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

• DOI: 10.1002/alz.077327
• Publication Date: 2023-12-25T15:09:06Z
• Authors: Ryan S Falck; Teresa Liu‐Ambrose; Liisa A Galea; Roger Tam
• Keywords: Cognitive Decline

Abstract Background Circadian rhythms (i.e., the ∼24‐hour biological clock) are critical to maintenance of sleep‐wake cycle, and disturbances common in people at risk for cognitive decline dementia. Several studies have identified circadian factors associated with using actigraphy (a field measure indexing cycle). However, there currently untapped opportunities use power artificial intelligence, specifically machine learning (ML), improve our ability identify signs from data. As a first step towards this goal, we examined utility two supervised ML models predicting performance data UK Biobank study. Method A cross‐sectional analysis participants study (40‐69 years entry) valid complete (N = 49,469). Participants completed computerized versions Trail Making Test B‐A (TMT) Digit Symbol Substitution (DSST). Actigraphy were collected over 7 days, average hourly movement being indexed. Along 24‐hour data, included following features each model: age, sex, household income, educational attainment, smoking alcohol intake, ethnicity, body mass index, Townsend Deprivation Index. Seventy percent randomized training set, remaining 30% held out as test set. We developed separate predict performance: 1) linear regression approach; 2) 3‐hidden layer (40 hidden units per layer) neural network. Model accuracy was compared coefficient determination (R 2 ). Result Mean age 55 (SD 8 years) 56% female. Average TMT time 27.22 seconds 20.10 seconds), mean DSST score 20.06 5.04). Our had modest predictive 8%) 21%); network similar capability both 21%). Conclusion approaches show capability. Further work is needed how can be used function biometric