglect behavior and 52 percent reported more than one form.
Those numbers deserve a careful reading, because the word neglect does heavy and sometimes unfair work. In this research tradition it usually describes unmet need, a dose skipped because the day fell apart, a bath that did not happen because the person refused and the caregiver had nothing left. It is a description of a household under strain, not a verdict on a family's character. That distinction is exactly what a triage flag can flatten, and it is the thing most worth watching as tools like this move from abstract to clinic. A summary that reads "financial neglect risk" in a physician's inbox carries weight that a line in a nursing note does not.
And the evidence here is early by any measure. This is a conference abstract, not a full peer-reviewed paper, covering 18 patients at a single academic health system. The authors report no accuracy rate, no comparison against clinician chart review or any gold standard, no false positive count, and no test of whether a single flag changed a single care plan. Grouping notes by similarity is a sensible way to cut through duplication, but the underlying record is still only what clinicians happened to write down, and copy-forward text can propagate an error as faithfully as a fact. Large language models also produce confident, well-structured output whether or not the source supports it, which is precisely the failure mode you cannot detect in a sample of 18 without a reference standard.