Inform Health Soc Care, 2014; 39(3–4): 249–261 ! Informa UK Ltd. ISSN: 1753-8157 print / 1753-8165 online DOI: 10.3109/17538157.2014.931851
A prospective field study for sensor-based identification of fall risk in older people with dementia Matthias Gietzelt,1 Florian Feldwieser,2 Mehmet Go ¨ vercin,2 Elisabeth Steinhagen-Thiessen,2 and Michael Marschollek3 1
Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig – Institute of Technology and Hannover Medical School, Braunschweig, Germany, 2 Geriatric Research Group, Department of Geriatric Medicine, Charite´ – Universita¨tsmedizin Berlin, Berlin, Germany, 3 Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig – Institute of Technology and Hannover Medical School, Hannover, Germany Objective: Aim of this study was to make a fall prognosis in a cohort of older people with dementia in short-term (2 month), mid-term (4 month) and long-term (8 month) intervals using accelerometry during the subjects’ everyday life. Methods: The study was designed as a longitudinal cohort study. The subjects were recruited from a nursing home and geriatric assessment tests were conducted at baseline. Each subject underwent four visits and was measured at each visit for one week. Gait episodes were detected and gait parameters were extracted from these episodes. These gait parameters were combined with the falls occurred during the study. A decision tree induction method was used to analyze the data. Results: Forty subjects participated in the study, whereby 12 drop-outs were registered. The geriatric assessment tests were unable to distinguish between the groups (AUC50.6). The evaluation of the models induced with the decision tree classification showed a rate of correctly classified gait episodes of 88.4% for short-term, 74.8% for midterm, and 88.5 % for long-term monitoring. Discussion and conclusions: We concluded that it is possible to classify gait episodes of fallers and non-fallers in people with dementia during everyday life using accelerometry. Keywords Accelerometer, dementia, fall risk, health-enabling technologies
INTRODUCTION Thirty-seven percent of women and 25% of men aged 65 years and older suffer from at least one fall event per year. This percentage increases with age, so that about 80% of all people older than 80 years fall at least once a year (1). The estimated health-related costs of falls and their consequences in people older than 65 years in the US vary between US$3476 and US$10 749 per fall.
Correspondence: Matthias Gietzelt, Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig – Institute of Technology and Hannover Medical School, Mu¨hlenpfordtstr. 23, D-38106 Braunschweig, Germany. Tel: +49 531 391 9505. Fax: +49 531 391 9502. E-mail: [email protected]
250
M. Gietzelt et al.
In Germany, fall related costs are valued at one billion Euros per year (2). Within hospital wards, specific geriatric assessments are typically performed for the estimation of fall risk. The most frequently used tests are the Timed ‘‘Up & Go’’ test [TUG] (3), the Tinetti test (4), the Functional Reach test (5), and the STRATIFY score (6). Common fall risk factors are: abnormal gait, lower limb weakness, presence of confusion or disorientation, altered mental status, agitation, urinary incontinence, and fall history (7). However, the commonly used clinical fall risk assessment tools that are based on these consistently recurrent risk factors, only achieve moderate to poor results in at least one of the statistical parameters sensitivity, specificity, positive and negative predictive value [PPV and NPV] (7,8). This fact leads to a massive over- or underestimation of fallers or non-fallers. Specialized fall prevention programs seem to be more cost-effective than treatment and care after a fall event has been occurred (9). It is still difficult to correctly identify those who are at risk of falling, which is why fall risk assessment is still a subject of research. In the context of cost-effectiveness it is important that prevention programs are only administered to persons that are at risk of falling and not across a broad population. Currently, the trend tends to technical solutions rather than to classic fall risk assessment scores and tests, because of the potentially expected improvements in objectiveness and reliability. A common approach is to use wearable sensors [accelerometers and gyroscopes] (10), because of their small size and low price. Other approaches use camera systems (11), the GAITRite M2 sensor floor [CIR Systems Inc., Sparta, NJ], or the S4 Sensors Walkway [Tactex and S4 Sensors, Victoria, BC, Canada] (12). Related work Most authors tend to use recent fall history of subjects in order to make a fall prognosis (13). A recent fall history might be a good indicator for a higher risk of falling, but is not a robust predictor that a fall will actually happen in the near future. Other authors use a variety of fall risk assessment tools and scores for prognosis (10,14,15). These tools, however, may have a limited capability for fall prognosis. Studies that documented falls prospectively are rare and to the authors’ best knowledge currently only four studies exist, which make a prospective fall prognosis based on falls documented after the sensor-based measurements took place (16–19). These studies used an accelerometry-based gait analysis and their hypothesis was that there are archetypical gait patterns, which are predictive for falls. The subjects were called by phone in a followup after one (16–18) or two years (19), respectively. This might be a methodological flaw, because the subjects may have overlooked a fall happened during that time and there is no possibility to control the information of the subjects. Furthermore, subjects tend to exclude or ignore fall events, which caused no injuries. Therefore, it seems to be necessary to have access to a professional documentation of fall events for these kinds of studies. The four mentioned studies used sensor-based measurements and were conducted in a supervised setting. It is known that subjects tend to change
Falls risk in older people with dementia
their behavior, if they know that they are being observed. Therefore, it would be advantageous, if the sensor-based part of the measurements could take place in an unsupervised setting during the subjects’ everyday live. This may reduce the effect of feeling observed by the study personal.
OBJECTIVE Aim of this study was to make a fall prognosis in a cohort of older people with dementia in short-term [2 month], mid-term [4 month] and long-term [8 month] intervals using accelerometry-based gait analysis. The measurements were performed in the subjects’ everyday life.
METHODS Study design The study was designed as a longitudinal cohort study. This design allows following the subjects over a period of time. The subjects were recruited from a nursing home [Vivantes ‘‘Hauptstadtpflegehaus John F. Kennedy’’, Berlin, Germany] specialized in the care of people with dementia. A nursing home was chosen, because falls were routinely documented as part of their Quality Assurance measures. The inclusion criteria were:
Age 65 years when joining the study, TUG 415 seconds, Mini Mental State Examination (MMSE) 524 points (20), Recurrent falls, Signed written informed consent by the subjects’ legal guardians.
Study participants who were not able to walk independently were excluded. Overall, there were six visits for each study participant. At the first visit, the residents were informed about the study and they were screened for the inand exclusion criteria. Geriatric assessments were performed [MMSE, Barthel index (21), TUG, Tinetti test, Functional Reach and STRATIFY score], and basic body dimensions [size, weight] were collected. At visits 2–5 the accelerometric measurements took place, with a two-month interval between the visits. For convenience, we want to use the term visit also for the accelerometric measurement after the geriatric assessments. The visits 2–5 took one week each. A one week measurement period per subject was chosen, because of the sensor’s limited battery life. At visit 6 a final discussion with the nursing staff and the study participants took place. The nursing staff of the institution was trained prior to the measurements. They were asked to routinely control the correct assignment of sensor and subject. During the measurements of visits 2–5, the nursing staff collected all sensors and sensor pockets in the evening, checked the correct assignment of sensors and subjects, and searched for sensors lost during the day. In the morning, all sensors were attached to the subjects again. The study was approved by the ethics committee of the Charite´, Berlin, Germany and conducted in the context of the GAL project, a German acronym for ‘‘Design of Environments for Ageing’’ (22).
