Abstract
Besides eliminating pain, restoring activity is a major goal in orthopaedic interventions including joint replacement or trauma surgery following falls in frail elderly, both treatments of highest socio-economic impact.
In joint replacement and even more so in frail elderly at risk of falling, turns are assessed in clinical tests such as the TUG (Timed Get-up-and-Go), Tinetti, or SPPB so that classifying turning movements in the free field with wearable activity monitors promises clinically valuable objective diagnostic or outcome parameters.
It is the aim of this study to validate a computationally simple turn detection algorithm for a leg-worn activity monitor comprising 3D gyroscopes.
A previously developed and validated activity classification algorithm for thigh-worn accelerometers was extended by adding a turn detection algorithm to its decision tree structure and using the 3D gyroscope of a new 9-axis IMU (56×40×15mm, 25g, f=50Hz,).
Based on published principles (El-Gohary et al. Sensors 2014), the turn detection algorithm filters the x-axis (thigh) for noise and walking (Butterworth low-pass, 2ndorder with a cut-off at 4Hz and 4thorder with a cut-off at 0.3Hz) before using a rotational speed threshold of 15deg/s to identify a turn and taking the bi-lateral zero-crossings as start and stop markers to integrate the turning angle.
For validation, a test subject wore an activity monitor on both thighs and performed a total of 57 turns of various types (walking, on-the-spot, fast/slow), ranges (45 to 360deg) and directions (left/right) in free order while being video-taped. An independent observer annotated the video so that the algorithmic counts could be compared to n=114 turns. Video-observation was compared to the algorithmic classification in a confusion matrix and the detection accuracy (true positives) was calculated.
In addition, 4-day continuous activity measures from 4 test subjects (2 healthy, 2 frail elderly) were compared.
Overall, only 5/114 turns were undetected producing a 96% detection accuracy. No false positives were classified. However, when detection accuracy was calculated for turning angle intervals (45°: 30–67.5°; 90°: 67.5–135°; 180°: 135–270°; 360°: 270–450°), accuracy for all interval classifications combined dropped to 83.3% with equal values for left and right turns. For the 180° and 360°, accuracy was 100% while for the shorter 45° and 90° turns accuracy was 75% and 71% only, mainly because subsequent turns were not separated.
Healthy subjects performed between 470 (office worker) and 823 (house wife) turns/day while frail elderly scored 128 (high fall risk) to 487 turns/day (low fall risk). Turns/day and steps/day were not correlated. In healthy subjects ca. 50% of turns were in the 45° category compared to only ca. 35% in frail elderly.
Turn detection for a thigh-worn IMU activity monitor using a computationally simple algorithm is feasible with high general detection accuracy. The classification and separation of subsequent short turns can be further improved.
In multi-day measurement, turns/day and the distribution of short and long turns seem to be a largely independent activity parameter compared to step counts and may improve objective assessment of fall risk or arthroplasty outcome.
