Falls are a major source of disability, morbidity, and mortality. Recent in-home technology advancements—such as new sensor and accelerometer technology—are promising areas for falls reduction and alerts. If such in-home technologies could be used to detect falls in real time, and to report the fall to a remote caregiver or emergency personnel, much of the harm caused by in-home falls can be reduced.
A main challenge in real-world, real-time falls detection relates to computing. If a user is wearing an accelerometer that is sending data to a remote automated computer system, how can the system distinguish a fall from, for example, plopping down on the couch to relax, bending down to play with a pet, or even just lying down to sleep?
To solve this problem, researchers and engineers develop algorithms that process the movement of users—their movement up or down and right or left, and the speed of their movement. Based on this data, the algorithm calculates whether the movement was likely to be the result of a fall. These algorithms are tested and based on simulated falls in a laboratory setting, partly because it has been difficult to obtain real-world fall data. It is unclear whether these algorithms based on simulated laboratory falls—usually simulated by physically fit volunteers, often with expertise in martial arts or other physical activities—are useful for detecting real-world falls.
A recent study conducted by an international team of researchers evaluates the effectiveness of these algorithms to analyze falls within a unique database of real-world falls. This database included accelerometer measures that captured the movement of a group of participants, each for a period of 48 hours. Participants reported any falls that occurred to the researchers. In total, there were 29 real-world falls for which the researchers had accelerometer data.
The researchers then tested 13 different algorithms to see if they were able to identify the real-world falls. Unfortunately, none of the algorithms scored particularly high in both sensitivity (the ability to correctly identify falls that actually occurred) and specificity (the ability to correctly identify a movement as a non-fall). Two performed fairly well on both of these measures, but each would create too many “false alarms” if used as an automated detector of falls. Despite these findings, the authors stated that by accumulating a larger accelerometer database of real-world falls, researchers may soon be able to develop an accurate remote falls monitoring system. This would help solitary individuals who suffer falls in their homes, and improve our understanding of how to minimize falls risk.