Gait & Posture 39 (2014) 991–994
Contents lists available at ScienceDirect
Gait & Posture journal homepage: www.elsevier.com/locate/gaitpost
Short Communication
Can we use accelerometry to monitor balance exercise performance in older adults? James Y. Tung a,b,*, Helen Ng a, Cameron Moore a, Lora Giangregorio a,c a
Department of Kinesiology, University of Waterloo, Waterloo, ON N2L 3G1, Canada David Cheriton School of Computer Science, University of Waterloo, Waterloo, ON, N2L 3G1, Canada c Toronto Rehabilitation Institute, University Health Network, Toronto, ON, M5G 2A2, Canada b
A R T I C L E I N F O
A B S T R A C T
Article history: Received 31 May 2013 Received in revised form 31 October 2013 Accepted 25 November 2013
While home-based balance exercises are recommended to reduce the risk of falling and fractures in older adults, adherence to exercise remains suboptimal. The long-term objective of this research is to advance body-worn sensor techniques to measure at-home exercise performance and promote adherence. In this study, a method of distinguishing 5 types of walking using hip- and ankle-worn accelerometers was developed and evaluated in a target clinical population. A secondary objective was to evaluate the method’s sensitivity to sensor placement. Eighteen community-dwelling, older females (50 years) with low bone mass wore triaxial accelerometers at the left hip and each ankle while performing 5 walking tasks at home: 4 walking balance exercises (figure 8, heel-toe, sidestep, backwards) and straight-line walking. Sensor data were separated into low (0.5–2 Hz) and high (2–10 Hz) frequency bands, and rootmean-square values (energy) were computed for each sensor, axis, and band. These 18 energy estimates were used as inputs to a neural network classifier with 5 outputs, corresponding to each task. Using a leave-one-out cross-validation protocol, the neural network correctly classified 82/90 test instances (91% accuracy). Compared to random selection accuracy of 20% (i.e., 1 in 5), the results indicated excellent separation between tasks. Reducing the sensor set to one hip and one ankle resulted in 6.7– 8.9% reduction in accuracy. Our findings can be used in the development of tools used to deliver exercise performance metrics (e.g., % completed) or recognize walking and balance exercise activities using bodyworn accelerometers. ß 2013 Elsevier B.V. All rights reserved.
Keywords: Falls Fractures Balance exercise Osteoporosis Low bone mass
1. Introduction Considering injuries arising from falls are a major cause of mortality and morbidity in the elderly, preventing falls is a high priority in geriatric medicine [1]. While exercise programs have been shown to reduce fall risk [2–4], maintaining these benefits relies on sustained adherence to exercise. However, long-term adherence to home-based exercise programs for older adults remains suboptimal. In a review of 7 trials (n = 747) examining the effectiveness of the Otago Exercise Program, less than 40% of participants continued exercising at 12 months [2]. The Lifestyle integrated Functional Exercise program, designed to promote activity by integrating exercises into daily activities, demonstrated similar adherence levels (50 years) with self-reported low bone mass (i.e., ‘‘Has a doctor ever told you that you have osteopenia, low bone mass, or osteoporosis?’’) or prior non-traumatic fracture were recruited from local osteoporosis support groups. The population was chosen to inform activity monitoring for future trials in individuals with osteoporosis. Individuals who reported chest pain on exertion or uncontrolled hypertension, history of neurologic conditions, or contraindications to exercise were excluded from participation. This cross-sectional study was approved by the Office of Research Ethics (University of Waterloo).