251
252
M. Gietzelt et al.
Gait analysis The accelerometric signals were calibrated, and gait episodes were extracted from these signals using an auto-correlation method (23). This method was evaluated in a former study with gait data of 6 young and healthy adults in a supervised setting. They were asked to walk distances of 20 m and 60 m at self-selected speed with 5 repetitions. All gait episodes were detected by the method and there were no false-positives. The delay of gait detection was 2.7 s on average. The gait episodes in the current study had to feature a length of at least 20 seconds, because shorter gait episodes are known to be unreliable (12). After that, the accelerometric signals were aligned to the human body in order to allow for providing information about a certain body axis (24). Figure 1 shows the definition of human axes. Finally, the following gait parameters were computed (23):
Gait velocity v: computed by the anterior–posterior accelerations aap of the subjects using vðtÞ ¼
Z
T
aap ðtÞdt
ð1Þ
0
whereas T is the duration of the gait episode. Simpson’s rule was used as numerical integration method. ~ avg : the kinetic energy transformed during gait. Average kinetic energy E Thereby, the norm of the signal jjaðtÞjj was used. In order to eliminate the
Figure 1. Definition of human axes: anterior–posterior (ap), left-right (lr), and superiorinferior (si).
Falls risk in older people with dementia
influence of gravity (1 g 9:81 m s2 ), the norm was corrected to jjaðtÞjj 1g. This yields the formula: ~ avg ¼ 1 m~v2 ¼ 1 m E 2 2
Z
2
T
j jjaðtÞjj 1 g jdt
:
ð2Þ
0
Compensation movements compaxis of the acceleration signal a: compaxis ¼
Z
T
jaaxis ðtÞ ^aaxis jdt,
ð3Þ
0
whereas ‘‘axis’’ can be one of ap, lr, si, or the norm of the signal. ^aaxis means the arithmetic average of the acceleration signal regarding ‘‘axis’’. Variance varaxis in all three axes and the norm of the acceleration signal. Step frequency freq: the main frequency of the frequency spectrum derived by the discrete Fourier transformation. Number of dominant peaks numpeaks: number of dominant peaks in the frequency spectrum.
Data analysis For analysis purposes the gait parameters were imported into a MySQL data base and linked up with the fall protocols. In the study protocol it was prospectively decided to use the decision tree induction method C4.5 (25) for analysis, because of the hypothesis that gait parameters have a hierarchical interrelationship regarding falls. Before this, gait parameters were selected regarding their predictive power using the Correlation-based Feature Selection Subset Evaluation method [CfsSubsetEval] (26). This method eliminates redundant parameters, which correlate strongly with other parameters and selects those parameters, which have a correlation with the target variable [falls]. The halting criterion of pruning the tree was a minimum of 20 instances per node, so that no further splitting was done. The classification [risk or safe] of a leaf was done according to the majority of instances. The data analysis was done in a prospective manner, so that occurred falls were only assigned to gait episodes, which were documented earlier. That is, in the context of the visits, it was possible to derive a prognosis for short-term [2 month], mid-term [4 month] and long-term [8 month] fall risk based on the gait parameters. For the long-term prediction only gait episodes from visit 2 could be used. The mid-term prediction used data from visits 2 to predict falls in measurement phases 1 and 2, and from visit 4 to predict falls in phases 3 and 4. For the short-term prediction gait episodes from all visits were used. The evaluation was done using a 10-fold cross-validation. The analysis was instance-based, so that each gait episode was treated and classified separately without any time-related analysis. Hard- and software In this study, the SHIMMER platform [Figure 2] and the integrated tri-axial accelerometer MMA7260QT was used to measure movements (27).
253
254
M. Gietzelt et al.
Figure 2. The Shimmer platform used as accelerometric sensor platform in this study compared to a 2 Euro coin.
The accelerometric signals were recorded on an internal microSD-card for later analysis. Accelerometric signals were calibrated, gait episodes were extracted, and gait parameters were computed using a selfdeveloped Java-framework. WEKA version 3.6.6 served as software for data analysis (28).
RESULTS Subjects The study was conducted during September 2012 and July 2013 with 40 subjects [20 male, 20 female]. The development of the study population over time is shown in Figure 3. The measurement phases in Figure 3 are the two month intervals between the visits. Until the end of the study, 12 drop-outs [30%] were registered. Reasons for drop-outs were moves, mobility-related diseases or deaths. The mean age of the subjects was 76.0 ± 8.3 years [mean ± standard deviation], the mean body height was 168.7 ± 8.6 cm and the mean weight was 73.4 ± 14.1 kg. Altogether 26 falls occurred, whereby 13 subjects have fallen. Eight subjects fell once, two fell twice, one subject experienced four falls and two subjects had five falls. The subjects had severe to most severe cognitive impairments [MMSE 9.3 ± 8.0 points] and showed a general need for help at activities of daily living [Barthel index 53.1 ± 21.9 points]. Furthermore, distinctive limitations in mobility [TUG 37.9 ± 36.2 s; Tinetti test 14.2 ± 8.0 points], and a higher risk of falling [STRATIFY score 3.6 ± 1.4 points] were identified.
Falls risk in older people with dementia t = 0 month 64 persons screened, 40 subjects included in the study
38 subjects
Phase 1, t = 0...2 month 2 drop-outs 6 falls occurred
18 data sets missing
33 subjects
Phase 2, t = 2...4 month 5 drop-outs 11 falls occurred
2 data sets missing
30 subjects
Phase 3, t = 4...6 month 3 drop-outs 8 falls occurred
5 data sets missing
28 subjects
Phase 4, t = 6...8 month 2 drop-outs 1 fall occurred
11 data sets missing
t = 8 month 28 subjects remained
Figure 3. The development of the number of study participants over time.
Table 1. Summary of the AUC of demographic data and geriatric assessment tests regarding a long-term fall prognosis.
