Falls among older people are a serious health issue and can result in hip fractures, traumatic brain injuries, and even fatalities.1 , 2 In Australia, the cost of fall-related injuries in older people is more than $200 million per year and is increasing as the population ages.3 Traditional methods of self-reporting or caregiver reporting of daily activities and fall-related events may be inaccurate due to recall bias, denial, and/or inability for constant monitoring. Remote monitoring using wearable devices is a low-cost alternative and can provide new insights into the complex interactions between active lifestyles, healthy aging, and increased exposure to situations in which falls occur.4
Fall detection using wearable devices has been the focus of substantial recent research and systematic review.5–10 Body-worn accelerometers detect impacts and changes in orientation associated with falls.5–8 Accuracy may be improved by using multiple sensors. For example, barometers can detect height changes associated with falls,9 and Android-based smartphones apps have also used gyroscopes and global position systems to detect falls.10 These technologies aim to provide rapid detection of falls and, therefore, prevent frail older adults suffering “long lies” due to not being able to get up after a fall. However, detecting a fall does not prevent injuries that can result from a fall, including hip fractures and traumatic brain injury.
Preventing falls may be facilitated by identifying people at risk of falling and early intervention. Fall risk assessments have also been the focus of substantial research and systematic review.11–16 Existing fall risk tools have generally included clinical assessments of multiple domains, for example, balance, mobility, physiology (strength, vision), psychology (fear of falling), cognition, local environmental risk, and medication use. However, widely accepted tools such as the Timed Up and Go test have low specificity (61%-68%),11 , 12 with different thresholds (11.4-12.5 seconds) recommended for identifying older people at increased fall risk. Clinicians must, therefore, determine which instrument or combination of instruments may best target the risk of falling for a given older adult.13 , 14
Near-fall detection could provide new opportunities to identify older people at high risk of falling before a fall occurs. Near falls are defined as trips, slips, and missteps and involve a loss of balance that does not result in a fall because a corrective action is taken to recover balance.17–19 Near falls occur more frequently than actual falls,17 and older people who frequently experience near falls are at increased risk of future falls.19 , 20 Remote monitoring of near falls during daily activities could inform and provide important information for specifically targeting these falls and associated circumstances as part of fall prevention initiatives. An accurate algorithm for the detection of near falls could also enhance the quality of existing fall detection systems by reducing false alarms that may lead to staff nonresponse to alarms or “alarm fatigue” in clinical settings.
The purpose of this study was to systemically summarize and review the evidence regarding the detection of near falls using wearable devices in older people. Through this systematic review, we sought to answer the following 3 questions: (1) Can near falls in older people be detected using wearable devices? (2) Can near falls be distinguished from falls and other activities of daily living (ADLs)? (3) Does the incidence and type of near falls detected by wearable devices differ with age or other chronic conditions?
Registry of Systematic Review Protocol
This review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines.21 The review protocol was registered at the International Prospective Register of Systematic Reviews (PROSPERO) (registration number: CRD42016047693).
Literature Search Strategy
Five electronic databases, CINAHL, EMBASE, MEDLINE, Compendex, and Inspec, were inspected from their inceptions until September 2016. No language or other restrictions were applied to the initial search. The following terms were used to search the databases: (wearable* or accelerometer* or gyroscope*) and (slip* or trip* or near fall* or near-fall* or loss of balance* or balance loss* or stumble* or misstep* or miss step* or pre-impact* or pre impact* or “not stab*” or decrease stab*) and (old* or older adult* or elder* or aged) and (detect* or Monitor* or predict* or Distinguish* or accurac* or signal*) not (turbine* or mechanic* or aerodynamic* or drill* or ocean* or avionic*).
Articles published in English that measured near falls using wearable devices among younger and older adults (60+ years of age) or people with a chronic condition (eg, Parkinson's disease [PD] and stroke) were included. A near fall was defined as a slip (sliding of the support leg), trip (impact of the swinging leg with an external object), or loss of balance where the person started to fall but not including falls to the ground or other lower surface.20 , 22 Studies that artificially induced near falls in laboratories and those that monitored the natural occurrence of near falls in daily life were included. Articles were excluded if they were (1) case studies, (2) an abstract of a conference, (3) published in a language other than English, or (4) did not use a wearable device.
