Accelerometer-based wireless body area network to estimate intensity of therapy in post-acute rehabilitation

Abstract

Background: It has been suggested that there is a dose-response relationship between the amount of therapy and functional recovery in post-acute rehabilitation care. To this day, only the total time of therapy has been investigated as a potential determinant of this dose-response relationship because of methodological and measurement challenges. The primary objective of this study was to compare time and motion measures during real life physical therapy with estimates of active time (i.e. the time during which a patient is active physically) obtained with a wireless body area network (WBAN) of 3D accelerometer modules positioned at the hip, wrist and ankle. The secondary objective was to assess the differences in estimates of active time when using a single accelerometer module positioned at the hip.

Methods: Five patients (77.4 +/- 5.2 y) with 4 different admission diagnoses (stroke, lower limb fracture, amputation and immobilization syndrome) were recruited in a post-acute rehabilitation center and observed during their physical therapy sessions throughout their stay. Active time was recorded by a trained observer using a continuous time and motion analysis program running on a Tablet-PC. Two WBAN configurations were used: 1) three accelerometer modules located at the hip, wrist and ankle (M3) and 2) one accelerometer located at the hip (M1). Acceleration signals from the WBANs were synchronized with the observations. Estimates of active time were computed based on the temporal density of the acceleration signals.

Results: A total of 62 physical therapy sessions were observed. Strong associations were found between WBANs estimates of active time and time and motion measures of active time. For the combined sessions, the intraclass correlation coefficient (ICC) was 0.93 (P < or = 0.001) for M3 and 0.79 (P < or = 0.001) for M1. The mean percentage of differences between observation measures and estimates from the WBAN of active time was -8.7% +/- 2.0% using data from M3 and -16.4% +/- 10.4% using data from M1.

Conclusion: WBANs estimates of active time compare favorably with results from observation-based time and motion measures. While the investigation on the association between active time and outcomes of rehabilitation needs to be studied in a larger scale study, the use of an accelerometer-based WBAN to measure active time is a promising approach that offers a better overall precision than methods relying on work sampling. Depending on the accuracy needed, the use of a single accelerometer module positioned on the hip may still be an interesting alternative to using multiple modules.

Figure 1

Time and motion observations and recording of body accelerations. The WBAN used in this study was comprised of three 3D accelerometers modules. Signals recorded by accelerometers were transmitted to a receiver located on the Tablet-PC. The Tablet-PC recorded WBAN's data in background, while an observer noted time and motion parameters of the session. All data was synchronized on a common timeline.

Figure 2

Estimation of active time with accelerometers signals. The three steps of signal transformation are presented in A: 1-Rectified signal, 2-Binary signal and 3-Temporal density. In B, the rectified signal is transformed in a binary signal: all samples above 0.015 Volts (dotted line) are given a value of "1", while samples below equal zero. In C, temporal density is obtained by filtering binary signal with a rolling window of 10 sec. Then, all samples above 0.5 (dotted line) is cumulated to give the active time estimate.

Figure 3

Measure and estimates of active time of therapy sessions throughout the length of stay for each subject.

Figure 4

Association between estimates of active time and measure of active time for observed sessions. Intraclass correlation coefficient between accelerometers' estimates and measurement of active time are presented in the lower right corner of each scatter plot. 95% Confidence interval of ICC was 0.89 to 0.96 for M3 and 0.68 to 0.87 for M1.

Figure 5

Bland-Altman plots of measure and estimate of active time for observed sessions. M3 and M1 are compared to time and motion (TM) analysis. On the Y-axis, differences between methods are expressed as: [(M-TM)/((TM+M)/2)*100]. On the X-axis, averaged active time is calculated as: [(M+TM)/2].