Personalised Accelerometer Cut-point Prediction for Older Adults' Movement Behaviours using a Machine Learning approach.

BACKGROUND AND OBJECTIVES: Body-worn accelerometers are the most popular method for objectively assessing physical activity in older adults. Many studies have developed generic accelerometer cut-points for defining activity intensity in metabolic equivalents for older adults. However, methodological diversity in current studies has led to a great deal of variation in the resulting cut-points, even when using data from the same accelerometer. In addition, the generic cut-point approach assumes that 'one size fits all' which is rarely the case in real life. This study proposes a machine learning method for personalising activity intensity cut-points for older adults.
METHODS: Firstly, raw accelerometry data was collected from 33 older adults who performed set activities whilst wearing two accelerometer devices: GENEActive (wrist worn) and ActiGraph (hip worn). ROC analysis was applied to generate personalised cut-point for each data sample based on a device. Four cut-points have been considered: Sensitivity optimised Sedentary Behaviour; Specificity optimised Moderate to Vigorous Physical Activity; Youden optimised Sedentary Behaviour; and Youden optimised Moderate to Vigorous Physical Activity. Then, an additive regression algorithm trained on biodata features, that concern the individual characteristics of participants, was used to predict the cut-points. As the model output is a numeric cut-point value (and not discrete), evaluation was based on two error metrics, Mean Absolute Error and Root Mean Square Error. Standard Error of estimation was also calculated to measure the accuracy of prediction (goodness of fit) and this was used for performance comparison between our approach and the state-of-the-art. Hold-out and 10-Fold cross validation methods were used for performance validation and comparison.
RESULTS: The results show that our personalised approach performed consistently better than the state-of-the-art with 10-Fold cross validation on all four cut-points considered for both devices. For the ActiGraph device, the Standard Error of estimation from our approach was lower by 0.33 (Youden optimised Sedentary Behaviour), 9.50 (Sensitivity optimised Sedentary Behaviour), 0.64 (Youden optimised Moderate to Vigorous Physical Activity) and 22.11 (Specificity optimised Moderate to Vigorous Physical Activity). Likewise, the Standard Error of estimation from our approach was lower for the GENEActiv device by 2.29 (Youden optimised Sedentary Behaviour), 41.65 (Sensitivity optimised Sedentary Behaviour), 4.31 (Youden optimised Moderate to Vigorous Physical Activity) and 347.15 (Specificity optimised Moderate to Vigorous Physical Activity).
CONCLUSIONS: personalised cut-point can be predicted without prior knowledge of accelerometry data. The results are very promising especially when we consider that our method predicts cut-points without prior knowledge of accelerometry data, unlike the state-of-the-art. More data is required to expand the scope of the experiments presented in this paper.