Under researcher supervision, participants performed 5 types of walking in their own home: 4 exercises to improving dynamic balance chosen from the Otago Exercise Program designed to improve balance and prevent falls [10] (figure 8, heel-toe (tandem), side-stepping, and backwards walking) and forward walking. Each task was performed 4 times along a hallway or wall, with total task duration ranging from 25 to 70 s per task. Participants performed the exercises 2–4 months following initial assessment. 2.5. Acceleration signal processing
2.2. Participant characteristics Participants completed a health history questionnaire to gather socio-demographic, fracture history, and falls (past 6 months) information to describe the study population. To describe functional mobility and lower leg strength, gait speed and five times sit-to-stand tests (FTSS) were performed. Gait speed was measured by taking the mean of 3 walking trips at participants’ preferred speed over a 15-foot pressure-sensitive walkway (GAITRite System, CIR Systems Inc., Clifton, NY). In the FTSS test, participants stood up from sitting on a chair and sat down again five times as quickly as possible, without using their arms to push off [9]. Characteristics were measured at the osteoporosis support group meetings as part of a related study. A subset of participants consented to participate in the second part of the study to wear accelerometers and perform dynamic balance exercises. 2.3. Accelerometers Participants wore 3-D accelerometers (X6-2 Mini, Gulf Coast Data Concepts Inc., 40 Hz sample rate, 6 g range, 12-bit resolution)
Walking types were characterized using acceleration data to indicate activity from each sensor and axis (approximating sagittal, frontal, and vertical planes). Signals were first high-pass filtered (0.25 Hz) to remove gravity effects, then separated into low (0.5– 2 Hz) and high (2–10 Hz) frequency bands. Energy, calculated as the root-mean-square values of each channel, was computed for each sensor, axis, and band for 18 energy estimates. Matlab R2012b (Mathworks, Inc.) was used process the data. 2.6. Neural network classifier To develop acceleration profiles corresponding to each walking type, a neural network classifier was constructed with 18 inputs (i.e., energy estimates), 12 hidden layers, and 5 outputs (i.e., walking types). To train and evaluate the classifier, a leave-one-out cross-validation procedure was performed. In this procedure, data from one participant was reserved for validation, and the remaining data (n = 17) was used to map acceleration profiles to walking types (i.e., training). The ability of the trained classifier to distinguish walking types was tested using the reserved validation
Table 1 Descriptive statistics of energy features (RMS values (m/s2)) by walking exercise. Figure 8s Mean Left hip Low-band (0.5–2 Hz) XRMS,low 0.72 YRMS,low 0.50 ZRMS,low 0.78 Hi-band (2–10 Hz) XRMS,hi 0.41 YRMS,hi 0.30 ZRMS,hi 0.43 Left ankle Low-band (0.5–2 Hz) XRMS,low 2.22 YRMS,low 1.40 ZRMS,low 1.78 Hi-band (2–10 Hz) XRMS,hi 1.42 YRMS,hi 0.78 ZRMS,hi 1.11 Right ankle Low-band (0.5–2 Hz) XRMS,low 2.36 YRMS,low 1.52 ZRMS,low 1.94 Hi-band (2–10 Hz) XRMS,hi 1.30 YRMS,hi 0.86 ZRMS,hi 1.20
Heel-toe
Sidestepping
Backwards
Normal
SD
Mean
SD
Mean
SD
Mean
SD
Mean
SD
0.16 0.16 0.15
0.55 0.26 0.53
0.13 0.08 0.12
0.68 0.62 1.28
0.25 0.35 0.39
0.72 0.39 0.76
0.23 0.17 0.18
1.03 1.14 0.77
0.35 0.31 0.14
0.10 0.09 0.08
0.32 0.16 0.31
0.07 0.05 0.07
0.38 0.41 0.71
0.14 0.25 0.23
0.42 0.23 0.43
0.12 0.09 0.11
0.60 0.67 0.45
0.22 0.21 0.09
0.67 0.45 0.49
1.68 0.97 0.91
0.19 0.26 0.17
1.05 0.56 1.77
0.47 0.19 0.55
1.67 0.78 0.90
0.36 0.18 0.26
3.05 2.23 1.44
1.55 0.79 0.73
0.48 0.27 0.33
1.08 0.59 0.52
0.14 0.20 0.13
0.66 0.34 0.86
0.24 0.08 0.29
1.11 0.45 0.56
0.27 0.08 0.21
1.86 1.16 0.80
0.92 0.42 0.45
0.70 0.47 0.61
1.60 0.98 1.02
0.42 0.25 0.30
1.25 0.61 1.66
0.51 0.19 0.45
1.52 0.98 1.10
0.42 0.44 0.50
2.94 2.18 1.57
1.30 0.72 0.94
0.51 0.29 0.41
0.92 0.60 0.58
0.21 0.14 0.21
0.74 0.37 0.85
0.32 0.14 0.24
0.85 0.61 0.62
0.31 0.32 0.28
1.55 1.22 0.85
0.73 0.45 0.57
J.Y. Tung et al. / Gait & Posture 39 (2014) 991–994
993
Table 2 Confusion table and accuracy results from leave-one-out cross-validation (n = 18). Type performed
Classified as Figure 8
Heel-toe
Sidestep
Backwards
Straight-line
Accuracy (%)
Figure 8 Heel-toe Sidestep Backwards Straight-line Overall
16 1 0 0 0
0 16 0 1 0
0 0 18 0 0
2 1 0 16 2
0 0 0 1 16
88.9 88.9 100.0 88.9 88.9 91.1
case. The procedure is then repeated 16 times using each participant once as the validation case (i.e., bootstrapping).