Data Extraction, Section, and Coding
Titles and abstracts of the studies were screened independently according to the inclusion criteria by reviewers 1 and 2. The full texts of these studies were retrieved and independently assessed for eligibility by the 2 reviewers. Any disagreement about the study inclusion or the data extracted was resolved by discussion with reviewer 3 as required. A standardized, piloted form was used to extract data from the included studies. Extracted data included incidence and type of near falls, specification of a wearable device (sample rate, placement position, and algorithm for detection), study setting, participant characteristics, and the accuracy of near-fall detection. A narrative synthesis of the findings was then completed.
Risk of Bias and Relative Quality Assessment
The heterogeneous methodologies of the included studies were not suited to a standard risk of bias assessment. Therefore, a relative quality assessment tool was designed to rank the studies included in this systematic review. Seven key criteria relating specifically to the quality of a study to contribute to the 3 review questions were chosen by discussion and consensus with all coauthors. A point was scored for each criterion fulfilled and a total score was summed for each study (maximum 7 points, with a higher score indicating a higher quality study). The criteria were (1) assessment of multiple types of near falls (eg, trips and slips), (2) an attempt to distinguish near falls from ADLs (eg, bending down [excluding standing and walking]), (3) study setting (inclusion of naturally occurring near falls), (4) an attempt to distinguish near falls from falls, (5) adequate sample size (n > 10), (6) calculation of near-fall detection accuracy (ie, sensitivity and specificity), and (7) collection of near-fall data from older participants (≥60 years of age). We considered points of 0 to 2, 3 to 5, and 6 to 7 as low-, moderate-, and high-quality research, respectively. The relative quality assessment was performed by reviewers 1 and 2. Discrepancies were identified and resolved by discussion with reviewer 3 as required.
No studies met the original eligibility criteria and, therefore, the registered scope of this systematic review was updated to include studies that included people younger than 60 years and full-length conference papers. Figure 1 shows the flowchart of the study selection process. The initial database search identified 165 articles. After removing 87 duplicates, 72 articles were screened by their titles and abstracts and 52 articles did not fulfill the inclusion criteria. The remaining 20 were obtained as full-text articles and 9 were identified as eligible and included in this systematic review.23–31
Table 1 summarizes the methodologies and outcomes of the 9 studies included. Three studies involved young adults,23 , 24 , 28 no studies involved only older adults, and 3 studies involved both young and older adults in their experiments.27 , 29 , 31 One study recruited people with PD29 and 2 studies recruited healthy older people.27 , 31 Three studies did not report participant ages.25 , 26 , 30 In the included studies, only 39% (76 of 193) of participants were 60 years of age or older and no studies reported accuracy separately for older versus younger people. Two studies included 40 to 91 participants,27 , 29 4 included 10 to 15 participants,23 , 24 , 28 , 31 2 included less than 10 participants,25 , 30 and 1 did not specify the number of participants.26
Accuracy of Near-Fall Detection Using Wearable Devices
Five studies reported accuracy and/or sensitivity and specificity of 97% or greater,23–27 and 3 studies reported a range of values between 85% and 97%.28 , 29 , 31 One study did not report accuracy.30 Two studies found that accuracy was improved by increasing the number of wearable devices.24 , 26 Two studies compared different single device locations and found the chest25 and the right thigh24 to be the most accurate.
Study Quality and Risk of Bias
The results of the relative quality assessment are presented in Table 2. The quality of studies included in this systematic review was low to moderate, which may have also increased the risk of bias. The scores for the included studies ranged from 0 to 5, with a median of 3 (interquartile range = 3, upper quartile = 4, lower quartile = 1) and an average of 2.6. Methodological issues that reduced study quality included the following:
- No studies assessed the accuracy to detect naturally occurring near falls.
- Only 3 studies examined more than 1 type of near falls.23 , 24 , 28
- Only 3 studies attempted to distinguish near falls from ADLs.24 , 26 , 28
- Only 3 studies attempted to distinguish near falls from actual falls.23 , 26 , 28
- Only 3 studies included older adults27 , 29 , 31 and none of these studies reported the accuracy to detect near falls in older adults only.
