Fall Detection as a Digital Biomarker

Fall detection is the automated identification of a fall from wearable motion sensors, usually reported as a count of events over a monitoring period and, in consumer implementations, as an immediate alert to emergency contacts.

The clinical problem is measurement, not intervention. Falls are conventionally counted from patient recall or from diaries, and both under record: people forget falls that caused no injury, and cognitive impairment, which raises fall risk, also degrades recall. A method that observes falls rather than asking about them addresses that directly.

Falls also have an unusual position in this library. The measure is absent from the public registry of digital endpoints entirely, yet fall detection is a shipped feature on mainstream consumer watches and has a near perfect questionnaire counterpart in the Falls Efficacy Scale International. That combination, real world deployment with no registered endpoint use, is exactly what a registry based library cannot see.

Detection uses accelerometry, often with a gyroscope, and looks for the characteristic signature of a fall: a period of near free fall, a high magnitude impact, and a subsequent change in orientation with little movement. Thresholds and machine learning classifiers are both used.

Validation has a specific difficulty. Real falls are rare and cannot be staged ethically in the populations that fall most, so much of the algorithm development literature uses scripted falls performed by young volunteers onto mats, which do not resemble real falls in older adults. Work evaluating waist mounted algorithms against continuous unscripted daily activity showed how much performance drops once the recording includes ordinary life, and methodological reviews have set out how real world evaluation should be done.

In research the value is an objective fall count as an outcome in trials of exercise, balance training, medication review and deprescribing, where the conventional endpoint is a diary. It also supports observational work linking gait and balance measures to actual falls rather than to fall risk scores.

Outside research the dominant use is safety alerting, particularly for people living alone, where the time spent on the floor after a fall is itself a determinant of outcome.

The measure is reported alongside the Falls Efficacy Scale International, which captures fear of falling, and alongside fall risk and balance instruments such as STRATIFY, the Berg Balance Scale and the Tinetti assessment. That pairing matters because fear of falling changes behaviour independently of whether falls occur, and a study that reduces falls by making people move less has not succeeded.

Accuracy in the real world is well below what scripted validation suggests. Sensitivity is imperfect, particularly for slow slides to the floor that lack a clear impact, and false positives are generated by dropping the device, vigorous activity and sitting down heavily.

Consumer implementations are tuned for alerting rather than for counting, so their thresholds favour avoiding false alarms and will miss less severe falls that a research study would want recorded. Detection also depends on the device being worn at the moment of the fall, and falls cluster around night time toileting when watches are often off the wrist and charging.

Across the domain, an absent detection is not evidence that no fall occurred, and studies should continue to collect participant reported falls alongside.

• Yardley L, et al. Development and initial validation of the Falls Efficacy Scale-International (FES-I). Age Ageing. 2005. pubmed.ncbi.nlm.nih.gov
• Bourke AK, et al. Evaluation of waist-mounted tri-axial accelerometer based fall-detection algorithms during scripted and continuous unscripted activities. J Biomech. 2010. pubmed.ncbi.nlm.nih.gov
• Chaudhuri S, et al. Real-world accuracy and use of a wearable fall detection device by older adults. J Am Geriatr Soc. 2015. pubmed.ncbi.nlm.nih.gov
• Broadley RW, et al. Methods for the real-world evaluation of fall detection technology: a scoping review. Sensors. 2018. pubmed.ncbi.nlm.nih.gov