Two prediction models published in 2026 aim to tell clinicians which stroke patients face cognitive trouble, and they do not answer the same question. In The Lancet Healthy Longevity on 4 March 2026, Andrea Kusec, Kym Snell and Nele Demeyere reported a prediction model development and validation study, now presented as a Stroke Cognition Calculator, which estimates the chance of impairment across cognitive domains six months after stroke. It was built on 430 people from the Oxford Cognitive Screening Programme (OCS-Recovery) who completed the Oxford Cognitive Screen both in acute care and at six months, recruited at the John Radcliffe Hospital, Oxford, and Abingdon Community Hospital between 20 March 2012 and 9 March 2020. The overall binary model had an optimism-adjusted C-statistic of 0.76 (95% CI 0.71 to 0.80) and 0.74 (0.68 to 0.80) on external validation in a separate multi-site English cohort, OCS-Care (264 participants, mean age 68.9, mean NIHSS 2.8, reported in the authors' preprint of the same work).
The second model, published in Neurology on 12 January 2026 by Jessica Lo and colleagues for the Stroke and Cognition Consortium (STROKOG), predicts something else: dementia within five years. It pooled individual participant data from 12 studies across 10 countries, 2,663 people with a mean age of 67.0 years, 40 per cent female, followed for a median of 2.0 years (IQR 1.0 to 5.0), during which 655 developed dementia, a rate of 8.7 per 100 person-years. Dementia was diagnosed mainly by expert panel consensus or algorithm, and the analysis used Fine-Gray models that treat death as a competing event. The final model uses age, sex, education, previous stroke, diabetes and stroke severity, plus two interactions and study-level variables including national health expenditure. Discrimination was 0.81 (95% CI 0.75 to 0.87) in the full development dataset.
The distinction matters for a family at a bedside. Impairment on a stroke-specific screen at six months is common and often partly recovers; dementia is a clinical diagnosis with a different threshold and a longer horizon. The Oxford model deliberately uses a stroke-specific outcome rather than dementia screening tools such as the Mini Mental State Examination, which the authors note are poorly suited to left hemisphere stroke because they lean heavily on language. The STROKOG model is aimed at longer-term risk stratification and follow-up planning.
Both papers are honest about what is missing. The Oxford model's external check was in a milder cohort, and 436 of the 866 people screened acutely did not complete six-month follow-up (82 died), with those lost having had more severe strokes and greater acute impairment. The STROKOG model was validated by internal-external cross-validation within its own pooled studies, where the pooled C-index fell to 0.70 (0.67 to 0.73), higher in post-2010 cohorts (0.79) and European cohorts (0.74), with risks slightly overestimated in Asian cohorts; its authors state that external validation in diverse populations is still needed. Neither team claims a tested clinical benefit. Demeyere said the next step is to test how the predictions "can be used responsibly to improve follow-up and support for patients and families".