Types of Near Falls
Three studies distinguished between multiple types of near falls.23 , 24 , 28 Two studies measured slips, trips, incorrect weight transfer while rising from sitting, and being bumped by another person.24 , 28 The remaining 6 studies measured single types of near falls; trips, stumbles, slips, missteps, or excessive leaning individually.25–27 , 29–31 Contrasting experimental setups were used to study near falls: Two studies asked participants to watch a video clip and perform the near-fall scenarios,24 , 28 1 study instructed participants to simulate different types of falls onto a crash pad,23 2 studies used accelerating or decelerating treadmills speeds,27 , 31 1 study placed an obstacle on a walking path,25 1 study applied soap film to a wooden walkway,30 and 1 study used rubber plates, wooden sticks, and a trip wire.29 No studies assessed the accuracy to detect naturally occurring near falls outside the laboratory environment.
Distinguishing Near Falls From ADLs
Three studies attempted to distinguish near falls from ADLs: Two studies asked the participants to walk, stand, rise from sitting, descend from standing to sitting and transfer from standing to lying, pick up an object from ground, and negotiate stairs.24 , 28 One study included walking, standing, and lying.26 All 9 studies used a wearable device to detect near falls in the laboratory environment. Two studies asked participants to wear a device for between 3 and 10 days to measure real-life ADLs.23 , 29
Distinguishing Near Falls From Actual Falls
Three studies attempted to distinguish near falls from actual falls: Two studies asked the participants to fall from an upright position onto a crash pad or couch.23 , 26 One study showed participants videos of people falling to train their movements in combination with using both trip and slip hazards.28
Near Falls in High-Risk Populations
One study included people with PD with an average 5.3 years since diagnosis.29 In this study, 40 participants were asked to walk over a laboratory walkway under 6 different hazard conditions. Twenty-nine missteps were provoked and used to train an empirical decision tree. The algorithm was then used to remotely monitor misstep behavior over 3 days in the community setting. People with PD and a history of falls (2+ falls in the past 6 months) had 23% more missteps than the people with PD and no history of falls.
Types of Wearable Sensors
All studies used a body-worn accelerometer to measure accelerations during a near fall. To increase device accuracy, 5 studies added gyroscopes that measured angular velocity,24 , 26 , 28–30 1 study included insole sensors,30 and 1 study further integrated the accelerometer data to obtain a measure of preimpact velocity.28
Wearable Device Placement
Positions where wearable devices were attached included the head,24 chest,24–26 waist,23–25 , 27–29 , 31 left and right thighs,24 , 30 left and right ankles,24 , 25 right and left trousers pockets,25 and shanks.30 Four studies used multiple wearable device placements.24–26 , 30
Near-Falls Detection Algorithms
Three studies used machine learning algorithms.23 , 24 , 26 Four studies used 4 different threshold detection algorithms that were based on (1) statistical distribution models,24 (2) stationary wavelet transforms,27 (3) preimpact velocity,28 and (4) the receiver operating curve.31 One study used an Extended Kalman filter30 and 1 study used a novel algorithm.29 The threshold values for the wavelet-based algorithm were adjusted over time for each individual to help minimize false positives.27 The different types of machine-learning algorithms included Support Vector Machines (SVM), Naïve Bayesian Classifiers (NBC), Radial Basis Functions (RBF), SVM, Decision Trees, k-Nearest Neighbors, and Ripple Down Rule Learners.23 , 24 , 26 Different algorithms performed better in different experiments. Choi et al26 compared 5 different machine-learning algorithms: NBC, SVM, RBF, C4.5, and a Ripple Down Rule Learner and found that the NBC performed best. In contrast, Albert et al23 found that the SVM outperformed the NBC. Finally, Iluz et al29 reported that all machine-learning algorithms tested had poor sensitivity (62.5%) to detect missteps in people with PD and, therefore, developed a novel multistage algorithm.