4. Discussion
Participant mean (SD) age was 74.4 7.2 years. Osteopenia and osteoporosis were reported by 44% and 56%, respectively, with 56% reporting a fracture after age 40, and 17% reporting a fall in the past 6 months. While gait speeds (1.24 0.20 m/s, range = 0.91–1.60) indicate good walking ability, the FTSS times (12.0 2.8 s, range = 8.72–18.20) indicated that some participants lacked transfer capabilities. Six of 18 participants (33%) performed the FTSS more slowly than the 12 s cut-off indicating elevated fall risk [9].
We developed and tested a method to distinguish between 5 walking types using hip and ankle accelerometers in older adults at risk for fractures. Considering a random selection accuracy of 20% (i.e., 1 in 5 choices), the 91.1% overall accuracy suggests that accelerometers could be used to identify exercises aimed at improving dynamic balance. In contrast with other algorithms [8], these results were achieved without tailoring the classifier to each participant. Importantly, the algorithm was validated using data from our target population: community-dwelling older adults with low bone mass. Our findings will inform the development of tools to track at-home exercise performance to monitor adherence, provide feedback, and inform coaching strategies. We have demonstrated a trade-off between sensor placement and classification accuracy. The most accurate combination was using data from all three sources (hip, left and right ankles). The cost of going from three sensors to two (1 hip, 1 ankle) is 6.7–8.9%, likely attributable to reduced ability to capture turning where leg motions are directionally-dependent. While most applications of accelerometers employ a single monitor, applications using multiple sensors are emerging [7]. Our study represents a step towards a robust multi-sensor monitor to track not only activity levels, but exercise performance. Considering the potential impact of orientation errors, further evaluation of the sensitivity to variable sensor placement is warranted.
3.2. Neural network classifier accuracy
Acknowledgment
Overall, the algorithm accurately classified 82/90 (91.1%) test instances (Table 2). The majority (5/8) of errors were incorrectly classified as backwards walking. In contrast, all sidestepping trials were accurately classified.
We acknowledge the assistance of Stephanie Cistrone and Carly Skidmore with collection of data for this study.
2.7. Data analyses For participant characteristics, mean and standard deviation (SD) were used to describe continuous data and count (%) were used to represent categorical data. Classifier accuracy was measured using the count and percent of correct classifications. To examine the impact of sensor placement on the ability to distinguish tasks, the neural network training and cross-validation procedure was performed using all possible combinations of sensors (Table 1). 3. Results 3.1. Participant characteristics
Conflict of interest statement
3.3. Sensor placement sensitivity The authors have no conflicts of interest to declare. Generally, reducing the number of sensors resulted in decreased accuracy ranging from 91.1% (3 sensors), 73.3–84.4% (2 sensors), to 64.4–70.0% (1 sensor, Table 3). Amongst 2 sensor combinations, the left + right ankle combination (73.3%) produced less accurate results than the hip and ankle sensor combination (hip + left, 84.4%; hip + right, 82.2%).