Age
Body height
Weight
MMSE
Barthel index
TUG
Tinetti test
Functional Reach
STRATIFY score
0.58
0.53
0.67
0.58
0.55
0.52
0.53
0.55
0.56
Geriatric assessments and fall prognosis The results of the geriatric assessment tests were explored for any predictive ability to distinguish between the subgroups faller and non-faller. Therefore, the Receiver Operating Characteristic [ROC] was used. From this characteristic, the Area Under the Curve [AUC] can be determined, which is a measure of the linear discriminative ability of the geriatric assessments. In case of a perfect discriminative ability between the subgroups faller and non-faller, the AUC becomes 1, in the worst case 0.5. Table 1 summarizes the results of the geriatric assessment tests. Gait parameters and fall prognosis Short-term prognosis The C4.5 algorithm induced a tree with 13 nodes for short-term prognosis. The resulting tree is shown in Figure 4 and can be interpreted as follows: Beginning from the root of the tree, in each node a decision is made. If the decision is positive, then one has to follow the left edge; otherwise the right one. This is done until one reaches a leaf specifying ‘‘RISK’’ or ‘‘SAFE’’, which means that the gait parameters extracted from the current gait episode indicate a short-term risk of falling or not. Before inducing the tree, the CfsSubsetEval algorithm chose the following predictive gait parameters:
compensation movements of the signal’s norm (compNorm ) compensation movements lr-direction (complr ) variance of the signal’s norm (varNorm )
255
256
M. Gietzelt et al.
Figure 4. The decision tree for short-term fall prognosis. The decisions in the nodes identify, which edge will be visited: If the decision is positive, the left edge will be visited; otherwise the right one. The content of the leaves indicate, whether the gait episode showed a risk for short-term falling or not.
Table 2. Summary of the results of the cross-validation for short-term fall prognosis. Short-term fall prognosis cross table and statistical measurements Model prognosis
Reality
Positive Negative
Positive 8084 971
Negative 136 377
CCR Sensitivity Specificity
88.4% 98.3% 28.0%
PPV
89.3%
NPV Cohen’s k (29) AUC
73.5% 0.36 0.76
The abbreviations of the statistical measurements mean the rate of correctly classified instances (CCR), and the positive and negative predictive value (PPV and NPV). Positive and negative refers to fall related and non-fall related cases.
Altogether 9.586 gait episodes were detected for short-term prognosis, which were understood as instances for classification. The 10-fold crossvalidation showed a rate of correctly classified instances [CCR] of 88.4%. Table 2 summarizes the results of the cross-validation.
Falls risk in older people with dementia Table 3. Summary of the results of the cross-validation for mid-term fall prognosis. Mid-term fall prognosis cross table and statistical measurements Model prognosis
Reality
Positive Negative
positive 3513 1181
negative 981 2918
CCR Sensitivity Specificity PPV
74.8% 78.2% 71.2% 74.8%
NPV Cohen’s k AUC
74.8% 0.49 0.80
Mid-term prognosis The analysis of mid-term fall prognosis induced a decision tree with 109 nodes, whereas 8.593 gait episodes were provided for classification. Due to the large size, we have decided against a presentation of the tree in this paper. The following parameters were identified as predictive using the CfsSubsetEval method:
duration of gait episode velocity (v) compensation movements of the signal’s norm (compNorm ) compensation movements si-direction (compSi ) variance of the signal’s norm (varNorm ) variance si-direction (varSi ) variance lr-direction (varlr ) step frequency (freq) Table 3 summarizes the results of the 10-fold cross-validation.
Long-term prognosis For the long-term prognosis the following parameters were chosen:
velocity (v) compensation movements of the signal’s norm (compNorm ) compensation movements lr-direction (complr ) compensation movements si-direction (compSi ) variance of the signal’s norm (varNorm ) variance si-direction (varSi ) step frequency (freq) number of dominant peaks (numpeaks)
3.974 gait episodes were identified for classification. Figure 5 presents the induced decision tree. Table 4 summarizes the results of the 10-fold crossvalidation.
DISCUSSION AND CONCLUSION In this paper we present a system and a field study to assess the risk of falling in people with dementia in their everyday life environment using accelerometry-based gait analysis. The results of the short- and long-term prognosis
257
258
M. Gietzelt et al.
Figure 5. The decision tree for long-term prognosis.
Table 4. Summary of the results of the cross-validation for long-term fall prognosis.. Long-term fall prognosis cross table and statistical measurements Model prognosis
Reality
Positive Negative
Positive 176 109
Negative 348 3341
CCR Sensitivity Specificity PPV
88.5% 33.6% 96.8% 61.8%
NPV Cohen’s k AUC
90.6% 0.38 0.79
showed that very small decision trees could be induced, especially in the focus of the large amount of available instances [gait episodes] to classify. Looking at the CCRs, it seems that these two prognoses have achieved the best results, but this can lead to misinterpretations. Hereby, one has to pay attention and, however, has to consider the other statistical measurements, too. It emerges that the short-term prognosis has a low specificity and the long-term prognosis a low sensitivity, which can lead to both over- and underestimations of fallers or non-fallers. Only the mid-term prognosis shows values for sensitivity, specificity, PPV and NPV, which are nearly at the same level. In addition, Cohen’s k showed the largest value of all three models. One can conclude that the mid-term prognosis shows the best results. That means it is possible to
Falls risk in older people with dementia
classify gait episodes in being associated with a faller or a non-faller for a midterm period with an accuracy of about 75% only using an accelerometry-based gait analysis. It might be possible to improve the prognosis with additional information about age and other fall-related risk factors. The number of gait episodes for short- mid-, and long-term analysis varied. This was due to subjects, who dropped-out of the study, because of moves, mobility-related diseases or deaths. In these cases fall protocols were not available, and thus, the time of e.g. 8 month for long-term monitoring could not be fully covered. Many data sets had to be excluded from analysis, because of assignment problems between sensor and subject. This points out that this study was a field study in an unsupervised setting, and that there was no influence to control the measurement process. This is an important result for future field studies. In this paper, only a decision tree induction method was used for classification, because this was planned prospectively in the study protocol. The choice of the decision tree was based on our hypothesis that gait parameters have a hierarchical interrelationship regarding gait problems and hence to the fall prognosis. That means that a high step frequency associated with small steps may indicate festination, which can typically be observed in patients with Parkinson’s disease. One can also imagine that a gait may be unstable, if there is a low velocity and a low step frequency. The most interesting point with decision trees is that not all parameters have to be computed at once, because only the parameters on the path of the decision tree are needed. This might save computing time on small sensor nodes or smart phones with limited resources. Because of the instance-based approach to classify gait episodes, the information about intra-individual changes of gait could not be covered. This may contain further information and may lead to better results. Another approach should be the individualization of the analysis, so that an intraindividual prognosis for each subject is done instead of the current interindividual classification of gait episodes. This, however, requires a learning period, in which enough fall occurrences are present. Limitations There are a number of limitations to be stated regarding the current study:
The nursing home documents fall events because of reasons of Quality Assurance using fall protocols. Although this documentation is done in a professional manner, it might have happened that some falls remain undiscovered und undocumented. Furthermore, the quality of the fall protocols can vary due to experience and medical expertise of the assessor. The preselection of the subjects [before screening for in- and exclusion criteria] was not random. They were selected by the nursing staff. The criteria of preselection were recurrent falls and the expected cooperativeness of the residents. This makes it clear that this study was a technical feasibility study. The nursing home was not randomly selected, too. The choice of the nursing home in northern Berlin depended on organizational reasons.