Detection of near falls is an emerging area of research that has grown from the rapid development of wearable device technology as well as accumulated evidence on fall detection.5–10 The clinical utility of this technology involves the unobtrusive and continual monitoring of activities of daily life in populations at high risk of falling4 to identify issues that might be targeted in fall prevention interventions to prevent falls and associated injuries.
This study systematically reviewed the evidence regarding detection of near falls using wearable devices. Importantly, no studies investigated the accuracy to detect near falls by older people at home or in the community. The registered scope of this systematic review was, therefore, updated to include studies that included people younger than 60 years. Furthermore, 7 of the 9 included studies were pilot in nature (15 or less participants), were largely conducted in younger people, and involved laboratory-induced near falls. The results demonstrated that wearable devices could be used to detect these artificially induced near falls with high accuracy, sensitivity, and specificity (85.7%-100%) in controlled settings. However, daily life and laboratory movements are likely to be different.32 The literature suggests that wearable device detection of near falls may be possible, but further validation with respect to detecting naturally occurring events in high-risk populations is now required. Given the current exponential advances and uptake of wearable technology, remote monitoring of near falls could provide new opportunities to identify older people who are at increased risk of falls before falls and related injuries occur.20 The new information could enable exercise and fall prevention programs to specifically target the types of near falls experienced and the daily activities or times of day during which near falls are experienced. The waist was the most commonly used location (7 studies) and accuracy was improved by multiple devices. However, the included studies were all of low to moderate quality and compared with the current state of knowledge on fall detection,5–10 near-fall detection is far less developed.
Can Near Falls in Older People Be Detected Using Wearable Devices?
Although the studies generally reported high accuracy to detect artificially induced near falls in younger adults, of the 3 studies that included people older than 60 years,27 , 29 , 31 none of them included multiple types of near falls or attempted to distinguish between ADLs and actual falls. These 3 studies in older people used various thresholds to identify “stumbles” during constant speed walking with high accuracy (85.7%-99.9%), but daily activities (such as sitting down or bending over) were not included in the analysis. Because daily activities involve changes in device orientation, they could have been mistaken for near falls and increased the false-positive rates.28 If daily activities had been included, it may have reduced the near-fall detection accuracy reported in older people. One study in older people reduced false positives by using an empirical decision tree that required a period of continuous gait before a stumble could be detected.29 However, this approach could reduce sensitivity to stumbles that occur during gait initiation. Furthermore, in this study, accuracy was reported using a “hit ratio” (93.1%), which makes comparisons to other studies that reported sensitivity problematic. In summary, there is yet insufficient evidence to determine whether naturally occurring near falls can be accurately detected in older people and distinguished from both actual falls and the movements that comprise many ADLs.
Can Near Falls Be Distinguished From Falls and Other ADLs?
Of the 4 studies that attempted to distinguish near falls from ADLs and/or actual falls,23 , 24 , 26 , 28 only 2 included both ADLs and actual falls,26 , 28 of which 1 study provided no information about the participants26 and so was difficult to interpret. The other study was rated most highly on the quality scale for this systematic review (5/7)28 and reported high sensitivity (95.2%) and specificity (97.6%). This study comprised 11 young (mean age of 27.6 years) healthy participants who simulated 5 different types of near falls, actual falls, and ADLs. Data from a waist-mounted device were combined with a preimpact velocity threshold algorithm to detect near falls. This suggests that in young people, laboratory-induced near falls can be distinguished from actual falls and other ADLs, but more studies in older people are required to confirm this.
Does the Incidence and Type of Near Falls Detected by Wearable Devices Differ With Age or Other Chronic Conditions?
Only 1 study investigated near falls in a high-risk cohort.29 This study comprised 40 older people (aged 41-81 years) with PD. Missteps were simulated in the laboratory and then remotely assessed during daily activities over 3 days. Data from a waist-mounted device were combined with a novel algorithm to detect near falls with good accuracy in the laboratory (hit ratio 93.1%). Accuracy during daily life was not assessed and because of the greater variability present in daily life data, accuracy during daily life is likely to be lower. With respect to clinical validity, participants with PD and a history of falls were found to have 23% more missteps than participants with PD and no history of falls.29 This suggests that remote assessments of near falls could provide a way to monitor age, health, and condition-related changes in physical functioning. However, only 1 study investigated 1 type of near fall in 1 high-risk cohort.29 The incidence and type of different types of near falls experienced by people with stroke, multiple sclerosis, age-related frailty, and other conditions should be the focus of further research.
