Pull up a real dementia chart four years in and it does not look like data. It looks like prose: a SOAP note from a Tuesday clinic, an MMSE score dropped mid-paragraph, a line about a medication switch that nobody flagged, a daughter's remark about the stove recorded in the subjective section and never coded anywhere. A neurologist reads this by flipping backwards through visits and holding the arc in their head. That arc is rarely a clean downhill line; people have good months, infections that look like decline, and quiet plateaus. A preprint posted to arXiv on January 6, 2026, and revised on March 20, asks whether a language model can be trained to do that flipping-backwards, and then to show its working.
The system is called Dementia-R1, from Choonghan Kim, Hyunmin Hwang, Hangeol Chang, Jaemin Kim, Jinse Park, Jae-Sung Lim and Jong Chul Ye. It is a 7 billion parameter model, small by current standards, trained with reinforcement learning in two stages. The first stage is the interesting one: instead of rewarding the model only for getting a final diagnosis right, which is a thin and distant signal, the authors first reward it for predicting verifiable clinical indices (MMSE, scored 0 to 30; GDS, 1 to 7; CDR, 0 to 3) from the history that came before. Learn to forecast the next score, then learn to forecast the person. The training notes came from about 3,000 patients with neurocognitive disorders seen at Asan Medical Center in Seoul between January 1, 2021 and September 30, 2023, selected on ICD-10 codes covering Alzheimer's disease, vascular dementia and mild cognitive impairment. Splits were made at patient level, only notes written before the prediction anchor were visible, and anything less than six months ahead was excluded so the model could not simply read a diagnosis that had effectively already been made.
On that hospital's held-out notes the model reached an AUROC of 84.02 percent. AUROC is worth translating, because it is not an accuracy figure a family can apply to themselves: it means that if you pick one patient who went on to meet the dementia label and one who did not, the model ranks them in the right order roughly 84 times out of 100. It says nothing about any individual. Tested on an entirely different hospital, Haeundae Paik, using longitudinal records of Parkinson's disease patients (ICD-10 G20, 2010 to 2024) to predict Parkinson's disease dementia, it reached 78.37 percent, ahead of a general model ten times its size. On ADNI, the long-running public research dataset, flattened into chronological text logs, it reached 83.17 percent, the best among the language model baselines tested, with an F1 of 74.91 percent against GPT-4o's 76.05 percent.
For this section, the output matters more than the score. The paper's worked example shows the model reconstructing a patient's MMSE trajectory in order (23, then 21, then 17, then 17) and correctly noting a switch between two named drugs, laid out in separate sections for cognitive history, diagnosis and current status. The contrast case is the warning. A much larger general-purpose model, given the same note, wrote "Ebixa (donepezil)" when Ebixa is memantine, a different drug class entirely, and read GDS scores of 3 to 4 as mild depression, confusing a dementia staging scale with a mood inventory. That is exactly the kind of error that sounds fluent, survives a skim, and lands in a summary a family might later be handed. Two board-certified neurologists ran blinded head-to-head comparisons across both hospitals, 10 cases per model pairing for 100 judgment pairs in total, scoring temporal reasoning, evidence grounding, evidence selection, medical soundness and overall clinical utility. One hundred pairs and two readers is a pilot, and should be read as one.
The authors are unusually frank about where the floor is soft. The clinical scores used as training rewards were not pulled from structured fields; they were extracted from free text by another language model, and a clinician audit of 200 randomly sampled notes (600 score slots) found 98.5 percent slot-level accuracy, with failure modes including an extractor that grabbed a Barthel Index number sitting near the GDS line. The Korean notes were translated into English before training, and the paper concedes that subtleties of language, the very hesitations and syntax errors that can themselves signal cognitive change, may not survive the crossing. All of the free-text data came from one country and one ethnic group, and from two hospitals that document in their own house style. The target label is a recorded clinical diagnosis, not a biological truth.
What this is, then, is a plausible engine, tested backwards on records that already exist, not a tool anyone has tried in a live clinic. The appeal is obvious to anybody who has ever assembled a caregiver's diary and a stack of visit summaries and wished someone would read all of it in one sitting. The distance between that wish and a deployed system runs through prospective testing, other languages, other health systems, and the much harder question of what a clinician does when a confident paragraph of machine reasoning disagrees with the person sitting in front of them.