259
260
M. Gietzelt et al.
The flows in the everyday work routine may differ between both units of the nursing home. This may had an influence on the observational equivalence. The study design was chosen as field study. Thereby, the impossibility of controlling the setting is a huge problem. It remains unclear, if all detected gait episodes were actually gait episodes and if they could be assigned to the correct subject. Another goal of the study design was to reduce the impression of the subjects of being observed. It is hard to estimate whether this was the case and what influence this had on the participants of the study. Subjects, who are walking more often than others, can bias the result, because they produce more gait episodes. This may lead to an overrepresentation of particular archetypical gait patterns of subjects with increased motor activity.
DECLARATION OF INTEREST The authors report no conflicts of interest. The authors alone are responsible for the content and writing of the paper. The Lower Saxony research network ‘‘Design of Environments for Ageing’’ acknowledges the support of the Lower Saxony Ministry of Science and Culture through the ‘‘Niedersa¨chsisches Vorab’’ grant programme (grant ZN 2701).
REFERENCES 1. von Renteln-Kruse W, Krause T, Dieckmann P, Vogel J. Geriatric patients’ mobility status as reflected by the relevant items of the Barthel index and in-hospital falls. J Am Geriatr Soc 2006;54:1012–13. 2. Scanaill CN, Garattini C, Greene BR, McGrath MJ. Technology innovation enabling falls risk assessment in a community setting. Ageing Int 2011;36:217–31. 3. Podsiadlo D, Richardson S. The timed ‘‘Up & Go’’: a test of basic functional mobility for frail elderly persons. J Am Geriatr Soc 1991;39:142–8. 4. Tinetti ME. Performance-oriented assessment of mobility problems in elderly patients. J Am Geriatr Soc 1986;34:119–26. 5. Duncan PW, Weiner DK, Chandler J, Studenski S. Functional reach: a new clinical measure of balance. J Gerontol 1990;45:M192–7. 6. Oliver D, Britton M, Seed P, et al. Development and evaluation of evidence based risk assessment tool (STRATIFY) to predict which elderly inpatients will fall: casecontrol and cohort studies. BMJ 1997;315:1049–53. 7. Oliver D, Daly F, Martin FC, McMurdo ME. Risk factors and risk assessment tools for falls in hospital in-patients: a systematic review. Age Ageing 2004;33:122–30. 8. Myers H. Hospital fall risk assessment tools: a critique of the literature. Int J Nurs Pract 2003;9:223–35. 9. Merom D, Cumming R, Mathieu E, et al. Can social dancing prevent falls in older adults? A protocol of the Dance, Aging, Cognition, Economics (DAnCE) fall prevention randomised controlled trial. BMC Public Health 2013;13:477. 10. Giansanti D. Investigation of fall-risk using a wearable device with accelerometers and rate gyroscopes. Physiol Meas 2006;27:1081–90. 11. Sˇkrba Z, O’Mullane B, Greene BR, et al. Objective real-time assessment of walking and turning in elderly adults. Proceedings of the 31st Annual International Conference of the IEEE EMBS; 2009 Sep 2–6; Minneapolis, MN, USA; 2009:807–10.
Falls risk in older people with dementia
12. Scanaill CN, Greene BR, Doheny EP, et al. Clinical gait assessment of older adults using open platform tools. Proceedings of the 33rd Annual International Conference of the IEEE EMBS; 2011 Aug 30–Sep 3; Boston, MA, USA; 2011:462–5. 13. Doheny EP, Greene BR, Foran T, et al. Diurnal variations in the outcomes of instrumented gait and quiet standing balance assessments and their association with falls history. Physiol Meas 2012;33:361–73. 14. Narayanan MR, Redmond SJ, Scalzi ME, et al. Longitudinal falls-risk estimation using triaxial accelerometry. IEEE Trans Biomed Eng 2010;57:534–41. 15. Senden R, Savelberg HH, Grimm B, et al. Accelerometry-based gait analysis, an additional objective approach to screen subjects at risk for falling. Gait Posture 2012;36:296–300. 16. Laessoe U, Hoeck HC, Simonsen O, et al. Fall risk in an active elderly population – can it be assessed? J Neg Results Biomed 2007;6:2. doi: 10.1186/1477-5751-6-2. 17. Marschollek M, Rehwald A, Wolf KH, et al. Sensor-based fall risk assessment – an expert ‘to go’. Methods Inf Med 2011;50:420–6. 18. Marschollek M, Rehwald A, Wolf KH, et al. Sensors vs. experts – a performance comparison of sensor-based fall risk assessment vs. conventional assessment in a sample of geriatric patients. BMC Med Inform Decis Mak 2011;11:48. 19. Greene BR, Doheny EP, Walsh C, et al. Evaluation of falls risk in communitydwelling older adults using body-worn sensors. Gerontology 2012;58:472–80. 20. Folstein MF, Folstein SE, McHugh PR. ‘‘Mini-mental state’’. A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 1975;12: 189–8. 21. Mahoney FI, Barthel DW. Functional evaluation: The Barthel index. Md State Med J 1965;14:61–5. 22. Haux R, Hein A, Eichelberg M, et al. The Lower Saxony research network design of environments for ageing: towards interdisciplinary research on information and communication technologies in ageing societies. Inform Health Soc Care 2010;35: 92–103. 23. Gietzelt M, Wolf KH, Kohlmann M, et al. Measurement of accelerometry-based gait parameters in people with and without dementia in the field – a technical feasibility study. Methods Inf Med 2013;52:319–25. 24. Gietzelt M, Schnabel S, Wolf KH, et al. A method to align the coordinate system of accelerometers to the axes of a human body: The depitch algorithm. Comput Methods Programs Biomed 2012;106:97–103. 25. Quinlan JR. C4.5: programs for machine learning. Revised ed. San Mateo (CA): Morgan Kaufmann Publishers; 1993. 26. Hall MA. Correlation-based feature selection for machine learning [dissertation]. Hamilton, New Zealand: University of Waikato; 1999. 27. Burns A, Greene BR, McGrath MJ, et al. SHIMMERTM – a wireless sensor platform for noninvasive biomedical research. IEEE Sensors J 2010;10:1527–34. 28. Hall M, Frank E, Holmes G, et al. The WEKA data mining software: an update. SIGKDD Explorations 2009;11:10–18. 29. Cohen J. A coefficient of agreement for nominal scales. Educ Psychol Meas 1960;20: 37–46.