Wearable Device Sensors, Placement, and Number
The heterogeneous methodologies (eg, types of near falls and sensors) in the reviewed studies, lack of detail, and replicable description of device placement made it difficult to determine the most appropriate location, type, or number of sensors. For example, 1 study reported that the right thigh was the best location to accurately detect near falls.24 In contrast, another study found that unilateral placements were less sensitive to near falls occurring on the other side of the body and reported a preference for the chest because it was less affected by noise and better able to detect upper body orientation changes resulting from a trip.25 Using multiple device locations may also improve accuracy,24 , 26 but acceptability or usability was not assessed and it was not clear whether the marginal improvement in detection accuracy by the increased number of wearable devices outweighed the increased cost of wearing them. More research is, therefore, required to determine the ideal device location and the tradeoff between an optimal technical solution and what is acceptable to older people.
Near-Fall Detection Algorithms
Findings from the included studies indicated that accurate detection of laboratory-induced near falls in younger adults was possible using several different algorithms. However, the 3 studies that used machine-learning algorithms23 , 24 , 26 were developed in small samples (n ≤ 15) using younger healthy people and, therefore, may have been overtrained. The generalizability of these 3 algorithms to accurately detect near falls in at-risk older adults or people with atypical movement patterns was not clear. Three studies used fixed threshold values to identify near falls.24 , 28 , 31 However, such single thresholds may detect only 1 type of near fall about a specific axis,29 and the fixed thresholds may be adversely affected by sensor placement, gait speed, and movements in different people.27 Some of these issues were successfully addressed in 2 studies by using multiple detection thresholds for different types of near falls29 and adaptive thresholds27 based on the individual's changing movement patterns.
Study Quality and Risk of Bias
Although the reported high accuracy of near-fall detection was encouraging, the included studies had several methodological issues that reduced their quality and may have increased the risk of bias. Importantly, no studies have yet reported accuracy for the detection of naturally occurring near falls in older people, and the registered scope of this systematic review was updated to include studies with people younger than 60 years. Methodological issues include small sample sizes, few near-fall events, lack of variation in the training data, and using young people in controlled environments to simulate the experiences of high-risk people in daily life.
These methodological issues increase the risk of overtraining or overfitting of the reported algorithms. During daily life, at least 5 types of near falls (slip, trip, incorrect weight transfer, misstep, and hit and bump) have been observed in older people.24 , 28 The lack of variation within the experimental protocols of several of the included studies could have led to overtraining of the algorithm—for example, to detect one type of near fall with increased accuracy at the expense of reduced accuracy to detect other types of near falls. Recording a greater number of near fall events would also have helped maintain the events per variable ratio greater than 10.33 The small number of near-fall events recorded by several studies could also have increased the risk of overfitting the data—for example, resulting in high reported accuracy for a specific cohort at the expense of reduced generalizability or accuracy to detect near falls in general populations.
Including data from ADLs, actual falls, and near falls during the training of the detection algorithms is essential to reduce the false-positive rate in future remote-monitoring situations. False detection of actual or near falls during an ADL (eg, erroneously detecting bending down as a near fall) could also lead to “alarm fatigue” or ignorance to the frequency of alarms. This could reduce both user and clinical acceptability of any new remote-monitoring technologies. Future research should, therefore, include larger studies that assess multiple types of near falls and attempt to distinguish near falls from ADLs and actual falls. Participants should include frail older people, people with chronic diseases, and people who are at high risk of fall injuries.
Potential for Clinical Utility
The emerging evidence for wearable technology for fall5–10 and near-fall23–31 detection is promising for a simple, low-cost method of monitoring people at high risk of falls. Identification of falls and near-fall events using wearable technologies will likely overcome the issue of bias, recall, and logistical difficulties of self- or caregiver reporting. Remote monitoring provides information about daily life performances, which are different from clinical assessments.32 It could enable alerts of deteriorating balance control, for example, in an older person who may be unable to identify or communicate such changes and, therefore, assist in guiding appropriate intervention before an actual fall occurs. However, more studies are required to determine the validity and reliability of wearable devices to detect near falls in high-risk populations during daily life.
