DOI: 10.1002/alz.077327 ISSN: 1552-5260

Can we use machine learning to predict cognitive performance from actigraphy data? Preliminary results from the UK Biobank Study

Ryan S Falck, Teresa Liu‐Ambrose, Liisa A Galea, Roger Tam
  • Psychiatry and Mental health
  • Cellular and Molecular Neuroscience
  • Geriatrics and Gerontology
  • Neurology (clinical)
  • Developmental Neuroscience
  • Health Policy
  • Epidemiology



Circadian rhythms (i.e., the ∼24‐hour biological clock) are critical to the maintenance of the sleep‐wake cycle, and sleep‐wake disturbances are common in people at risk for cognitive decline and dementia. Several studies have identified circadian factors associated with cognitive decline using actigraphy (a common field measure for indexing the sleep‐wake cycle). However, there are currently untapped opportunities to use the power of artificial intelligence, specifically machine learning (ML), to improve our ability to identify signs of cognitive decline from actigraphy data. As a first step towards this goal, we examined the utility of two supervised ML models for predicting cognitive performance using data from the UK Biobank study.


A cross‐sectional analysis of participants in the UK Biobank study (40‐69 years at entry) with valid actigraphy data and complete cognitive data (N = 49,469). Participants completed computerized versions of Trail Making Test B‐A (TMT) and Digit Symbol Substitution Test (DSST). Actigraphy data were collected over 7 days, with average hourly movement being indexed. Along with 24‐hour actigraphy data, we included the following features in each model: age, biological sex, household income, educational attainment, smoking and alcohol intake, ethnicity, body mass index, and Townsend Deprivation Index. Seventy percent of participants were randomized to the training set, with the remaining 30% held out as a test set. We developed two separate ML models to predict cognitive performance: 1) a linear regression approach; and 2) a 3‐hidden layer (40 hidden units per layer) neural network. Model accuracy was compared using the coefficient of determination (R2).


Mean age was 55 years (SD = 8 years) and 56% of participants were female. Average TMT time was 27.22 seconds (SD = 20.10 seconds), and mean DSST score was 20.06 (SD = 5.04). Our supervised linear regression had modest predictive ability of TMT (R2 = 8%) and DSST performance (R2 = 21%); the 3‐hidden layer neural network had similar predictive capability for both TMT (R2 = 8%) and DSST (R2 = 21%).


ML approaches for predicting cognitive performance using actigraphy data show modest capability. Further work is needed to identify how ML can be used to predict cognitive function from biometric data.

More from our Archive