In addition to diagnosing and managing specific conditions, monitoring gait in the elderly population can serve as a proactive approach to maintaining overall health and independence. By establishing baseline gait parameters for individuals, healthcare providers can track changes over time, enabling early interventions and personalized treatment plans.
2. Gait Monitoring and Analysis
In the last few years, there have been several studies aimed at analyzing the gait of people, in particular of the fragile ones.
In particular, a study proposed by the Politecnico di Torino [
8] employs shoes with sensorized insoles and a chest-mounted device, both used to record data in a controlled laboratory setting. Other studies adopt video-based solutions [
9] or different kinds of external systems, such as sensorized walkways [
10]. Unfortunately, forcing a person to wear different devices or interact with unfamiliar technology to obtain accurate data from their gait could compromise the data themselves since the person can be uncomfortable and change the way he or she moves. Moreover, when a device makes the user uncomfortable, it can only be used for short continuous sessions, usually a few hours at best, thus entailing two main problems: to gather statistically meaningful amounts of data, several sessions need to take place, and short sessions are unlikely to feature relevant events such as stumbles and falls. The intensity of the symptoms caused by the disorders affecting the patients may vary during a given day, week, or year. Short and infrequent sessions could miss important fluctuations or key events and not fully represent the patient’s real condition.
Smaller and lighter devices can also create issues if the patient is not used to them and can only be used in a specific laboratory environment under qualified personnel supervision. In [
11], the patients wore an inertial sensor attached to a waist-mounted strap situated on the lower back while performing a supervised timed up and go test to extract useful features from the signal for fall risk assessment. In [
12], the authors placed a small accelerometer on the patients’ lower back before asking them to perform a series of tasks in a supervised laboratory environment to identify early markers of Parkinson’s disease (PD). Laboratory environments usually have smooth and flat surfaces without external influences, making these conditions significantly different from the ones people find in their daily lives, characterized by irregular surfaces and obstacles. The presence of other people can also make the patients feel anxious.
The data collected in these circumstances, for the reasons presented above, may not represent the real condition of the patients, who can be prone to alter their actions subconsciously and, thus, their gait. Many studies attempted to use more widespread and accessible devices, such as torso-mounted smartphones [
13,
14,
15], which can accurately capture the forces involved in a walk with minimal noise from other sources. However, such devices still encounter low acceptance on behalf of the patients due to their weight, size, location, or stigmatization [
16,
17]. Smartwatches, like other wrist-worn devices and accessories, are significantly easier to wear, less bulky, and more likely to be accepted by the user. Unsurprisingly, the scientific literature has been moving in this direction for a few years. The goals in this scope are many and diverse.
In [
18], the authors exploited the acceleration data available in the UK Biobank database to determine whether acceleration data collected during daily living can serve as a prodromal marker for PD. Although the number of subjects is extremely high, the authors only focused on generic information extracted from the data, such as average acceleration, without analyzing the patients’ gait. Another study using data from the same database is [
19], in which the goal was to use free-living acceleration data to detect PD early. The authors split the raw data into segments of either gait or low acceleration magnitude and found notable differences between PD patients and healthy controls. Both studies only focused on PD without proposing a way to track a patient’s health status applicable to a more general population. Other studies, such as [
20], also analyzed the gait for an entirely different reason. The authors wanted to extrapolate gait features from the acceleration data, which could authenticate the user wearing the device. While this study represents additional proof of how gait features can be observed at the wrist level, the data were collected by having subjects walk a corridor six times at three different paces, thus creating a controlled environment. Moreover, the study did not focus on the health status of the subjects or the quality of their gait.
A key aspect of gait analysis is differentiating what the user is doing, a task known as activity recognition. Studies such as [
21,
22] are useful in this scope. In the first one, the authors aimed at detecting gait segments in acceleration data collected in real-life scenarios. In contrast, the second one evaluates activity recognition and fall detection methods on different datasets. While understanding what the user is doing and isolating gait sections is the first step of assessing gait quality, the studies did not explore which features could be extracted from these sections and actions to obtain an approximate measure of the quality of the movements or the user’s health status.
In [
23], the authors investigated the correlation between wrist movements and freezing events in patients with PD. This event detection is more closely related to health status than the two previous studies. However, the proposal is still too specific to be applied to a larger population. To create a reliable process to extrapolate features and detect gait events from acceleration data, the authors of [
24] proposed a deep-learning-based automatic pipeline validated against a sensorized walkway. Although studies such as this one offer an interesting foundation, they do not explore how the features can be used to assess quality and health status.
Helpful information can be obtained by performing a spectral analysis of the data coming from body sensors. This is performed by applying the Fourier transform to the acceleration data. This procedure allows to compute some indicators, such as the mean value of the frequency, its variance, and its entropy. Having the data of single individuals over a pretty long period of time, the variation of those indicators can give information on the progress of the disease.
In [
25], the authors performed a Fourier analysis of the stride-by-stride estimates of the linear acceleration coming from body-worn inertial sensors. The forces measured by the triaxial accelerometer were transposed from a local to a global frame using a quaternion-based orientation estimation algorithm and by detecting when each stride began using a gait segmentation algorithm. The authors of [
26] developed models for reducing the complexities of extracting features from data coming, among others, from wearable sensors. The extraction of the features is based on spectral analysis, and, in turn, the frequency analysis is based on the fast Fourier transform. Additionally, in [
27], a method for analyzing data from a human body sensor was performed. It developed an algorithm that could identify the gait behavior with high precision. As in the previous ones, also in [
27], the Fourier transform represents a fundamental tool in the classification process. Other applications of the Fourier transform for the analysis of frequencies can be found in [
28], where the authors studied the data coming from people having different kinds of neuromuscular disorders. The authors of [
28] developed software that could detect quantitative observations that could improve the subjects’ gait. For other details about the applications of the Fourier transform to human gait activity, see also the references contained in [
25,
26,
27,
28].