261
Copyright of Informatics for Health & Social Care is the property of Taylor & Francis Ltd and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.
A prospective field study for sensor-based identification of fall risk in older people with dementia Matthias Gietzelt,1 Florian Feldwieser,2 Mehmet Go ¨ vercin,2 Elisabeth Steinhagen-Thiessen,2 and Michael Marschollek3 1
Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig – Institute of Technology and Hannover Medical School, Braunschweig, Germany, 2 Geriatric Research Group, Department of Geriatric Medicine, Charite´ – Universita¨tsmedizin Berlin, Berlin, Germany, 3 Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig – Institute of Technology and Hannover Medical School, Hannover, Germany Objective: Aim of this study was to make a fall prognosis in a cohort of older people with dementia in short-term (2 month), mid-term (4 month) and long-term (8 month) intervals using accelerometry during the subjects’ everyday life. Methods: The study was designed as a longitudinal cohort study. The subjects were recruited from a nursing home and geriatric assessment tests were conducted at baseline. Each subject underwent four visits and was measured at each visit for one week. Gait episodes were detected and gait parameters were extracted from these episodes. These gait parameters were combined with the falls occurred during the study. A decision tree induction method was used to analyze the data. Results: Forty subjects participated in the study, whereby 12 drop-outs were registered. The geriatric assessment tests were unable to distinguish between the groups (AUC50.6). The evaluation of the models induced with the decision tree classification showed a rate of correctly classified gait episodes of 88.4% for short-term, 74.8% for midterm, and 88.5 % for long-term monitoring. Discussion and conclusions: We concluded that it is possible to classify gait episodes of fallers and non-fallers in people with dementia during everyday life using accelerometry. Keywords Accelerometer, dementia, fall risk, health-enabling technologies
INTRODUCTION Thirty-seven percent of women and 25% of men aged 65 years and older suffer from at least one fall event per year. This percentage increases with age, so that about 80% of all people older than 80 years fall at least once a year (1). The estimated health-related costs of falls and their consequences in people older than 65 years in the US vary between US$3476 and US$10 749 per fall.
Correspondence: Matthias Gietzelt, Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig – Institute of Technology and Hannover Medical School, Mu¨hlenpfordtstr. 23, D-38106 Braunschweig, Germany. Tel: +49 531 391 9505. Fax: +49 531 391 9502. E-mail: [email protected]
250
M. Gietzelt et al.
In Germany, fall related costs are valued at one billion Euros per year (2). Within hospital wards, specific geriatric assessments are typically performed for the estimation of fall risk. The most frequently used tests are the Timed ‘‘Up & Go’’ test [TUG] (3), the Tinetti test (4), the Functional Reach test (5), and the STRATIFY score (6). Common fall risk factors are: abnormal gait, lower limb weakness, presence of confusion or disorientation, altered mental status, agitation, urinary incontinence, and fall history (7). However, the commonly used clinical fall risk assessment tools that are based on these consistently recurrent risk factors, only achieve moderate to poor results in at least one of the statistical parameters sensitivity, specificity, positive and negative predictive value [PPV and NPV] (7,8). This fact leads to a massive over- or underestimation of fallers or non-fallers. Specialized fall prevention programs seem to be more cost-effective than treatment and care after a fall event has been occurred (9). It is still difficult to correctly identify those who are at risk of falling, which is why fall risk assessment is still a subject of research. In the context of cost-effectiveness it is important that prevention programs are only administered to persons that are at risk of falling and not across a broad population. Currently, the trend tends to technical solutions rather than to classic fall risk assessment scores and tests, because of the potentially expected improvements in objectiveness and reliability. A common approach is to use wearable sensors [accelerometers and gyroscopes] (10), because of their small size and low price. Other approaches use camera systems (11), the GAITRite M2 sensor floor [CIR Systems Inc., Sparta, NJ], or the S4 Sensors Walkway [Tactex and S4 Sensors, Victoria, BC, Canada] (12). Related work Most authors tend to use recent fall history of subjects in order to make a fall prognosis (13). A recent fall history might be a good indicator for a higher risk of falling, but is not a robust predictor that a fall will actually happen in the near future. Other authors use a variety of fall risk assessment tools and scores for prognosis (10,14,15). These tools, however, may have a limited capability for fall prognosis. Studies that documented falls prospectively are rare and to the authors’ best knowledge currently only four studies exist, which make a prospective fall prognosis based on falls documented after the sensor-based measurements took place (16–19). These studies used an accelerometry-based gait analysis and their hypothesis was that there are archetypical gait patterns, which are predictive for falls. The subjects were called by phone in a followup after one (16–18) or two years (19), respectively. This might be a methodological flaw, because the subjects may have overlooked a fall happened during that time and there is no possibility to control the information of the subjects. Furthermore, subjects tend to exclude or ignore fall events, which caused no injuries. Therefore, it seems to be necessary to have access to a professional documentation of fall events for these kinds of studies. The four mentioned studies used sensor-based measurements and were conducted in a supervised setting. It is known that subjects tend to change
Falls risk in older people with dementia
their behavior, if they know that they are being observed. Therefore, it would be advantageous, if the sensor-based part of the measurements could take place in an unsupervised setting during the subjects’ everyday live. This may reduce the effect of feeling observed by the study personal.
OBJECTIVE Aim of this study was to make a fall prognosis in a cohort of older people with dementia in short-term [2 month], mid-term [4 month] and long-term [8 month] intervals using accelerometry-based gait analysis. The measurements were performed in the subjects’ everyday life.
METHODS Study design The study was designed as a longitudinal cohort study. This design allows following the subjects over a period of time. The subjects were recruited from a nursing home [Vivantes ‘‘Hauptstadtpflegehaus John F. Kennedy’’, Berlin, Germany] specialized in the care of people with dementia. A nursing home was chosen, because falls were routinely documented as part of their Quality Assurance measures. The inclusion criteria were:
Age 65 years when joining the study, TUG 415 seconds, Mini Mental State Examination (MMSE) 524 points (20), Recurrent falls, Signed written informed consent by the subjects’ legal guardians.
Study participants who were not able to walk independently were excluded. Overall, there were six visits for each study participant. At the first visit, the residents were informed about the study and they were screened for the inand exclusion criteria. Geriatric assessments were performed [MMSE, Barthel index (21), TUG, Tinetti test, Functional Reach and STRATIFY score], and basic body dimensions [size, weight] were collected. At visits 2–5 the accelerometric measurements took place, with a two-month interval between the visits. For convenience, we want to use the term visit also for the accelerometric measurement after the geriatric assessments. The visits 2–5 took one week each. A one week measurement period per subject was chosen, because of the sensor’s limited battery life. At visit 6 a final discussion with the nursing staff and the study participants took place. The nursing staff of the institution was trained prior to the measurements. They were asked to routinely control the correct assignment of sensor and subject. During the measurements of visits 2–5, the nursing staff collected all sensors and sensor pockets in the evening, checked the correct assignment of sensors and subjects, and searched for sensors lost during the day. In the morning, all sensors were attached to the subjects again. The study was approved by the ethics committee of the Charite´, Berlin, Germany and conducted in the context of the GAL project, a German acronym for ‘‘Design of Environments for Ageing’’ (22).
