Assessing physical activity and functional fitness level using convolutional neural networks

Abstract

Older adults are related to a reduction in physical functionality, as a result of a musculoskeletal system degeneration. In that way, physical exercise has been stated as a suitable intervention to prevent such health problems. Therefore, an adequate assessment of the physical activity and functional fitness levels is needed to plan the individualized intervention. A broad test used to assess the functional fitness level is the 6-minutes walk test (6MWT). It has been previously measured using accelerometer sensors. In views of this background, the main aim of the present study is to use deep learning to extract automatically and to predict the physical activity and functional fitness levels of the older adults through the acceleration signals recorded by a smartphone during the 6MWT. A total of 17 participants were recruited. Anthropometric measurements (weight, height, and body mass index), physical activity, and functional fitness levels from each participant were recorded. Consecutively, two deep learning-based methods were applied to determine the prediction. According to the results, the proposed method can predict physical activity and functional fitness levels with high accuracy, even using only one cycle. Thus, the approach described in the present work could be implemented in future mobile health systems to identify the physical activity profile of older adults.

Introduction

One of the most problematic conditions in older adults population is the fragility [1]. According to a relatively recent study [2], the global prevalence of fragility was found from 4.9% to 27.3%, and prefragility was established from 34.6% to 50.9%. In views of this framework, it is logical that fragility has been considered as a prior public health problem [3], [4]. In that way, older adults are associated with a reduction in the functionality of the musculoskeletal system. It is related to a bone mass and strength loss, and a decrease in hormone production [5]. Therefore, all these physiological changes involve an increase in the risk of developing different clinical conditions [6].

Besides, physical activity-based programs seem to be an effective intervention to prevent fragility [7], [8], [9]. Nevertheless, to know the functional level of the older adults, and consequently plan the intervention, there are several tests. One of the most used tests to evaluate the functional exercise capacity is the 6-minutes walking test (6MWT). It is considered as a simple, non-invasive, and reproducible test that reflects the physical condition status of the tested subject through an objective measurement [10], [11]. In that sense, the use of accelerometers, such as Actigraph, Mini-Mitter or IM systems, to assess the physical activity is widely employed in the research literature [12]. Hence, the revolution of using smartphone applications is providing new opportunities for physical activity and functional fitness assessment.

Smartphones incorporate different sensors that allow them to perform different tasks [13], [14]. Among these sensors, they are provided with accelerometers, making available the physical activity quantification. Furthermore, smartphone applications, also well-known as apps, makes their use more suitable for a broader population, allowing them to monitor and manage several chronic conditions [15]. Therefore, these devices can get the sensor’s information and store and share it (by WiFi, 3G, LTE, etc.) with another remote device to submit it to a post-processing stage [16].

Classification of time-series data such as signals generated by inertial sensors are difficult to be classified in a sample-as-features fashion: the high dimensionality of the feature space along with the limited number of samples (as it is usual in biomedical problems) produces the so-called curse of the dimensionality problem, limiting the generalization capabilities of the model. In that way, a post-processing stage of computing relevant information to classify inertial signals by extracting statistical temporal and spectral descriptors reduces the dimensionality of the feature space improving the generalization capability of the model and reducing the computational burden. While temporal features are based on statistics computed directly over the time signal, spectral features require computing the Power Spectrum (PSD) of the signal to extract information (i.e., power) in different sub-bands. This has been traditionally carried out by Fourier [17] or Wavelet Analysis [17], [18], [19]. However, these analysis methods present some drawbacks related to the non-stationary nature of the inertial signals.

Furthermore, the use of features computed from the time signal (such as mean, variance, or amplitude-related features) as well as features derived from the PSD spectrum (peak power, spectral centroid, spectral kurtosis, etc.) could neither be descriptive enough or capture the pattern related to class discrimination. Thus, we hypothesize that the physical activity and functional levels of older adults could be predicted using two complementary methods that extract discriminative features from the inertial signals recorded during the 6MWT.

In this paper, we present two different approaches based on deep learning architectures. The first consist on using a Convolutional Autoencoder (CAE) to extract features from the acceleration data. These features are then classified using a support vector machine. The second approach presented in this paper consists of using a convolutional neural network (CNN) that firstly extract features from the acceleration data and then, uses a perceptron-like network (fully connected layers) to implement a classifier. In this case, the same network extracts features and classify them. Moreover, we compare the results obtained using the deep learning-based approaches presented in this paper, to the previously presented in [20], which uses only signal processing techniques (specifically, Empirical Mode Decomposition) to decompose the acceleration signals and then computes classical time and frequency statistical descriptors. These descriptors eventually classified using a support vector machine.

Thus, the novelty of this paper is twofold. Firstly, we used the 6MWT to explore human movement patterns related to fragility. Secondly, we proposed a method that avoids either to use of classical statistical signal processing techniques or the computation of predefined statistics as features to describe the signals. Instead, we propose two deep learning architectures that automatically compute specific features through a learning process. In this way, we present two methods that can be seen as complementary. The CAE approach uses unsupervised learning to compute representative features that can be eventually classified. The CNN consists of a convolutional stage that extracts features and a classification network. These two approaches perform equally (given the statistical analysis), but provides two ways using different learning paradigms to extract features from inertial signals using machine learning.

The rest of the paper is organized as follows. Section 2 describes the current studies available in which the 6MWT is measured using accelerometers. Next, Section 3 presents the database used to assess the method proposed in this work along with a description of the methodology, including the feature extraction process from the original inertial signals. Then, the results obtained are shown in Section 4 and finally, conclusions and future work is presented in Section 5.