Table 3 Sensor placement sensitivity analysis. Sensors
Accuracy rate (%)
False positive rate (%)
Hip + left + right Hip + left Hip + right Left + right Left Hip Right
91.1 84.4 82.2 73.3 70.0 64.6 64.4
2.2 3.9 4.4 6.7 7.5 8.6 8.9
References [1] Rubenstein LZ. Falls in older people: epidemiology, risk factors and strategies for prevention. Age Ageing 2006;vol. 35(September (Suppl. 2)):ii37–41. [2] Thomas S, Mackintosh S, Halbert J. Does the ‘Otago exercise programme’ reduce mortality and falls in older adults?: a systematic review and metaanalysis. Age Ageing 2010;39(November (6)):681–7. [3] Clemson L, Fiatarone Singh MA, Bundy A, Cumming RG, Manollaras K, O’Loughlin P, et al. Integration of balance and strength training into daily life activity to reduce rate of falls in older people (the LiFE study): randomised parallel trial. BMJ 2012;345(January):e4547. [4] Gillespie LD, Robertson MC, Gillespie WJ, Lamb SE, Gates S, Cumming RG, et al. Interventions for preventing falls in older people living in the community. Cochrane Database Syst Rev 2009;2:CD007146. [5] Preece SJ, Goulermas JY, Kenney LPJ, Howard D, Meijer K, Crompton R. Activity identification using body-mounted sensors – a review of classification techniques. Physiol Meas 2009;30(April (4)):R1–33. [6] Taraldsen K, Chastin SFM, Riphagen II, Vereijken B, Helbostad JL. Physical activity monitoring by use of accelerometer-based body-worn sensors in older adults: a systematic literature review of current knowledge and applications. Maturitas 2012;71(January (1)):13–9.
994
J.Y. Tung et al. / Gait & Posture 39 (2014) 991–994
[7] Cheung VH, Gray L, Karunanithi M. Review of accelerometry for determining daily activity among elderly patients. Arch Phys Med Rehabil 2011;92(June (6)):998–1014. [8] Dobkin BH, Xu X, Batalin M, Thomas S, Kaiser W. Reliability and validity of bilateral ankle accelerometer algorithms for activity recognition and walking speed after stroke. Stroke 2011;42(August (8)):2246–50.
[9] Tiedemann A, Shimada H, Sherrington C, Murray S, Lord S. The comparative ability of eight functional mobility tests for predicting falls in communitydwelling older people. Age Ageing 2008;37(July (4)):430–5. [10] Gardner MM, Buchner DM, Robertson MC, Campbell J. Practical implementation of an exercise-based falls prevention programme. Age Ageing 2001;30(January (1)):77–83.
Contents lists available at ScienceDirect
Gait & Posture journal homepage: www.elsevier.com/locate/gaitpost
Short Communication
Can we use accelerometry to monitor balance exercise performance in older adults? James Y. Tung a,b,*, Helen Ng a, Cameron Moore a, Lora Giangregorio a,c a
Department of Kinesiology, University of Waterloo, Waterloo, ON N2L 3G1, Canada David Cheriton School of Computer Science, University of Waterloo, Waterloo, ON, N2L 3G1, Canada c Toronto Rehabilitation Institute, University Health Network, Toronto, ON, M5G 2A2, Canada b
A R T I C L E I N F O
A B S T R A C T
Article history: Received 31 May 2013 Received in revised form 31 October 2013 Accepted 25 November 2013
While home-based balance exercises are recommended to reduce the risk of falling and fractures in older adults, adherence to exercise remains suboptimal. The long-term objective of this research is to advance body-worn sensor techniques to measure at-home exercise performance and promote adherence. In this study, a method of distinguishing 5 types of walking using hip- and ankle-worn accelerometers was developed and evaluated in a target clinical population. A secondary objective was to evaluate the method’s sensitivity to sensor placement. Eighteen community-dwelling, older females (50 years) with low bone mass wore triaxial accelerometers at the left hip and each ankle while performing 5 walking tasks at home: 4 walking balance exercises (figure 8, heel-toe, sidestep, backwards) and straight-line walking. Sensor data were separated into low (0.5–2 Hz) and high (2–10 Hz) frequency bands, and rootmean-square values (energy) were computed for each sensor, axis, and band. These 18 energy estimates were used as inputs to a neural network classifier with 5 outputs, corresponding to each task. Using a leave-one-out cross-validation protocol, the neural network correctly classified 82/90 test instances (91% accuracy). Compared to random selection accuracy of 20% (i.e., 1 in 5), the results indicated excellent separation between tasks. Reducing the sensor set to one hip and one ankle resulted in 6.7– 8.9% reduction in accuracy. Our findings can be used in the development of tools used to deliver exercise performance metrics (e.g., % completed) or recognize walking and balance exercise activities using bodyworn accelerometers. ß 2013 Elsevier B.V. All rights reserved.
