1. Introduction
Falls represent a serious issue which can have catastrophic consequences and the need for ICT (Information and Communication Technologies) based solutions is hence emerging to improve the life quality and autonomy of frail people, with particular regards to falls’ detection and classification.
In the last decades, the interest of the scientific community in the field of assistive technology is acquiring more and more visibility, especially due to the phenomenon of population ageing.
Just to mention few examples, in United Kingdom in 1997 around one in every six people (15.9%) were aged 65 years and over, increasing to one in every five people (18.2%) in 2017 and is projected to reach around one in every four people (24%) by 2037 [
1]. In 2018 the Italian resident population amounted to about 60 million units. The average age is 45.2 years, reflecting a structure by age where only 13.4% of the population is under 15, 64.1% between 15 and 64 and 22.6% is 65 and older [
2].
Most countries will see the number of elderly (60+) doubled in the coming 30 years, but also the number of the oldest-old (80+) will grow drastically in the long run. Specifically, considering the European country, the share of the elderly in the total population is expected to increase from 21% now to around 34% in 2050. In absolute terms, 37 million people are expected to be aged 80 and over in 2050, an increase by almost 160% compared with 1995 [
3].
From the aforementioned statistics, it is straightforward to deduce the need for solutions assuring an improved quality of life and independence of the elderly.
To address this issue, in the past few years a significant research activity, focusing on advanced and reliable solutions to monitor frail people, has been developed [
4,
5,
6,
7,
8,
9,
10,
11,
12,
13,
14]. These systems would directly benefit elderly people, by providing them with more autonomy, reducing the need for moving to institutionalized care center, enabling timely and effective intervention in case of need, and ultimately reducing the emotional and financial burden for the elderly and their families. The use of technology would also reduce the overall costs of health and social care [
15].
Valid monitoring systems can be extremely beneficial in analyzing deviations with respect to the normal behaviour of an individual which can be strategic for early detection and treatment of worsening health conditions [
16].
Celler et al. [
17], during a studies of behavioural monitoring, discovered that the health condition of a subject can be estimated by check a set of simple parameters such as the mobility, sleep patterns, washing and toilet facilities, that have been demonstrated to be representative of the interaction of the subject with his environment. Commonly analyzed human activities are actions typically performed during the day, such as dressing, standing, sitting, walking, climbing stairs and laying down. Many different technologies have been proposed in the literature for human activities and falls monitoring. These include wearable multi-sensor architectures, developed by customized solutions, smartphones, or a combination of these systems.
The main aim of this paper is to present an overview of different approaches for falls detection, with a specific focus on methodologies developed at the SensorLab of the University of Catania, Italy, for fall classification. Robustness of developed methodologies against other user behaviors, such as sitting, is also demonstrated. A detailed discussion and more experimental results related to the classification paradigms developed can be found in journal papers already published by the authors [
18,
19].
4. A Case of Study
The classes of events, E, considered through this paper are:
Backward falls (FB) (50 repetitions);
Forward falls (FF) (50 repetitions);
Lateral falls (LF) (50 repetitions);
Sitting events (SI) (50 repetitions).
Each event has been acquired for 10 s, with a sampling frequency, , of 500 Hz. Although a sampling rate of 100 Hz has been demonstrated to be suitable in case of Fall detector, since the aim of this work is the classification of different kind of Falls, a higher sampling rate has been used to avoid any loss of information provided by the accelerometer signals. The acquisition were performed using ultralow-power/high-performance/three-axis nanoac- celerometer, ST LIS3DH. The accelerometer has dynamically capable of measuring accelerations with output data rates user selectable full scales of ±2g/±4g/±8g/±16g and it is ranging from 1 Hz to 5 kHz. It provides a 16-bit information by an Inter Integrated Circuit/Serial Peripheral Interface digital output interface. The sensor has been positioned on the user’s right hip since this is close to the center of the body mass. The inertial monitoring of such body point will provide reliable information on the body movements, which are minimally affected by sudden limb motion artifacts.
In particular, 10 users, with different stature and weights and ranging between 25 and 44 years old, with a mean of 37 years and a standard deviation of 5.56 years, have been selected to simulate both falls and sitting events, 5 times each. Characteristics of users involved in the experimental trials are summarized in
Table 1.
It is mandatory to underline that the aim of this case study is not to distinguish falls from the sitting event but rather to classify different kinds of falls (forward fall and backward fall as an example). However, the sitting event has been included in order to the test the robustness of the methodology proposed against a non-fall event which shows a fall-similar dynamic.
It is mandatory to clarify that this work reports on laboratory tests performed by people belonging to the Research Team with different heights, ages and weights in safe conditions (falls have been simulated using a mattress as a common practice adopted also by other research groups). Each participant was requested to sign an informed consensus regarding the purpose of the study and working conditions. Nevertheless, every precaution was taken to ensure user safety during experiments. Moreover, it must be considered that the Device Under Test is not belonging to the class of medical devices, being an external inertial unit used to monitor the dynamic of the user body.
