The health-tech industry is investing a lot of resources in automatic fall detection, to give the right serenity to the elderly and caregivers. But the issue is not easy to solve
Falling is human
“Do you know why we fall, Bruce? To learn how to get back on our feet,” Thomas Wayne told his son, the future Batman in Batman Begins. Unfortunately, it’s not always that simple, especially when you’re old and need help, and the most important thing is that you get there as soon as possible. That’s where to automatically detect falls comes to help.
The percentage of seniors over 65 who fall at least once a year is between 28% – 35%; it reaches 32% – 42% for those over 70.
Since our society is “aging” fast – in the sense that average life expectancy is rising worldwide and it is estimated that by 2030 the number of people over 60 will reach one and a half billion – it is necessary to take precautionary measures as soon as possible, for limiting the physical damage falling incidents usually cause.
One of the related phenomena to be remedied is the so-called “the long lying”. It refers to the time the person spends on the ground waiting for rescue; it can lead to dehydration, loss of consciousness and even hypothermia in some cases. This is why early intervention can really make a difference.
Devices and seniors, or how perceptions of fall risk are changing
The urgency to find a solution to the detection of falls does not end only in optimizing the time of a posthumous intervention: according to some studies [1] on the subject, the fear of falling is an integral part of the problem, since this feeling would negatively affect the quality of life of the elderly person, forcing him to give up various physical activities and thus creating the psychological conditions that lead to isolation and depression.
The interviewees, who wear fall detectors, were elderly people who fell in the last six months. After the trial period – which lasted 17 weeks – most of those who had worn it occasionally said they felt safer and more independent.
So there seem to be no contraindications in the adoption of systems to automatically detect falls. Their potential is fully understood. Why then have not yet come to satisfactory solutions (bearing in mind that 100% accuracy is still impossible)?
Before seeing the critical issues together, it is useful to get an idea of what are the solutions tested so far and how they work.
How do you automatically detect falls?
Over the years there have been several (but not many) proposed categorizations of the various systems; here we refer in particular to one of the most recent and complete [2].
There are two main categories of these different systems: the first is based on detection by sensors; the second on the use of analysis algorithms. The instruments belonging to the two categories intervene at different moments of the fall detection. The analytical algorithms intervene on the data collected by the sensors, developing their own predictive rules on the available information.
Sensors: hunting for data
In order for the analysis algorithms to work, there must be data to process. Sensors are the source of this information but they are not all the same. Different classes of sensors provide different types of data, each with its own strengths and weaknesses. Some of them will surely amaze you.
Inertial sensors
In this category fall accelerometers, gyroscopes, barometers. They are instruments capable of detecting sudden movements while in direct contact with the person. That made them very popular and widely used in consumer electronics (smartphone, wristwatch).
Today they show excellent accuracy, and generally they aren’t an excessive threat to privacy or source of sensitive data. What hinders their use is their intrinsic intrusiveness; plus their placement on the user’s body influences their effectiveness and the results.
At the moment, the hips seem to be the best and most reliable place to detect falls. We must also keep in mind an essential factor: the fall is a rare and in its own way unpredictable event.
This makes it very difficult to use real and validated data to establish the right detection thresholds and, in case you use machine learning techniques, to train the recognition algorithms.
This represents a difficulty very recurrent unfortunately and which we try to overcome with functional methods only in the short term.
Contextual sensors
Within this category we find those sensors that trace the movements of the body perceiving it from the surrounding environment, based for example on acoustic or visual signals.
These include microphones, cameras, thermal and infrared sensors, which recently became very popular.
These systems are generally less invasive but have to deal with the problem of coverage of the monitored areas. Usually the installation can only take place in indoor mode and the presence of other people/objects in the place creates a large amount of false positives.
In particular, the solutions that are based on visual tracking must also respond to issues related to the treatment of privacy and the technical difficulties of calibration and installation of the cameras.
Radiofrequency sensors
This is an innovative and promising sub-category of “context” sensors.
The speed of body movements brings anomalous changes in radio frequencies that can be picked up by these sensors. In fact, these are waves that, like WiFi, are already everywhere, which makes the use of these techniques not intrusive. But it takes over – albeit to a lesser extent – a problem of coverage.
Integrated sensors
I.e. born from the “fusion” of several sensors. Used to overcome the poor accuracy of individual sensors or the excessive amount of false positives, the performance is still unsatisfactory at the moment.
The algorithms or process and classify data
Once the sensors have recorded and collected data, different “filters” pre-process and clean them. That is done in order to highlight the distinctive features of a particular event such as a fall.
Generally the recognition algorithms are based on some assumptions about how, for example, a fall should “appear” in mathematical terms in the collected data.
Simplifying a lot, if my only sensor consists of an accelerometer, I can expect a fall to appear as a sharp peak. A peak can mean many things: one with the right timing and intensity is probably a fall. When developing a fall recognition algorithm, our task is to describe this event as robustly as possible.
A particular case are the algorithms that use automatic learning. They are able to identify independently the parameters that can characterize an event, provided that you have many examples to train them.
Depending on the classification algorithm in question, we identify three different classes:
Threshold-based algorithms
like the simple example above. The threshold must be well fixed to avoid false alarms or, even worse, missing them!
They are divided into fixed (which apply to individuals and situations); and adaptable (they automatically adapt based on the user’s motility history).
Automatic learning algorithms
(also called machine learning), which are able to identify independently the parameters that can characterize a fall. Usually, these approaches take into account various characteristics more complex than a threshold; so that sometimes it is difficult even for those who have designed them to actually understand their logic.
Integrated systems
I.e. the result of the two techniques. For example, by using machine learning techniques to identify the best threshold (fixed or adaptive) to use. In this way, it is possible to obtain clear results but at the same time improve the performance of techniques based only on the threshold.
Automatic falls detection: difficult, not impossible
Now that we have a clearer picture of the “tools of the trade” and the fundamental notions to understand how the problem of automatically detect falls is addressed, we can better understand the difficulties that researchers face.
No sensor or algorithm is perfect. Often it’s necessary to resort to compromises or solutions that integrate different techniques to obtain good results.
However, this is a critical application, and trying aggressively not to miss any fall, inevitably leads to a high rate of false alarms (or false positives). That dramatically worsens the user experience and undermines confidence.
Obstacles and goals
So, why is it so hard to automatically detect falls?
– Actual data on falls are few and difficult to acquire;
– Often to overcome this inconvenience, as we have already pointed out, the data on which the algorithms are trained are built ad hoc by the researchers themselves, affecting the target audience (the elderly) and collecting data for training too different from those that will then be detected in real application contexts;
– We have just concluded that the integrated systems work better, but still contain redundant information due to very complex, “heavy” computational processes that increase the cost of production and the difficulty of implementation;
– There is no general evaluation system common to the different approaches;
– Finally, we should not underestimate the issue of acceptance by end-users or the elderly. The sense of a violation of privacy may in some cases prevail over the feeling of security and protection; the wearable device has a strong physical value that we cannot overlook; implementations on smartphones instead take little account of the data that still tell us about low adoption rates and the high risk of false positives caused by the fall of the device without implying the fall of the person himself.
In short, the automatic detection of falls is still an open challenge and of the utmost urgency, always confirming itself as one of the main “hot topics” of elderly care.
[1] Automatic fall detectors and the fear of falling, Brownsell and Hawley (2004)