Keywords: Falls Fractures Balance exercise Osteoporosis Low bone mass
1. Introduction Considering injuries arising from falls are a major cause of mortality and morbidity in the elderly, preventing falls is a high priority in geriatric medicine [1]. While exercise programs have been shown to reduce fall risk [2–4], maintaining these benefits relies on sustained adherence to exercise. However, long-term adherence to home-based exercise programs for older adults remains suboptimal. In a review of 7 trials (n = 747) examining the effectiveness of the Otago Exercise Program, less than 40% of participants continued exercising at 12 months [2]. The Lifestyle integrated Functional Exercise program, designed to promote activity by integrating exercises into daily activities, demonstrated similar adherence levels (50 years) with self-reported low bone mass (i.e., ‘‘Has a doctor ever told you that you have osteopenia, low bone mass, or osteoporosis?’’) or prior non-traumatic fracture were recruited from local osteoporosis support groups. The population was chosen to inform activity monitoring for future trials in individuals with osteoporosis. Individuals who reported chest pain on exertion or uncontrolled hypertension, history of neurologic conditions, or contraindications to exercise were excluded from participation. This cross-sectional study was approved by the Office of Research Ethics (University of Waterloo).
Under researcher supervision, participants performed 5 types of walking in their own home: 4 exercises to improving dynamic balance chosen from the Otago Exercise Program designed to improve balance and prevent falls [10] (figure 8, heel-toe (tandem), side-stepping, and backwards walking) and forward walking. Each task was performed 4 times along a hallway or wall, with total task duration ranging from 25 to 70 s per task. Participants performed the exercises 2–4 months following initial assessment. 2.5. Acceleration signal processing
2.2. Participant characteristics Participants completed a health history questionnaire to gather socio-demographic, fracture history, and falls (past 6 months) information to describe the study population. To describe functional mobility and lower leg strength, gait speed and five times sit-to-stand tests (FTSS) were performed. Gait speed was measured by taking the mean of 3 walking trips at participants’ preferred speed over a 15-foot pressure-sensitive walkway (GAITRite System, CIR Systems Inc., Clifton, NY). In the FTSS test, participants stood up from sitting on a chair and sat down again five times as quickly as possible, without using their arms to push off [9]. Characteristics were measured at the osteoporosis support group meetings as part of a related study. A subset of participants consented to participate in the second part of the study to wear accelerometers and perform dynamic balance exercises. 2.3. Accelerometers Participants wore 3-D accelerometers (X6-2 Mini, Gulf Coast Data Concepts Inc., 40 Hz sample rate, 6 g range, 12-bit resolution)
Walking types were characterized using acceleration data to indicate activity from each sensor and axis (approximating sagittal, frontal, and vertical planes). Signals were first high-pass filtered (0.25 Hz) to remove gravity effects, then separated into low (0.5– 2 Hz) and high (2–10 Hz) frequency bands. Energy, calculated as the root-mean-square values of each channel, was computed for each sensor, axis, and band for 18 energy estimates. Matlab R2012b (Mathworks, Inc.) was used process the data. 2.6. Neural network classifier To develop acceleration profiles corresponding to each walking type, a neural network classifier was constructed with 18 inputs (i.e., energy estimates), 12 hidden layers, and 5 outputs (i.e., walking types). To train and evaluate the classifier, a leave-one-out cross-validation procedure was performed. In this procedure, data from one participant was reserved for validation, and the remaining data (n = 17) was used to map acceleration profiles to walking types (i.e., training). The ability of the trained classifier to distinguish walking types was tested using the reserved validation
