The typical study in this field lasted about as long as a holiday. Across 49 papers pooled by a systematic review published on February 23, 2026 in the Journal of Medical Internet Research, the median wear time for a wrist device was seven days, with some studies running only three or four days and just one longitudinal cohort stretching to 182.5 days. Seven days of movement and rest, converted into features, fed into a model, and asked to say something about a brain that may have been changing for a decade. That gap between the length of the recording and the length of the disease is the quiet fact at the centre of this stocktake.
The review, by Ander Cejudo, Markel Arrojo, Cristina Martin and Aitor Almeida of the Vicomtech Foundation, the University of Deusto and the Biodonostia Health Research Institute in Spain, searched PubMed, Scopus, the ACM Digital Library and SpringerLink for work published between January 2020 and December 1, 2025. After screening 5590 records, 49 studies survived. Together they covered more than 200,000 participants, but the distribution is lopsided: sample sizes ran from 14 to 91,948, with a median of just 145, and five very large cohorts dragged the averages upward. Populations were mostly people with dementia (35 of 49 studies), people with mild cognitive impairment (29 of 49) and healthy comparison groups (36 of 49). Only four studies included people with subjective cognitive decline, which is to say the group most likely to be wearing a watch and worrying.
The hardware picture also differs from the one the consumer market suggests. Research grade actigraphy dominated, appearing in 43 of the 49 studies (87.8 percent), with commercial wearables in only 7 (14.3 percent). Broken down by device class, actigraphs accounted for 29 studies, accelerometers for 12, smartwatches for 7 and multimodal research devices, adding signals such as heart rate or electrodermal activity, for 2. The review notes that differences in device, sensor and preprocessing mean the resulting measures do not map onto a common estimand: a sleep fragmentation number from one wristband is not the same quantity as a sleep fragmentation number from another. For a family, that is the difference between a research finding and a reading on a screen at home.
What about the AI in the title? Less of it than the framing implies. Analysis was predominantly conventional statistics in 36 of 49 studies (73.5 percent), with machine learning in 7 (14.3 percent) and deep learning in 6 (12.2 percent), plus two studies using clustering. The machine learning studies had a median of 99 participants; the deep learning studies a median of 108. Where models were built, they used logistic regression, decision trees, support vector machines, random forests and gradient boosting, and on the deep learning side neural networks, convolutional networks, generative adversarial approaches and autoencoders. Reported classification performance, area under the curve, sat roughly between 0.70 and 0.95, and several studies forecasting clinical Alzheimer disease onset from motor activity features reported C-index values of about 0.80 to 0.84. Those are numbers from within the studies that built the models, not from independent deployment.
The review's own critical read is the useful part. Most included studies were rated low risk of bias (34 of 49), 13 medium and 2 high. But 38 of the 49 contributed what the authors class as indirect evidence: associations between sleep, circadian rhythm, activity or physiological measures and cognitive outcomes, rather than direct predictive tests. Only about a quarter, 13 of 49 (26.5 percent), addressed early detection or prevention through longitudinal risk estimation or predictive modelling. Designs split almost evenly between cross-sectional analyses (21 studies) and longitudinal cohorts (24 studies), some with follow-up extending 8 to 15 years, and the outcomes they predicted were themselves mostly brief screening tests: the Mini-Mental State Examination in 18 studies, the Montreal Cognitive Assessment in 7, ICD-10 coded clinical diagnoses in 7. The authors name the field's limits plainly: small samples, short monitoring, and limited external validation. They also flag their own, including English language restriction and the omission of databases such as Embase and PsycINFO.
Read in the register that matters at a kitchen table, this is neither a breakthrough nor a debunking. Disrupted sleep, fragmented rest and activity rhythms and irregular daily movement do track with worse cognitive outcomes, consistently, with modest to moderate effect sizes. That is a real signal, and it is exactly the lived-context signal a wrist device is well placed to record. What the evidence base has not yet built is the scaffolding that would turn signal into a test: agreed measures across devices, monitoring long enough to see a trajectory rather than a week, and models validated in people other than the ones they were trained on.