Sammendrag
Background: Physical activity monitoring combined with machine learning (ML) methods can contribute to more detailed knowledge about physical behavior. However, extant models are typically developed and validated on datasets from young, healthy adults. It is an open question to what extent such models accurately identify daily living activities in heterogenous older adults. This study 1) evaluates the performance of an existing activity type recognition ML model, based on data from healthy adults and classifying activity categories with high accuracy (Bach et al., 2022), in a sample of fit-to-frail older adults, and 2) uses the sample dataset to further develop the ML model. Methods: The sample included 18 older adults aged 70-95 years (79.6±7.6 years; 50% female) with a wide range of physical function, including 5 participants using walking aids. Activity was recorded using two Axivity AX3 accelerometers (thigh and lower back) and a chest-mounted camera pointing downwards during a semi-structured free-living protocol that included repetitions of walking, standing, sitting, and lying. Video recordings were labelled according to pre-defined activity definitions and were used as gold standard. Results: The overall accuracy of the original model on this dataset was 85% (87% without walking aids and 80% with walking aids), which improved to 94% (94% without walking aids and 92% with walking aids) after data of older adults was included in the training. Conclusions: Classification of daily physical behavior in older adults was considerably more accurate when the model was trained on data from older adults, especially for the frailest participants using walking aids. This improvement illustrates that it is necessary to use training data sets that are representative for the population of interest. The resulting validated ML model for fit-to-frail older adults may contribute to accurate and detailed knowledge of daily physical behavior that is essential for future research.
Vis fullstendig beskrivelse