Limitations of This Review
This systematic review should be considered in light of several limitations. A meta-analysis was not possible due to the small number of included studies, substantial risk of bias, and heterogeneous methodologies. The registered scope of this systematic review was updated to include studies with people younger than 60 years and full-length conference papers because a search with the initial criteria yielded no papers. The relative quality assessment tool was developed specifically for this review but has not been validated and would need to be adapted to be suitable for other systematic reviews.
The findings of this systematic review indicate that wearable devices can be used to detect laboratory-induced near falls in younger adults that occur in controlled environments with high accuracy. For this group and setting, wearable devices also appear to be able to distinguish the type of near fall, the type of daily activity, and the time of day that each near fall occurs. This information could be used to help develop more targeted exercise and fall prevention programs aimed at preventing fall injuries. However, most included study protocols comprised single or few types of near falls and did not attempt to distinguish near falls from actual falls and other ADLs. Although the waist was the most common location, due to the heterogeneous methodologies, it was difficult to determine the optimal placement or number of sensors. Similarly, the reported performance differences between algorithms were difficult to assess. It is not yet clear whether wearable devices can accurately detect near falls in older adults or other high-risk populations while at home or other community settings. Future larger and higher-quality studies should investigate multiple types of naturally occurring near falls in real-life settings with older people and people with different chronic conditions known to affect balance.
1. Tinetti ME, Speechley M, Ginter SF. Risk factors for falls among elderly persons living in the community. N Engl J Med. 1988;319(26):1701–1707.
2. Cali CM, Kiel DP. An epidemiologic study of fall
-related fractures among institutionalized older people. J Am Geriatr Soc. 1995;43(12):1336–1340.
3. Davis JC, Robertson MC, Ashe MC, Liu-Ambrose T, Khan KM, Marra CA. International comparison of cost of falls in older adults living in the community: a systematic review. Osteoporos Int. 2010;21(8):1295–1306.
4. Brodie MA, Okubo Y, Annegarn J, Wieching R, Lord SR, Delbaere K. Disentangling the health benefits of walking from increased exposure to falls in older people using remote gait monitoring and multi-dimensional analysis. Physiol Meas. 2017;38(1):45–62.
5. Ozdemir AT. An analysis on sensor locations of the human body for wearable fall
detection devices: principles and practice. Sensors. 2016;16(8):pii: E1161.
6. Kangas M, Konttila A, Lindgren P, Winblad I, Jamsa T. Comparison of low-complexity fall
detection algorithms for body attached accelerometers. Gait Posture. 2008;28(2):285–291.
7. Nyan MN, Tay FE, Murugasu E. A wearable
system for pre-impact fall
detection. J Biomech. 2008;41(16):3475–3481.
8. Chaudhuri S, Thompson H, Demiris G. Fall
detection devices and their use with older adults: a systematic review. J Geriatr Phys Ther. 2014;37(4):178–196.
9. Bianchi F, Redmond SJ, Narayanan MR, Cerutti S, Lovell NH. Barometric pressure and triaxial accelerometry-based falls event detection. IEEE Trans Neural Syst Rehabil Eng. 2010;18(6):619–627.
10. Casilari E, Luque R, Moron MJ. Analysis of android device-based solutions for fall
detection. Sensors. 2015;15(8):17827–17894.
11. Criter RE, Honaker JA. Identifying balance measures most likely to identify recent falls. J Geriatr Phys Ther. 2016;39(1):30–37.
12. Hirase T, Inokuchi S, Matsusaka N, Nakahara K, Okita M. A modified fall
risk assessment tool that is specific to physical function predicts falls in community-dwelling elderly people. J Geriatr Phys Ther. 2014;37(4):159–165.
13. Fabre JM, Ellis R, Kosma M, Wood RH. Falls risk factors and a compendium of falls risk screening instruments. J Geriatr Phys Ther. 2010;33(4):184–197.
