Detection of gait and postures using a miniaturised triaxial accelerometer-based system: Accuracy in community-dwelling older adults

SIR—Ageing is associated with a decrease in physical functioning and mobility [1, 2]. The performance of basic activities of daily living can become progressively difficult [2, 3]. Furthermore, the prevalence of illness increases with advancing age, which may also cause functional limitations. Maintaining physical activity in later life is related to a slower rate of motor decline, less disability and a lower incidence of specific chronic diseases [4–7]. Therefore, accurate knowledge of daily physical activity patterns is essential for assessment and intervention purposes. Activity monitors based on body-fixed accelerometers and gyroscopes are available for continuous measurements of mobility-related activities [8], but few studies demonstrate evidence-based clinical applications in older people [9]. Often, systems consist of multiple sensors attached to different body parts [10–18]. More sensors provide more information, but minimal instrumentation is preferable for unobtrusive and long-term monitoring in real life. Recently, the DynaPort MoveMonitor was introduced (McRoberts BV, The Hague, The Netherlands), a method to monitor physical activity with a high time resolution using a single and small wireless triaxial accelerometer, the MiniMod. The aim of the present study was to determine the accuracy of the DynaPort system for detecting gait (walking, shuffling) and postures (lying, sitting, standing) in community-dwelling older adults during activity sequences in a standardised setting (Study 1) as well as in a home environment (Study 2). The results were compared to video observation of the actual performance. Twenty older adults (mean age = 68.5 years, SD = 7.4) participated in Study 1. First, a fixed activity sequence including walking trajectories (1.4, 2.3, 4.5 m, taking a three-step stair) and postures (sitting, standing, lying) were performed five times in a laboratory set-up. Thereafter, subjects were allowed to move freely for 3 min with the only instruction that taking the stairs, sitting and lying had to be completed at least once. For Study 2, five older adults were monitored 30 min in their home environment while doing similar activities as in Study 1 and usual domestic activities. During all measurements, subjects were video-recorded while wearing the MiniMod (84 × 50 × 8 mm; 44.5 g) at their lower back. In the video analysis, the start and end of each activity was scored by an observer. Inter-rater reliability of this procedure was determined for the fixed sequence task by two independent raters in 10 subjects. Intraclass correlation coefficients (ICC) for the duration of walking, sitting, standing and lying were respectively 0.95, 0.78, 0.99 and 0.98. The category shuffling was added to the video analysis of the home data because more situations were expected where subjects would not walk, but also would not completely stand still. Results of blinded analysis of the acceleration data were compared to the video observation scores (‘criterion measure’) with a time resolution of 0.1 s. The percentage correspondence and non-correspondence between the two methods have been calculated for all activity data (i.e. over all subjects). Subsequently, sensitivity, specificity, positive predictive value and agreement were calculated per subject. Appendix 1 (see Supplementary data available in Age and Ageing online) provides a detailed description of the methods. Based on the video analysis, a total of 3,368 activities with a total duration of 312.3 min were assessed. Mean activity durations were shorter in the laboratory (fixed and free) than in the home environment, i.e. for lying (12.4 and 11.9 vs 24.2 s), sitting (2.3 and 4.1 vs 29.9 s), standing (3.0 and 2.7 vs 11.5 s) and walking (3.6 and 6.5 vs 7.8 s). Mean duration of shuffling (only at home) was 4.0 s. Table 1 presents the percentages of overall correspondence and non-correspondence between the output of the DynaPort MoveMonitor and the video analysis. Mean sensitivity, specificity and positive predictive values are given in Table 2. In Study 1, agreement per subject ranged between 74.4 and 92.2% (mean = 81.4%, SD = 4.2) for the fixed sequence and 65.1–98.9% (mean = 88.7%, SD = 7.7) for the free sequence. In Study 2, agreement per subject ranged between 68.3 and 85.9% (mean = 79.8%; SD = 6.9). Appendix 2 (see Supplementary data available in Age and Ageing online) provides additional information on the results. Table 1 presents the percentages of overall correspondence and non-correspondence between the output of the DynaPort MoveMonitor and the video analysis. Mean sensitivity, specificity and positive predictive values are given in Table 2. In Study 1, agreement per subject ranged between 74.4 and 92.2% (mean = 81.4%, SD = 