Meeting challenges of activity recognition for ageing population in real life settings

Recognizing Physical Activity of Older People from Wearable Sensors and Inconsistent Data

“…The current study builds upon and extends our previous work [8,12]. The workflow includes conventional steps, such as noise reduction, feature extraction, data imputation to fill in missing values, data standardization and classification with or without embedded dimensionality reduction.…”

“…Additionally, this end-to-end system is compared with three advanced CNN architectures, each of them based on different underlying assumptions (such as existence of correlation across sensors or correlation across axes within the same sensor). In more detail, the individual contributions of this paper can be summarized as follows:A SVM-based classification method is developed and assessed exclusively on older people’s recordings from a wearable sensor in everyday life conditions.Variations of the basic model are proposed to address device-relative issues that are due to data acquisition based on two different wearable devices, as well as their possible misplacement, during monitoring of physiological activity.The subject-specific prediction models of our previous work [12] are replaced with subject-independent models to avoid the laborious pre-training phase for every new-coming subject.Temporal consistency criteria are enforced to improve the predictions’ robustness.Three different convolutional neural network architectures are developed and applied to the same ADL recognition problem improving classification accuracy over our standard approach [8]. Advanced Bayesian optimization is exploited for efficient hyper-parameter tuning.…”

Tensor Decomposition for Multiple‐Instance Classification of High‐Order Medical Data

“…Equation 3denotes that each instance can be approximated as a linear combination of R two-dimensional components, V * ,r ∘ W * ,r ∈ ℝ J×K which correspond to the latent factors of the data. Thus, we can choose as features representing an instance i, the R coefficients u ir , r = 1, 2, … , R , that correspond to the ith row of factor matrix U in (2). Furthermore, we can see the latent factors V * ,r ∘ W * ,r as a high-order dictionary describing the data.…”

“…The measurements are recorded using two different devices, a fact that makes this dataset especially challenging. More details on the problem objective and the incorporated devices can be found in [1,2].…”

“…Applications that provide massive multidimensional data are vast. Some examples include monitoring patients by multisensor technologies [1,2], noninvasive lesion detection and diagnosis using hyperspectral sampling [3], cancer diagnosis based on tissue microarray data [4,5], color segmentation of skin lesions using histology-stained microscopic images [4,5], classification of EEG signals for seizure detection [6], or for Alzheimer's disease analysis [7]. The main challenge is to extract discriminative features from high-dimensional data in a way that preserves their multidimensional structure while at the same time models the interdimensions' interaction.…”