14. Park SH. Tools for assessing fall
risk in the elderly: a systematic review and meta-analysis. Aging Clin Exp Res. 2018;30(1):1–16.
15. Lusardi MM, Fritz S, Middleton A, et al Determining risk of falls in community dwelling older adults: a systematic review and meta-analysis using posttest probability. J Geriatr Phys Ther. 2017;40(1):1–36.
16. Lord SR, Delbaere K, Gandevia SC. Use of a physiological profile to document motor impairment in ageing and in clinical groups. J Physiol. 2016;594(16):4513–4523.
17. Arnold CM, Faulkner RA. The history of falls and the association of the Timed Up and Go test to falls and near
-falls in older adults with hip osteoarthritis. BMC Geriatr. 2007;7(1):17.
18. Wiles CM, Busse ME, Sampson CM, Rogers MT, Fenton-May J, van Deursen R. Falls and stumbles in myotonic dystrophy. J Neurol Neurosurg Psychiatry. 2006;77(3):393–396.
19. Srygley JM, Herman T, Giladi N, Hausdorff JM. Self-report of missteps in older adults: a valid proxy of fall
risk? Arch Phys Med Rehabil. 2009;90(5):786–792.
20. Nagai K, Yamada M, Komatsu M, et al Near
falls predict substantial falls in older adults: a prospective cohort study. Geriatr Gerontol Int. 2017;17(10):1477–1480.
21. Moher D, Liberati A, Tetzlaff J, Altman DG. Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. Ann Intern Med. 2009;151(4):264–269.
22. Ryan JW, Dinkel JL, Petrucci K. Near
falls incidence. A study of older adults in the community. J Gerontol Nurs. 1993;19(12):23–28.
23. Albert MV, Kording K, Herrmann M, Jayaraman A. Fall
classification by machine learning using mobile phones. PloS One. 2012;7(5):e36556.
24. Aziz O, Park EJ, Mori G, Robinovitch SN. Distinguishing near
-falls from daily activities with wearable
accelerometers and gyroscopes using Support Vector Machines. Conf Proc IEEE Eng Med Biol Soc. 2012:5837–5840.
25. Hajj Chehade N, Ozisik P, Gomez J, Ramos F, Pottie G. Detecting stumbles with a single accelerometer
. Conf Proc IEEE Eng Med Biol Soc. 2012;2012:6681–6686.
26. Choi Y, Ralhan AS, Ko S. A study on machine learning algorithms for fall
detection and movement classification. In: Conference Proceedings from Information Science and Applications, 2011; Jeju Island, Korea.
27. Karel JM, Senden R, Janssen JE, et al Towards unobtrusive in vivo monitoring of patients prone to falling. Conf Proc IEEE Eng Med Biol Soc. 2010;2010:5018–5021.
28. Lee JK, Robinovitch SN, Park EJ. Inertial sensing-based pre-impact detection of falls involving near
scenarios. IEEE Trans Neural Syst Rehabil Eng. 2015;23(2):258–266.
29. Iluz T, Gazit E, Herman T, et al Automated detection of missteps during community ambulation in patients with Parkinson's disease: a new approach for quantifying fall
risk in the community setting. J Neuroeng Rehabil. 2014;11(1):48.
30. Trkov M, Chen K, Yi J, Liu T. Slip detection and prediction in human walking using only wearable
inertial measurement units (IMUs). In: Conference Proceedings from IEEE/ASME AIM, July 7, 2015, Busan, Korea.
31. Weiss A, Shimkin I, Giladi N, Hausdorff JM. Automated detection of near
falls: algorithm development and preliminary results. BMC Res Notes. 2010;3:62.
32. Brodie MA, Coppens MJ, Lord SR, et al Wearable
pendant device monitoring using new wavelet-based methods shows daily life and laboratory gaits are different. Med Biol Eng Comput. 2016;54(4):663–674.
33. Peduzzi P, Concato J, Kemper E, Holford TR, Feinstein AR. A simulation study of the number of events per variable in logistic regression analysis. J Clin Epidemiol. 1996;49(12):1373–1379.