251
252
M. Gietzelt et al.
Gait analysis The accelerometric signals were calibrated, and gait episodes were extracted from these signals using an auto-correlation method (23). This method was evaluated in a former study with gait data of 6 young and healthy adults in a supervised setting. They were asked to walk distances of 20 m and 60 m at self-selected speed with 5 repetitions. All gait episodes were detected by the method and there were no false-positives. The delay of gait detection was 2.7 s on average. The gait episodes in the current study had to feature a length of at least 20 seconds, because shorter gait episodes are known to be unreliable (12). After that, the accelerometric signals were aligned to the human body in order to allow for providing information about a certain body axis (24). Figure 1 shows the definition of human axes. Finally, the following gait parameters were computed (23):
Gait velocity v: computed by the anterior–posterior accelerations aap of the subjects using vðtÞ ¼
Z
T
aap ðtÞdt
ð1Þ
0
whereas T is the duration of the gait episode. Simpson’s rule was used as numerical integration method. ~ avg : the kinetic energy transformed during gait. Average kinetic energy E Thereby, the norm of the signal jjaðtÞjj was used. In order to eliminate the
Figure 1. Definition of human axes: anterior–posterior (ap), left-right (lr), and superiorinferior (si).
Falls risk in older people with dementia
influence of gravity (1 g 9:81 m s2 ), the norm was corrected to jjaðtÞjj 1g. This yields the formula: ~ avg ¼ 1 m~v2 ¼ 1 m E 2 2
Z
2
T
j jjaðtÞjj 1 g jdt
:
ð2Þ
0
Compensation movements compaxis of the acceleration signal a: compaxis ¼
Z
T
jaaxis ðtÞ ^aaxis jdt,
ð3Þ
0
whereas ‘‘axis’’ can be one of ap, lr, si, or the norm of the signal. ^aaxis means the arithmetic average of the acceleration signal regarding ‘‘axis’’. Variance varaxis in all three axes and the norm of the acceleration signal. Step frequency freq: the main frequency of the frequency spectrum derived by the discrete Fourier transformation. Number of dominant peaks numpeaks: number of dominant peaks in the frequency spectrum.
Data analysis For analysis purposes the gait parameters were imported into a MySQL data base and linked up with the fall protocols. In the study protocol it was prospectively decided to use the decision tree induction method C4.5 (25) for analysis, because of the hypothesis that gait parameters have a hierarchical interrelationship regarding falls. Before this, gait parameters were selected regarding their predictive power using the Correlation-based Feature Selection Subset Evaluation method [CfsSubsetEval] (26). This method eliminates redundant parameters, which correlate strongly with other parameters and selects those parameters, which have a correlation with the target variable [falls]. The halting criterion of pruning the tree was a minimum of 20 instances per node, so that no further splitting was done. The classification [risk or safe] of a leaf was done according to the majority of instances. The data analysis was done in a prospective manner, so that occurred falls were only assigned to gait episodes, which were documented earlier. That is, in the context of the visits, it was possible to derive a prognosis for short-term [2 month], mid-term [4 month] and long-term [8 month] fall risk based on the gait parameters. For the long-term prediction only gait episodes from visit 2 could be used. The mid-term prediction used data from visits 2 to predict falls in measurement phases 1 and 2, and from visit 4 to predict falls in phases 3 and 4. For the short-term prediction gait episodes from all visits were used. The evaluation was done using a 10-fold cross-validation. The analysis was instance-based, so that each gait episode was treated and classified separately without any time-related analysis. Hard- and software In this study, the SHIMMER platform [Figure 2] and the integrated tri-axial accelerometer MMA7260QT was used to measure movements (27).
253
254
M. Gietzelt et al.
Figure 2. The Shimmer platform used as accelerometric sensor platform in this study compared to a 2 Euro coin.
The accelerometric signals were recorded on an internal microSD-card for later analysis. Accelerometric signals were calibrated, gait episodes were extracted, and gait parameters were computed using a selfdeveloped Java-framework. WEKA version 3.6.6 served as software for data analysis (28).
RESULTS Subjects The study was conducted during September 2012 and July 2013 with 40 subjects [20 male, 20 female]. The development of the study population over time is shown in Figure 3. The measurement phases in Figure 3 are the two month intervals between the visits. Until the end of the study, 12 drop-outs [30%] were registered. Reasons for drop-outs were moves, mobility-related diseases or deaths. The mean age of the subjects was 76.0 ± 8.3 years [mean ± standard deviation], the mean body height was 168.7 ± 8.6 cm and the mean weight was 73.4 ± 14.1 kg. Altogether 26 falls occurred, whereby 13 subjects have fallen. Eight subjects fell once, two fell twice, one subject experienced four falls and two subjects had five falls. The subjects had severe to most severe cognitive impairments [MMSE 9.3 ± 8.0 points] and showed a general need for help at activities of daily living [Barthel index 53.1 ± 21.9 points]. Furthermore, distinctive limitations in mobility [TUG 37.9 ± 36.2 s; Tinetti test 14.2 ± 8.0 points], and a higher risk of falling [STRATIFY score 3.6 ± 1.4 points] were identified.
Falls risk in older people with dementia t = 0 month 64 persons screened, 40 subjects included in the study
38 subjects
Phase 1, t = 0...2 month 2 drop-outs 6 falls occurred
18 data sets missing
33 subjects
Phase 2, t = 2...4 month 5 drop-outs 11 falls occurred
2 data sets missing
30 subjects
Phase 3, t = 4...6 month 3 drop-outs 8 falls occurred
5 data sets missing
28 subjects
Phase 4, t = 6...8 month 2 drop-outs 1 fall occurred
11 data sets missing
t = 8 month 28 subjects remained
Figure 3. The development of the number of study participants over time.
Table 1. Summary of the AUC of demographic data and geriatric assessment tests regarding a long-term fall prognosis.
