0438 Automatic Nighttime Agitation and Sleep Disruption Detection Using a Wearable Ankle Device and Machine Learning. (27th May 2020)

• Record Type: Journal Article
• Authors: Kumar, R; Feltch, C; Richards, K; Morrison, J; Rangel, A; Janney, R; Shayesteh, S; Allen, R; Banerjee, N
• Is Part Of: Sleep. Volume 43(2020)Supplement 1
• Page Start: A168
• Page End: A168
• Publication Date: 2020-05-27
• Subjects: Sleep -- Physiological aspects -- Periodicals; Sleep disorders -- Periodicals; Sommeil -- Aspect physiologique -- Périodiques; Sommeil, Troubles du -- Périodiques; Sleep disorders; Sleep -- Physiological aspects; Sleep -- physiological aspects; Sleep Wake Disorders; Psychophysiology; Electronic journals; Periodicals; 616.8498
• DOI: 10.1093/sleep/zsaa056.435
• Languages: English
• ISSNs: 0161-8105

Abstract

Abstract: Introduction: Nighttime agitation behavior such as wandering and restlessness during awake and sleep in people with Alzheimer's disease (AD) is expensive to manage and adversely affects sleep. Nighttime agitation is mostly noted by subjective caregiver reports. An automated process for this assessment would improve clinical management. Here we report on the RestEaZe TM system that uses an ankle band and machine learning to automatically classify sleep status and nighttime agitation behaviors in older adults with AD. Methods: We collected data on 7 adults (mean: 81 years, SD: 10.6) with AD. They wore the RestEaZe TM ankle band with a 3-axis accelerometer, a 3-axis gyroscope, and three textile capacitive sensors. A trained Research Assistant (RA) continuously observed for wandering, restlessness, wake, and sleep between 5pm and 7am using the Cohen Mansfield Agitation Inventory (CMAI). We merged, and band-pass filtered the data and divided it into 10-second non-overlapping windows. CMAI labels and time-series features (scaled using StandardScaler) extracted from the RestEaZe TM data were used to train a Random Forest binary classifier. The significant features were extracted based on the impact on the p-value for the classifier. We used the Synthetic Minority Oversampling Technique (SMOTE) to balance the dataset and performed 5-fold cross-validation with a 67-33 train-test split. Results: We report the sensitivity, specificity, accuracy, and Area-under-the Curve (AUC) for the ROC curve for the classifiers: (1) Sleep/Awake: sensitivity=0.95, specificity=0.87, accuracy=0.92, AUC=0.97; (2) Wandering/Non-Wandering: sensitivity=0.85, specificity=0.99, accuracy=0.98, AUC=0.99; and (3) Restless/Non-Restless: sensitivity=0.84, specificity=0.84, accuracy=0.84, AUC=0.92. The significant features were related to the intensity of movements. Conclusion: Our preliminary results show the feasibility of using RestEaZe TM for quantitatively measuring nighttime agitation. These can provide clinically useful objective measures of agitation that can be automatically transmitted to clinical or research records with minimal staff time requirements. Support: The authors acknowledge the funding support from the National Institute on Aging under award R01AG051588 and Arbor Pharmaceuticals for support for Horizant and the matching placebo.