Wearables promise to catch your falls. Real-world proof in older adults is scarce

Fitness watches and wearable pendants make two distinct promises about falls: that they will automatically detect a fall that has already happened and summon help, and, increasingly, that they can predict who is likely to fall before it occurs. On the first, the independent evidence in older adults is mostly drawn from staged laboratory falls, with real-world validation rare; on the second, a 2026 meta-analysis found wearables predict future falls with high specificity but only moderate sensitivity — useful for flagging people for screening, not for catching most future fallers [s1][s3].

Two different tasks, often conflated

Detecting a fall and predicting one are separate engineering problems. Detection reads a sudden motion signature — the accelerometer and gyroscope spike of an impact — and decides whether to raise an alarm. Prediction reads gait and balance over time and estimates risk. A device marketed as "fall-aware" may do either, and the evidence for the two is not interchangeable [s1].

What the detection evidence actually shows

A 2026 scoping review in Frontiers in Public Health mapped the field, including 243 studies of wearable and sensor-based fall detection and pre-impact prediction [s1]. Wearable inertial sensors dominated, and post-fall detection was the most studied task — but more than half of the studies relied on laboratory-based simulated falls rather than real ones [s1]. When the reviewers narrowed to studies with real-world or mixed real-world validation in people aged 65 and over, only 21 of the 243 qualified; within that subgroup, 71.4% addressed post-fall detection, 19.0% pre-impact prediction, and 9.5% fall-risk modelling [s1]. Evidence on long-term adherence, integration into care workflows, and health-economic impact was limited [s1].

That staged-versus-real gap matters because a fall deliberately performed by a young volunteer onto a crash mat looks cleaner to an algorithm than the slumps, slides and near-catches of an older adult at home. A 2025 systematic review in Sensors of fall detection in ambient-assisted- living and smart-home settings screened 473 records and analysed 80 studies across wearable, non-wearable and hybrid sensors [s2]. Its statistical comparison found that wearable sensors performed the worst of the three categories on fall detection, while deep-learning methods outperformed simpler threshold approaches across accuracy, precision, sensitivity, specificity and F1-score [s2]. Wearables win on convenience and cost, in other words, not on raw detection performance [s2].

What the prediction evidence shows

The prediction side has firmer pooled numbers. A 2026 meta-analysis in Frontiers in Public Health combined 20 studies in a bivariate random-effects model, searching the literature to 9 October 2025 [s3]. Pooled sensitivity was 0.55 (95% CI 0.42-0.67) and specificity was 0.89 (95% CI 0.84-0.93); the positive likelihood ratio was 5.2, the negative likelihood ratio 0.50, and the diagnostic odds ratio 10.39 [s3]. The summary receiver-operating-characteristic curve returned an area under the curve of 0.85 (95% CI 0.81-0.88), rising to 0.90 in studies that used machine learning [s3].

Read plainly, those figures mean a wearable is fairly good at confirming that a flagged person is at elevated risk — high specificity — but misses close to half of the people who go on to fall, because sensitivity sits at 0.55 [s3]. The authors conclude the devices are better suited to early screening and risk stratification than to standalone prediction [s3]. They also caution that performance is not fixed: it shifted with the age structure of the population studied, the sample size, and where on the body the sensor was worn, which is part of why pooled sensitivity carries such a wide confidence interval [s3].

The scoping review frames the same unevenness as a maturity gradient across technologies: of the modalities it surveyed, wearable detection systems showed the strongest grounding in real-world conditions, while the predictive, multimodal and whole-ecosystem approaches were less mature and less tested outside controlled settings [s1]. The relatively well-developed corner of the field, in other words, is the simplest one — a watch or pendant noticing an impact — and even that rests on a thin base of real-world studies in older adults [s1].

What it means

The useful distinction for anyone weighing a fall-alert watch or pendant is between a feature that is plausible and one that is proven in people like the intended user. Automatic detection is widely demonstrated in the lab and occasionally lifesaving in anecdote, but the body of real-world validation in older adults is small, and wearables lag other sensing approaches on detection accuracy [s1][s2]. Prediction has better-quantified performance, yet at a sensitivity that should not be mistaken for a safety net [s3].

What to watch is whether the next wave of studies moves out of the motion-capture lab and into homes, and whether anyone demonstrates that an automatic alert changes an outcome — fewer long lies on the floor, faster help, lower costs — rather than simply firing when it should. That outcome evidence, the reviews agree, is still largely missing [s1].