Age
Body height
Weight
MMSE
Barthel index
TUG
Tinetti test
Functional Reach
STRATIFY score
0.58
0.53
0.67
0.58
0.55
0.52
0.53
0.55
0.56
Geriatric assessments and fall prognosis The results of the geriatric assessment tests were explored for any predictive ability to distinguish between the subgroups faller and non-faller. Therefore, the Receiver Operating Characteristic [ROC] was used. From this characteristic, the Area Under the Curve [AUC] can be determined, which is a measure of the linear discriminative ability of the geriatric assessments. In case of a perfect discriminative ability between the subgroups faller and non-faller, the AUC becomes 1, in the worst case 0.5. Table 1 summarizes the results of the geriatric assessment tests. Gait parameters and fall prognosis Short-term prognosis The C4.5 algorithm induced a tree with 13 nodes for short-term prognosis. The resulting tree is shown in Figure 4 and can be interpreted as follows: Beginning from the root of the tree, in each node a decision is made. If the decision is positive, then one has to follow the left edge; otherwise the right one. This is done until one reaches a leaf specifying ‘‘RISK’’ or ‘‘SAFE’’, which means that the gait parameters extracted from the current gait episode indicate a short-term risk of falling or not. Before inducing the tree, the CfsSubsetEval algorithm chose the following predictive gait parameters:
compensation movements of the signal’s norm (compNorm ) compensation movements lr-direction (complr ) variance of the signal’s norm (varNorm )
255
256
M. Gietzelt et al.
Figure 4. The decision tree for short-term fall prognosis. The decisions in the nodes identify, which edge will be visited: If the decision is positive, the left edge will be visited; otherwise the right one. The content of the leaves indicate, whether the gait episode showed a risk for short-term falling or not.
Table 2. Summary of the results of the cross-validation for short-term fall prognosis. Short-term fall prognosis cross table and statistical measurements Model prognosis
Reality
Positive Negative
Positive 8084 971
Negative 136 377
CCR Sensitivity Specificity
88.4% 98.3% 28.0%
PPV
89.3%
NPV Cohen’s k (29) AUC
73.5% 0.36 0.76
The abbreviations of the statistical measurements mean the rate of correctly classified instances (CCR), and the positive and negative predictive value (PPV and NPV). Positive and negative refers to fall related and non-fall related cases.
Altogether 9.586 gait episodes were detected for short-term prognosis, which were understood as instances for classification. The 10-fold crossvalidation showed a rate of correctly classified instances [CCR] of 88.4%. Table 2 summarizes the results of the cross-validation.
Falls risk in older people with dementia Table 3. Summary of the results of the cross-validation for mid-term fall prognosis. Mid-term fall prognosis cross table and statistical measurements Model prognosis
Reality
Positive Negative
positive 3513 1181
negative 981 2918
CCR Sensitivity Specificity PPV
74.8% 78.2% 71.2% 74.8%
NPV Cohen’s k AUC
74.8% 0.49 0.80
Mid-term prognosis The analysis of mid-term fall prognosis induced a decision tree with 109 nodes, whereas 8.593 gait episodes were provided for classification. Due to the large size, we have decided against a presentation of the tree in this paper. The following parameters were identified as predictive using the CfsSubsetEval method:
duration of gait episode velocity (v) compensation movements of the signal’s norm (compNorm ) compensation movements si-direction (compSi ) variance of the signal’s norm (varNorm ) variance si-direction (varSi ) variance lr-direction (varlr ) step frequency (freq) Table 3 summarizes the results of the 10-fold cross-validation.
Long-term prognosis For the long-term prognosis the following parameters were chosen:
velocity (v) compensation movements of the signal’s norm (compNorm ) compensation movements lr-direction (complr ) compensation movements si-direction (compSi ) variance of the signal’s norm (varNorm ) variance si-direction (varSi ) step frequency (freq) number of dominant peaks (numpeaks)
3.974 gait episodes were identified for classification. Figure 5 presents the induced decision tree. Table 4 summarizes the results of the 10-fold crossvalidation.
DISCUSSION AND CONCLUSION In this paper we present a system and a field study to assess the risk of falling in people with dementia in their everyday life environment using accelerometry-based gait analysis. The results of the short- and long-term prognosis
257
258
M. Gietzelt et al.
Figure 5. The decision tree for long-term prognosis.
Table 4. Summary of the results of the cross-validation for long-term fall prognosis.. Long-term fall prognosis cross table and statistical measurements Model prognosis
Reality
Positive Negative
Positive 176 109
Negative 348 3341
CCR Sensitivity Specificity PPV
88.5% 33.6% 96.8% 61.8%
NPV Cohen’s k AUC
90.6% 0.38 0.79
showed that very small decision trees could be induced, especially in the focus of the large amount of available instances [gait episodes] to classify. Looking at the CCRs, it seems that these two prognoses have achieved the best results, but this can lead to misinterpretations. Hereby, one has to pay attention and, however, has to consider the other statistical measurements, too. It emerges that the short-term prognosis has a low specificity and the long-term prognosis a low sensitivity, which can lead to both over- and underestimations of fallers or non-fallers. Only the mid-term prognosis shows values for sensitivity, specificity, PPV and NPV, which are nearly at the same level. In addition, Cohen’s k showed the largest value of all three models. One can conclude that the mid-term prognosis shows the best results. That means it is possible to
Falls risk in older people with dementia
classify gait episodes in being associated with a faller or a non-faller for a midterm period with an accuracy of about 75% only using an accelerometry-based gait analysis. It might be possible to improve the prognosis with additional information about age and other fall-related risk factors. The number of gait episodes for short- mid-, and long-term analysis varied. This was due to subjects, who dropped-out of the study, because of moves, mobility-related diseases or deaths. In these cases fall protocols were not available, and thus, the time of e.g. 8 month for long-term monitoring could not be fully covered. Many data sets had to be excluded from analysis, because of assignment problems between sensor and subject. This points out that this study was a field study in an unsupervised setting, and that there was no influence to control the measurement process. This is an important result for future field studies. In this paper, only a decision tree induction method was used for classification, because this was planned prospectively in the study protocol. The choice of the decision tree was based on our hypothesis that gait parameters have a hierarchical interrelationship regarding gait problems and hence to the fall prognosis. That means that a high step frequency associated with small steps may indicate festination, which can typically be observed in patients with Parkinson’s disease. One can also imagine that a gait may be unstable, if there is a low velocity and a low step frequency. The most interesting point with decision trees is that not all parameters have to be computed at once, because only the parameters on the path of the decision tree are needed. This might save computing time on small sensor nodes or smart phones with limited resources. Because of the instance-based approach to classify gait episodes, the information about intra-individual changes of gait could not be covered. This may contain further information and may lead to better results. Another approach should be the individualization of the analysis, so that an intraindividual prognosis for each subject is done instead of the current interindividual classification of gait episodes. This, however, requires a learning period, in which enough fall occurrences are present. Limitations There are a number of limitations to be stated regarding the current study:
The nursing home documents fall events because of reasons of Quality Assurance using fall protocols. Although this documentation is done in a professional manner, it might have happened that some falls remain undiscovered und undocumented. Furthermore, the quality of the fall protocols can vary due to experience and medical expertise of the assessor. The preselection of the subjects [before screening for in- and exclusion criteria] was not random. They were selected by the nursing staff. The criteria of preselection were recurrent falls and the expected cooperativeness of the residents. This makes it clear that this study was a technical feasibility study. The nursing home was not randomly selected, too. The choice of the nursing home in northern Berlin depended on organizational reasons.