Although users addressed by the solution should be frail people, the choice of using healthy subjects performing tests in safe conditions, has been taken to avoid injuries during this preliminary phase. To support this choice, it must be stressed out that the event-driven cross-correlation classification strategy developed is robust against light modifications of signal dynamics, thus confirming that the classification procedure can be successfully extended to a new data set generated by real users.
In particular, 40 of the 50 acquisitions have been used to generate the event-related signature while, the remaining 10, have been used for test purposes. The reason behind the larger number of data used for signatures generation is due to the need of generating typical time evolution of the acceleration module for each of the considered classes. To such aim, the availability of a dataset able to properly represent the typical dynamics of each class of events is mandatory. Signatures generated using data collected during the experiments are shown in
Figure 2.
As an example, features obtained are reported in
Table 2. For the sake of clarification, each row of the
Table 2a is an FF event taken from the test set. In particular, each element of the row is the correlation result between the FF event and the FF, BF, LF and SI signature. For example, the element in column 2 is the maximum correlation between the FF event and the BF signature.
As it clearly emerges, higher values of the features have been obtained in case of the correlation of the pre-processed unknown pattern with the its related signature: the first column of the
Table 2a shows a greater value of the features because it contains the correlation between FF events and the FF signature, while, the last column of the
Table 2d, shows a greater value of the features because it reports the correlation between SI events and the SI signature. The same reasoning apply to the other events.
It should be noticed also that high features values can be obtained also in case of cross-correlation between a pattern belonging to a class and the signature of a different class (this is particularly evident between the FF event and the BF one). This scenario, which could bring to mis-classifications, can be justified by similar dynamics of the two class of events.
Table 3 shows the optimal threshold, for each class of events here addressed, evaluated by using the ROC curves theory [
5,
18].
Each threshold value shown in
Table 3, is related to the corresponding column of
Table 2, which is the one associated to its signatures. As an example, the threshold value for the SI event (0.88) only applies to the SI column in
Table 2a. In the same way the threshold value for the BF event (0.95) only applies to the BF column in
Table 2a.
The result of the comparison between feature values and the adopted thresholds will produce 1 in cases the feature value is higher than the corresponding threshold, and 0 in the opposite case. According to the above mentioned procedure, results shown in
Table 4 have been obtained; observing this Table, occurrences introduced in
Section 3.1.3 can be identified. Considering
Table 4a containing results for the FF events, any classification has been performed in the third and sixth rows; considering the table containing the SI events, different multiple classifications can be observed.
These occurrences can be reduced taking into account the post-fall classification, implemented by the ATA paradigm. Results shown in
Table 5 demonstrate reduction of mis-classifications.
5. The Assessment Procedure
In the following notes, the measurement procedure developed to assess performances of the different classification strategies addressed in above sections, is presented.
In case of a generic Event Class E, the following quantities can be defined:
TP (true positive): events of type E correctly recognized as belonging to class E;
FN (false negative): events of type E recognized as belonging to a class different than E;
TN (true negative): events different from type E correctly recognized as belonging to a class different than E;
FP (false positive): events different from type E recognized as belonging to class E;
During the assessment procedure, the following performance indexes will be used:
Sensitivity (
): the capability of an algorithm to correctly identify TPs as such.
Specificity (
): the capability of the system to correctly identify TNs as such.
Basically, the aim of the assessment approach is to estimate the system performances in terms of reliability in fall classification.
In order to provide a fast and synthetic way to highlight the performances enhancement using the ATA as respect to simple TA algorithm, in
Table 6 a comparison between the two classifiers is given. In particular, the Table shows the indexes computed for each event considered through this work and, moreover, an average evaluation of the algorithm’s performances across all the classes of events (last column of the Table).
In conclusion, it can be said that there is a significant improvement in the classification performances when moving from the TA algorithm to the ATA paradigm. Values reported in
Table 6, then, validate the adopted strategy while confirming the reliability of the proposed approach.
6. Conclusions
The worldwide ageing population is pushing forward the development of reliable assistive solutions for the Active Assisted Living context. Particular emphasis is given to falls which represent a serious issue which could bring catastrophic consequences. It hence emerges the need for the development of reliable and robust solutions to address the requirements of frail people willing to live autonomously. To such aim and taking into account also the need for low cost solutions, the focus of research efforts should move from very expensive hardware solutions to effective signal processing and smart computational paradigms.
In this paper, after a brief review of the State of the Art on falls recognition and classification, a case of study addressing a classification methodology exploiting a event-driven correlation paradigm and a threshold based classifier is presented. An improvement to this solution is represented by the integration of the post-fall evaluation of the accelerometer axes.
Computing the threshold values by means of the ROC curves theory further allow to make the classification methodology robust to exogenous factors. Performances of the methodology proposed have been addressed by two performances indexes, and , both in the case of TA and ATA algorithms. The value moves from 0.81 (TA) to 0.97 (ATA). The value moves from 0.92 (TA) to 0.99 (ATA). Above results state for the high reliability of the methodology developed and encourage future efforts to further extend its applicability to a wide set of events.