Table 1 Descriptive statistics of energy features (RMS values (m/s2)) by walking exercise. Figure 8s Mean Left hip Low-band (0.5–2 Hz) XRMS,low 0.72 YRMS,low 0.50 ZRMS,low 0.78 Hi-band (2–10 Hz) XRMS,hi 0.41 YRMS,hi 0.30 ZRMS,hi 0.43 Left ankle Low-band (0.5–2 Hz) XRMS,low 2.22 YRMS,low 1.40 ZRMS,low 1.78 Hi-band (2–10 Hz) XRMS,hi 1.42 YRMS,hi 0.78 ZRMS,hi 1.11 Right ankle Low-band (0.5–2 Hz) XRMS,low 2.36 YRMS,low 1.52 ZRMS,low 1.94 Hi-band (2–10 Hz) XRMS,hi 1.30 YRMS,hi 0.86 ZRMS,hi 1.20
Heel-toe
Sidestepping
Backwards
Normal
SD
Mean
SD
Mean
SD
Mean
SD
Mean
SD
0.16 0.16 0.15
0.55 0.26 0.53
0.13 0.08 0.12
0.68 0.62 1.28
0.25 0.35 0.39
0.72 0.39 0.76
0.23 0.17 0.18
1.03 1.14 0.77
0.35 0.31 0.14
0.10 0.09 0.08
0.32 0.16 0.31
0.07 0.05 0.07
0.38 0.41 0.71
0.14 0.25 0.23
0.42 0.23 0.43
0.12 0.09 0.11
0.60 0.67 0.45
0.22 0.21 0.09
0.67 0.45 0.49
1.68 0.97 0.91
0.19 0.26 0.17
1.05 0.56 1.77
0.47 0.19 0.55
1.67 0.78 0.90
0.36 0.18 0.26
3.05 2.23 1.44
1.55 0.79 0.73
0.48 0.27 0.33
1.08 0.59 0.52
0.14 0.20 0.13
0.66 0.34 0.86
0.24 0.08 0.29
1.11 0.45 0.56
0.27 0.08 0.21
1.86 1.16 0.80
0.92 0.42 0.45
0.70 0.47 0.61
1.60 0.98 1.02
0.42 0.25 0.30
1.25 0.61 1.66
0.51 0.19 0.45
1.52 0.98 1.10
0.42 0.44 0.50
2.94 2.18 1.57
1.30 0.72 0.94
0.51 0.29 0.41
0.92 0.60 0.58
0.21 0.14 0.21
0.74 0.37 0.85
0.32 0.14 0.24
0.85 0.61 0.62
0.31 0.32 0.28
1.55 1.22 0.85
0.73 0.45 0.57
J.Y. Tung et al. / Gait & Posture 39 (2014) 991–994
993
Table 2 Confusion table and accuracy results from leave-one-out cross-validation (n = 18). Type performed
Classified as Figure 8
Heel-toe
Sidestep
Backwards
Straight-line
Accuracy (%)
Figure 8 Heel-toe Sidestep Backwards Straight-line Overall
16 1 0 0 0
0 16 0 1 0
0 0 18 0 0
2 1 0 16 2
0 0 0 1 16
88.9 88.9 100.0 88.9 88.9 91.1
case. The procedure is then repeated 16 times using each participant once as the validation case (i.e., bootstrapping).
4. Discussion
Participant mean (SD) age was 74.4 7.2 years. Osteopenia and osteoporosis were reported by 44% and 56%, respectively, with 56% reporting a fracture after age 40, and 17% reporting a fall in the past 6 months. While gait speeds (1.24 0.20 m/s, range = 0.91–1.60) indicate good walking ability, the FTSS times (12.0 2.8 s, range = 8.72–18.20) indicated that some participants lacked transfer capabilities. Six of 18 participants (33%) performed the FTSS more slowly than the 12 s cut-off indicating elevated fall risk [9].
We developed and tested a method to distinguish between 5 walking types using hip and ankle accelerometers in older adults at risk for fractures. Considering a random selection accuracy of 20% (i.e., 1 in 5 choices), the 91.1% overall accuracy suggests that accelerometers could be used to identify exercises aimed at improving dynamic balance. In contrast with other algorithms [8], these results were achieved without tailoring the classifier to each participant. Importantly, the algorithm was validated using data from our target population: community-dwelling older adults with low bone mass. Our findings will inform the development of tools to track at-home exercise performance to monitor adherence, provide feedback, and inform coaching strategies. We have demonstrated a trade-off between sensor placement and classification accuracy. The most accurate combination was using data from all three sources (hip, left and right ankles). The cost of going from three sensors to two (1 hip, 1 ankle) is 6.7–8.9%, likely attributable to reduced ability to capture turning where leg motions are directionally-dependent. While most applications of accelerometers employ a single monitor, applications using multiple sensors are emerging [7]. Our study represents a step towards a robust multi-sensor monitor to track not only activity levels, but exercise performance. Considering the potential impact of orientation errors, further evaluation of the sensitivity to variable sensor placement is warranted.
