Every dementia diagnosis has a paper trail that starts long before anyone says the word. A note about a missed appointment. A pharmacy record that stops making sense. A line in a consultation summary that reads "family reports forgetfulness", typed by a general practitioner who had eight minutes and moved on. For years the promise of the electronic health record has been that the trail is already there, waiting to be read by software. The harder question is which parts of the record actually carry the signal, and that is the question a team led by Joyce Siette at Western Sydney University set out to answer.
Their systematic review, published on 6 May 2026 in Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring, searched six databases up to April 2025 and included 50 original studies that used routinely collected health data to identify dementia. The team sorted every data element those studies had used into ten health data domains. The tally reads like an audit of habit rather than of evidence. Demographics, meaning little more than age and sex, appeared in 48 of the 50 studies (96 per cent). Medications and medical codes appeared in 25 (50 per cent). Cognitive symptoms, the domain that includes confusion, disorientation, forgetfulness, language difficulty, memory loss, misplacing items, poor judgement and repetitive actions, appeared in 19 (38 per cent). Physical symptoms were in 17 (34 per cent), pathology testing and management in 15 (30 per cent), mental health symptoms in 13 (26 per cent), and behavioural symptoms, covering aggression, agitation, personality change, wandering and withdrawal, in only 12 (24 per cent).
The review's central finding is the mismatch between that ranking and what performed. Cognitive and behavioural domains, two of the less commonly used, contributed most to specificity, with reported specificity across the included studies ranging from 57.5 per cent to 99.9 per cent. Specificity is the unglamorous half of detection: it is the measure of how well a tool leaves alone the people who do not have the condition. A chart-based flag with poor specificity is a machine for generating frightened phone calls. The domains best placed to prevent that, according to this synthesis, are the ones describing what a person is actually doing and thinking, not their date of birth.
Where the data came from mattered too. Primary care records were the most used source, in 24 of the 50 studies (48 per cent), followed by research registries (12 studies), tertiary care systems (10) and secondary care systems (8), with the work concentrated in the United Kingdom and the United States. Designs were mostly case-control and retrospective cohort studies, and logistic regression was the common workhorse. Structured cognitive and neuropsychological assessments such as the Mini-Mental State Examination, the Addenbrooke's Cognitive Examination and the Dementia Rating Scale were used in 9 of the 50 studies (18 per cent). Notes written by general practitioners, old age psychiatrists or memory clinics featured in 6 (12 per cent). Neuroimaging such as MRI or PET appeared in 3 (6 per cent), SNOMED-CT codes in 2 (4 per cent), and unstructured clinical text, the free writing where clinicians put the things that do not fit a tick box, in 2 (4 per cent). The review also reports that models built specifically for Alzheimer's disease had higher mean accuracy than general dementia models, 74.6 against 67.1.
Then comes the part of the paper that should be read slowly by anyone selling a chart-based detection tool. Of the 50 studies, only 18 had extractable accuracy data at all. Sensitivity was reported in 16 studies (32 per cent), specificity in 14 (28 per cent), area under the curve in 8 (16 per cent), positive predictive value in 7 (14 per cent), and negative predictive value in just 4 (8 per cent). The spread in the figures that were reported was wide: sensitivity from 29.9 per cent to 100 per cent, positive predictive value from 2.1 per cent to 88.0 per cent, area under the curve from 71 per cent to 94.4 per cent. The authors report these as ranges rather than pooled estimates, and they conclude that future work needs validation, standardisation and clinical implementation. In other words, the field has a map of which fields to mine and very little agreement on how well the mining works.
For a family, none of this changes what happens in a consulting room next week. What it changes is the shape of the argument behind the screen. If the strongest signals sit in cognitive and behavioural fields, then the value of the record depends on whether anyone writes those things down in a place software can find. The person who notices the wandering, the withdrawal, the repeated question, is usually not a clinician; it is whoever is at home. That observation reaches the chart only if it is said out loud and recorded, which is an argument for telling the general practitioner the small things, and for asking that they go in the notes.