Objective assessment of sleep parameters using multimodal AX3 data in older participants

Background Sleep disturbances are both risk factors for and symptoms of dementia. Current methods assessing sleep largely based on either polysomnography (PSG) which is costly inconvenient, or self- care-giver reports prone to measurement error. Low-cost monitor longitudinally at scale can be useful symptom development. Here, we develop deep learning models that use multimodal variables (accelerometers temperature) recorded by the AX3 accurately identify wake epochs derive parameters. Method Eighteen men women (65-80y) participated in a laboratory-based study multiple devices monitoring were evaluated. PSGs over 10-h period scored according established criteria per 30 sec epochs. Tri-axial accelerometers temperature signals captured with an Axivity AX3, 100Hz 1Hz, respectively, throughout 19-h period, including concurrent PSG recording 9-h wakefulness. We developed evaluated supervised algorithm detect determine parameters from raw data. validated our results gold standard measurements compared Biobank accelerometer analysis toolbox. Single modality (accelerometer multimodality (both signals) approaches using 3-fold cross-validation. Result The proposed model outperformed baseline such as toolbox conventional machine classifiers (Random Forest Support Vector Machine) up 25%. Using data improved classification performance (up 18% higher) single modality. In terms parameters, approach boosted accuracy estimations 11% average Conclusion older adults without dementia, combining allows satisfactory quantification This holds promise behaviour deriving accurate objectively low-cost wearable sensor. A limitation current participants healthy adults: future work will focus people living