3.2. Neural network classifier accuracy
Acknowledgment
Overall, the algorithm accurately classified 82/90 (91.1%) test instances (Table 2). The majority (5/8) of errors were incorrectly classified as backwards walking. In contrast, all sidestepping trials were accurately classified.
We acknowledge the assistance of Stephanie Cistrone and Carly Skidmore with collection of data for this study.
2.7. Data analyses For participant characteristics, mean and standard deviation (SD) were used to describe continuous data and count (%) were used to represent categorical data. Classifier accuracy was measured using the count and percent of correct classifications. To examine the impact of sensor placement on the ability to distinguish tasks, the neural network training and cross-validation procedure was performed using all possible combinations of sensors (Table 1). 3. Results 3.1. Participant characteristics
Conflict of interest statement
3.3. Sensor placement sensitivity The authors have no conflicts of interest to declare. Generally, reducing the number of sensors resulted in decreased accuracy ranging from 91.1% (3 sensors), 73.3–84.4% (2 sensors), to 64.4–70.0% (1 sensor, Table 3). Amongst 2 sensor combinations, the left + right ankle combination (73.3%) produced less accurate results than the hip and ankle sensor combination (hip + left, 84.4%; hip + right, 82.2%).
Table 3 Sensor placement sensitivity analysis. Sensors
Accuracy rate (%)
False positive rate (%)
Hip + left + right Hip + left Hip + right Left + right Left Hip Right
91.1 84.4 82.2 73.3 70.0 64.6 64.4
2.2 3.9 4.4 6.7 7.5 8.6 8.9
References [1] Rubenstein LZ. Falls in older people: epidemiology, risk factors and strategies for prevention. Age Ageing 2006;vol. 35(September (Suppl. 2)):ii37–41. [2] Thomas S, Mackintosh S, Halbert J. Does the ‘Otago exercise programme’ reduce mortality and falls in older adults?: a systematic review and metaanalysis. Age Ageing 2010;39(November (6)):681–7. [3] Clemson L, Fiatarone Singh MA, Bundy A, Cumming RG, Manollaras K, O’Loughlin P, et al. Integration of balance and strength training into daily life activity to reduce rate of falls in older people (the LiFE study): randomised parallel trial. BMJ 2012;345(January):e4547. [4] Gillespie LD, Robertson MC, Gillespie WJ, Lamb SE, Gates S, Cumming RG, et al. Interventions for preventing falls in older people living in the community. Cochrane Database Syst Rev 2009;2:CD007146. [5] Preece SJ, Goulermas JY, Kenney LPJ, Howard D, Meijer K, Crompton R. Activity identification using body-mounted sensors – a review of classification techniques. Physiol Meas 2009;30(April (4)):R1–33. [6] Taraldsen K, Chastin SFM, Riphagen II, Vereijken B, Helbostad JL. Physical activity monitoring by use of accelerometer-based body-worn sensors in older adults: a systematic literature review of current knowledge and applications. Maturitas 2012;71(January (1)):13–9.
994
J.Y. Tung et al. / Gait & Posture 39 (2014) 991–994
[7] Cheung VH, Gray L, Karunanithi M. Review of accelerometry for determining daily activity among elderly patients. Arch Phys Med Rehabil 2011;92(June (6)):998–1014. [8] Dobkin BH, Xu X, Batalin M, Thomas S, Kaiser W. Reliability and validity of bilateral ankle accelerometer algorithms for activity recognition and walking speed after stroke. Stroke 2011;42(August (8)):2246–50.
[9] Tiedemann A, Shimada H, Sherrington C, Murray S, Lord S. The comparative ability of eight functional mobility tests for predicting falls in communitydwelling older people. Age Ageing 2008;37(July (4)):430–5. [10] Gardner MM, Buchner DM, Robertson MC, Campbell J. Practical implementation of an exercise-based falls prevention programme. Age Ageing 2001;30(January (1)):77–83.