4.2) for the fixed sequence and 65.1–98.9% (mean = 88.7%, SD = 7.7) for the free sequence. In Study 2, agreement per subject ranged between 68.3 and 85.9% (mean = 79.8%; SD = 6.9). Appendix 2 (see Supplementary data available in Age and Ageing online) provides additional information on the results. Overall correspondence (bold) and non-correspondence (plain) between the video observation (rows) and DynaPort (columns) per activity category Values are expressed in percentages. Overall correspondence (bold) and non-correspondence (plain) between the video observation (rows) and DynaPort (columns) per activity category Values are expressed in percentages. Mean sensitivity, specificity and positive predictive values (%) Mean sensitivity, specificity and positive predictive values (%) The DynaPort MoveMonitor detected lying almost faultlessly in all circumstances. A high overall correspondence was found for walking too, particularly when subjects walked longer trajectories. A lower sensitivity (86.6%) was reached for the fixed protocol, in which 1.4-m walks were often classified as shuffling. Also, taking the three-step stair was several times classified as shuffling when subjects stepped slowly or paused at the stair. Whether this should have been considered as walking is debatable. However, at home, shuffling was not accurately detected and difficult to evaluate by video observation as well. For example, when sweeping, a subject repeatedly alternated standing and making a few steps. In these and similar situations, shuffling cannot always be clearly discriminated from standing or walking. Moderate results were found for monitoring sitting and standing. In the laboratory, the mean duration of these postures was 2–4 s. Such short-lasting activities are difficult to detect. Quite a lot of sitting periods, predominantly on the edge of the bed, were not detected. This explains the misclassification of sitting as standing or lying, which were the preceding or following activities. Also, the ICC for sitting was lower than for other activity categories due to inter-observer differences in sitting periods on the edge of the bed. Hence, short durations and observation disagreements may have influenced the correspondence figures. However, missing short sitting periods, e.g. as part of getting out of bed, will have a minor effect on overall performance during measurements of days or weeks. Short standing periods were detected well, but frequent small differences with the video observation led to a lower overall correspondence. The mean sensitivity for sitting and standing improved to ∼80% in the home environment, where the mean duration was up to four times longer. Higher sensitivities have been reported but were most often reached by using more than one sensor [12, 13, 15–22]. Furthermore, to our knowledge, validations of other methods did not include activities <5 s and also apply a 0.1 s time resolution. Our study of detecting gait and postures was based on a precise analysis of a large number of fairly short activities. Thus, this evaluation was quite challenging and unforgiving, above all during the fixed sequence task. The latter is illustrated by the fact that in conditions where participants could perform subsequent activities in a slower pace (i.e. in the free sequence and at home), detection was generally better. Except for shuffling, the DynaPort MoveMonitor reached a mean sensitivity of ≥80% in the home environment. Results may further improve during prolonged monitoring when older adults are expected to have longer periods of sedentary activity. It is concluded that the DynaPort MoveMonitor is a practical and valuable tool for objective, continuous assessment of walking and postures. In the present study, the time spent walking and lying was accurately detected. Monitoring of short periods of sitting and standing requires further fine-tuning. Since outcomes may be influenced by the specific evaluation procedures, it is suggested that some standardisation is desirable to better compare future validations of activity monitoring methods. Objective methods are needed for accurate assessment of daily physical activity patterns. Activity monitors based on a single sensor provide practical solutions for long-term unobtrusive monitoring of mobility-related activities in real life. The DynaPort MoveMonitor is a practical and valuable method for objectively evaluating gait and postures. The authors would like to thank Nienke Bruins, Marjolein den Ouden, Vincent Hendriks, Henry van den Crommert and McRoberts BV. This study has been financially supported by a grant of the Innovation Fund of the University Medical Center Groningen and the European Commission (FP6 project SENSACTION-AAL, INFSO-IST-045622). There are no conflicts of interest. Supplementary data mentioned in the text is available to subscribers in Age and Ageing online.