Tom Doris, KeepUs project founder and technical lead, responds to our recent post [TTA 28 Aug] critiquing Philips Lifeline with AutoAlert’s accelerometer and its possible failure to detect a fall which resulted in the death of a Massachusetts woman. His analysis concludes that accelerometers on their own are surprisingly inaccurate. The false positives/negatives may be minimal but they do exist, and they should not be the only indicator of a fall.
Falling Down is a Surprisingly Hard Problem
More than 250,000 people suffer a hip fracture in the US every year. More than 20 percent will die within 12 months as a consequence of their fall. One in three who lived independently before the fracture will need at least a year of rehabilitation in a nursing home. While rehabilitation methods are improving, the single most important factor influencing the long-term outcome is the length of time between the fall and getting medical attention at a hospital. A few hours more or less makes the difference between life and death.
People are living longer, and current projections make it clear that elderly people will have to live independently in their own homes for as long as possible. You just can’t provide residential care for 20 percent of the population. Smartphones and wearable technology have the potential to dramatically improve eldercare. A relatively cheap smartphone can track activity and location. Modern platforms analyze the data in real-time over the internet and can, in theory, immediately spot when something is wrong and raise an alert.
The theory doesn’t always work however. As devices and apps hit the market, we’re gradually seeing what works and what doesn’t. It turns out that some problems can be solved and we just need more accuracy and ease-of-use. Location tracking is a good example of this. The early systems were unreliable and clunky to use, but they’ve improved rapidly and geofencing systems now work very well when properly configured. Fall detection is one of the most important problems. At first glance it seems like an easy one to solve, but it hasn’t yielded to the new technology.
Toasters Save Lives
About 25 years ago, a lot of houses in my neighborhood had disabled one or more of their smoke alarms. Most people didn’t have toasters, and they used the oven grill to make toast. Of course, this meant that at least once a week the toast got burned and the smoke alarm sounded. So they unplugged or took out the battery. In retrospect, it’s highly irrational: parents were risking the lives of their kids because of the minor inconvenience of an occasional false alarm. Now that everyone has a toaster, false alarms are extremely rare–nobody disables them.
Alarms that sound when they shouldn’t are “false positives”. A smoke alarm that fails to sound when the house really is on fire has suffered a “false negative”. There really is a fire and it’s failed to detect it. In any safety system, false negatives are possibly the worst kind of failure. When the danger is a very rare event, like a house fire, we often fail to behave rationally about safety measures.
Balancing False Falls
No matter how clever a fall detection system is, some falls will look like sudden sit-downs and some sudden-sit-downs will look like falls. The designer has to select a threshold that tries to balance the false-positive rate with the false-negative rate. This is the dilemma: you can reduce the number of times a fall goes undetected, but by doing so, you increase the number of false alarms. The big problem is that there are dozens of events every day that look like falls. If you increase your threshold even a small amount, in the hope that you’ll avoid missing a real fall, you’re going to have many more false alarms. It’s as if you’re designing a smoke alarm for someone who burns their toast half a dozen times a day. Worse still, if the false alarm rate is too high, people will just disable the system or ignore it, no matter how important the alert might be.
No Silver Bullet in Sight
The best systems we have for fall detection aren’t yet good enough for everyday use. They have too many false positives and are prone to false negatives. In recent years there has been lots of work on fall detection systems. Industry and academia have tried to create better devices and algorithms, and some very clever people have tackled this problem. Accelerometers and gyroscopes that detect position and movement are smaller and more accurate. The best detection systems all achieve about the same accuracy rates, give or take. Researchers have tried a wide variety of approaches. So we shouldn’t expect a massive improvement any time soon.
Fall detection is a difficult problem even if you have state-of-the-art accelerometers and gyroscopes positioned in multiple places on the person (Accurate, Fast Fall Detection Using Gyroscopes and Accelerometer-Derived Posture Information–University of Virginia). Smartphone accelerometers are getting better, but they are still quite basic devices, and most are “saturated” at 2-g, meaning that any force higher than 2-g is just reported as 2-g. They also don’t have great response curves for the range that they actually do cover.
Stumbles Are the Worst Falls
The evidence we found suggested that a large fraction of hip fractures result from events that would not even be classified as a bad fall by human witnesses. They are more like awkward stumbles, with more than one impact on the way down, so that the g-forces are actually low, and extremely difficult to discern from normal activity.
False Sense of Security
It’s important for the telecare industry to recognize that fall detection is an extremely difficult problem, and start using terminology that reflects the real accuracy of each system in terms that make sense to the consumer. A hidden danger of telecare devices is that they give people a false sense of security. Carers and family will check in less often if they think the alerting device is always going to work. When we read that a system is “98 percent accurate”, we understand that to mean that failures are extremely rare. But statistics are confusing here, and the percentage accuracy rates don’t give a clear picture at all.
It’s a bit like the freshman probability question: if 1,000 employees take a mandatory drug test which is 95 percent accurate, and 5 of them are actually using drugs, then if someone tests positive, there’s only about a 1-in-10 chance they are actually a drug user (because there’ll be about 50 false-positive tests). This is confusing stuff, but it’s important to get right. (The False Positive Paradox).
How to Recognize the Solution When It Arrives
The next-generation fall detection system will probably need to use more than just accelerometer data to achieve the necessary accuracy and reliability levels. If someone has fallen and needs help, vital signs like pulse rate and breathing will indicate distress. Bracelet devices can gather these signs and can be worn all the time, which is very important since most falls happen getting in or out of bed and in or out of the shower.
When a company brings such a device to market, they’ll be talking about the unique research they did to make it work, it won’t have been easy, they will have spent a lot of time and money making it work, so they’ll want to tell everyone about it. We should be skeptical of any claims that aren’t backed up by hard evidence of a new and innovative approach.