259
260
M. Gietzelt et al.
The flows in the everyday work routine may differ between both units of the nursing home. This may had an influence on the observational equivalence. The study design was chosen as field study. Thereby, the impossibility of controlling the setting is a huge problem. It remains unclear, if all detected gait episodes were actually gait episodes and if they could be assigned to the correct subject. Another goal of the study design was to reduce the impression of the subjects of being observed. It is hard to estimate whether this was the case and what influence this had on the participants of the study. Subjects, who are walking more often than others, can bias the result, because they produce more gait episodes. This may lead to an overrepresentation of particular archetypical gait patterns of subjects with increased motor activity.
DECLARATION OF INTEREST The authors report no conflicts of interest. The authors alone are responsible for the content and writing of the paper. The Lower Saxony research network ‘‘Design of Environments for Ageing’’ acknowledges the support of the Lower Saxony Ministry of Science and Culture through the ‘‘Niedersa¨chsisches Vorab’’ grant programme (grant ZN 2701).
REFERENCES 1. von Renteln-Kruse W, Krause T, Dieckmann P, Vogel J. Geriatric patients’ mobility status as reflected by the relevant items of the Barthel index and in-hospital falls. J Am Geriatr Soc 2006;54:1012–13. 2. Scanaill CN, Garattini C, Greene BR, McGrath MJ. Technology innovation enabling falls risk assessment in a community setting. Ageing Int 2011;36:217–31. 3. Podsiadlo D, Richardson S. The timed ‘‘Up & Go’’: a test of basic functional mobility for frail elderly persons. J Am Geriatr Soc 1991;39:142–8. 4. Tinetti ME. Performance-oriented assessment of mobility problems in elderly patients. J Am Geriatr Soc 1986;34:119–26. 5. Duncan PW, Weiner DK, Chandler J, Studenski S. Functional reach: a new clinical measure of balance. J Gerontol 1990;45:M192–7. 6. Oliver D, Britton M, Seed P, et al. Development and evaluation of evidence based risk assessment tool (STRATIFY) to predict which elderly inpatients will fall: casecontrol and cohort studies. BMJ 1997;315:1049–53. 7. Oliver D, Daly F, Martin FC, McMurdo ME. Risk factors and risk assessment tools for falls in hospital in-patients: a systematic review. Age Ageing 2004;33:122–30. 8. Myers H. Hospital fall risk assessment tools: a critique of the literature. Int J Nurs Pract 2003;9:223–35. 9. Merom D, Cumming R, Mathieu E, et al. Can social dancing prevent falls in older adults? A protocol of the Dance, Aging, Cognition, Economics (DAnCE) fall prevention randomised controlled trial. BMC Public Health 2013;13:477. 10. Giansanti D. Investigation of fall-risk using a wearable device with accelerometers and rate gyroscopes. Physiol Meas 2006;27:1081–90. 11. Sˇkrba Z, O’Mullane B, Greene BR, et al. Objective real-time assessment of walking and turning in elderly adults. Proceedings of the 31st Annual International Conference of the IEEE EMBS; 2009 Sep 2–6; Minneapolis, MN, USA; 2009:807–10.
Falls risk in older people with dementia
12. Scanaill CN, Greene BR, Doheny EP, et al. Clinical gait assessment of older adults using open platform tools. Proceedings of the 33rd Annual International Conference of the IEEE EMBS; 2011 Aug 30–Sep 3; Boston, MA, USA; 2011:462–5. 13. Doheny EP, Greene BR, Foran T, et al. Diurnal variations in the outcomes of instrumented gait and quiet standing balance assessments and their association with falls history. Physiol Meas 2012;33:361–73. 14. Narayanan MR, Redmond SJ, Scalzi ME, et al. Longitudinal falls-risk estimation using triaxial accelerometry. IEEE Trans Biomed Eng 2010;57:534–41. 15. Senden R, Savelberg HH, Grimm B, et al. Accelerometry-based gait analysis, an additional objective approach to screen subjects at risk for falling. Gait Posture 2012;36:296–300. 16. Laessoe U, Hoeck HC, Simonsen O, et al. Fall risk in an active elderly population – can it be assessed? J Neg Results Biomed 2007;6:2. doi: 10.1186/1477-5751-6-2. 17. Marschollek M, Rehwald A, Wolf KH, et al. Sensor-based fall risk assessment – an expert ‘to go’. Methods Inf Med 2011;50:420–6. 18. Marschollek M, Rehwald A, Wolf KH, et al. Sensors vs. experts – a performance comparison of sensor-based fall risk assessment vs. conventional assessment in a sample of geriatric patients. BMC Med Inform Decis Mak 2011;11:48. 19. Greene BR, Doheny EP, Walsh C, et al. Evaluation of falls risk in communitydwelling older adults using body-worn sensors. Gerontology 2012;58:472–80. 20. Folstein MF, Folstein SE, McHugh PR. ‘‘Mini-mental state’’. A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 1975;12: 189–8. 21. Mahoney FI, Barthel DW. Functional evaluation: The Barthel index. Md State Med J 1965;14:61–5. 22. Haux R, Hein A, Eichelberg M, et al. The Lower Saxony research network design of environments for ageing: towards interdisciplinary research on information and communication technologies in ageing societies. Inform Health Soc Care 2010;35: 92–103. 23. Gietzelt M, Wolf KH, Kohlmann M, et al. Measurement of accelerometry-based gait parameters in people with and without dementia in the field – a technical feasibility study. Methods Inf Med 2013;52:319–25. 24. Gietzelt M, Schnabel S, Wolf KH, et al. A method to align the coordinate system of accelerometers to the axes of a human body: The depitch algorithm. Comput Methods Programs Biomed 2012;106:97–103. 25. Quinlan JR. C4.5: programs for machine learning. Revised ed. San Mateo (CA): Morgan Kaufmann Publishers; 1993. 26. Hall MA. Correlation-based feature selection for machine learning [dissertation]. Hamilton, New Zealand: University of Waikato; 1999. 27. Burns A, Greene BR, McGrath MJ, et al. SHIMMERTM – a wireless sensor platform for noninvasive biomedical research. IEEE Sensors J 2010;10:1527–34. 28. Hall M, Frank E, Holmes G, et al. The WEKA data mining software: an update. SIGKDD Explorations 2009;11:10–18. 29. Cohen J. A coefficient of agreement for nominal scales. Educ Psychol Meas 1960;20: 37–46.
261
Copyright of Informatics for Health & Social Care is the property of Taylor & Francis Ltd and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use.
