Autonomous identification of freezing of gait in Parkinson's disease from lower-body segmental accelerometry

Abstract

Background: We have previously published a technique for objective assessment of freezing of gait (FOG) in Parkinson's disease (PD) from a single shank-mounted accelerometer. Here we extend this approach to evaluate the optimal configuration of sensor placement and signal processing parameters using seven sensors attached to the lumbar back, thighs, shanks and feet.

Methods: Multi-segmental acceleration data was obtained from 25 PD patients performing 134 timed up and go tasks, and clinical assessment of FOG was performed by two experienced raters from video. Four metrics were used to compare objective and clinical measures; the intraclass correlation coefficient (ICC) for number of FOG episodes and the percent time frozen per trial; and the sensitivity and specificity of FOG detection.

Results: The seven-sensor configuration was the most robust, scoring highly on all measures of performance (ICC number of FOG 0.75; ICC percent time frozen 0.80; sensitivity 84.3%; specificity 78.4%). A simpler single-shank sensor approach provided similar ICC values and exhibited a high sensitivity to FOG events, but specificity was lower at 66.7%. Recordings from the lumbar sensor offered only moderate agreement with the clinical raters in terms of absolute number and duration of FOG events (likely due to musculoskeletal attenuation of lower-limb 'trembling' during FOG), but demonstrated a high sensitivity (86.2%) and specificity (82.4%) when considered as a binary test for the presence/absence of FOG within a single trial.

Conclusions: The seven-sensor approach was the most accurate method for quantifying FOG, and is best suited to demanding research applications. A single shank sensor provided measures comparable to the seven-sensor approach but is relatively straightforward in execution, facilitating clinical use. A single lumbar sensor may provide a simple means of objective FOG detection given the ubiquitous nature of accelerometers in mobile telephones and other belt-worn devices.

Figure 1

Screen image of the FOG tagging software developed by the investigators. The video image shows a PD patient performing the timed up-and-go (TUG) task, in which subjects start from a seated position and walk 5 m to a target box marked on the floor with tape, execute a turn, and return to the seated position. Clinicians reviewed the video and indicated a FOG event by holding down the 'T' key, resulting in a FOG tag (red rectangle) which could be resized by dragging the ends horizontally in time, which moved the video by a corresponding amount. The video cursor could also be dragged forwards and backwards in time to facilitate repeat viewing of FOG events.

Figure 2

A Vertical acceleration data from the right shank during a TUG task of 43-s duration. The grey rectangle indicates FOG as determined by the clinical raters. B At each point in time acceleration data was sampled from a window (black rectangle - set to a width of 5 s in this instance) and a power spectrum calculated (inset). The ratio of power (area under the curve) in the freeze (3-8 Hz) and locomotor (0-3 Hz) bands from these spectra was determined to form an index of FOG (iFOG). C The iFOG waveform (shown here truncated at 10) was used to identify freezing using a threshold (set to 3 in this example); any periods in which the iFOG waveform was above threshold were tagged as a FOG event. D Vertical acceleration from the lumbar sensor was attenuated relative to the shank but still exhibited similar high frequency characteristics, enabling FOG detection E.

Figure 3

Visual representation of the dataset generated by a single TUG trial. Each of the seven sensors generated an acceleration trace; each trace was processed with sampling windows of 2.5, 5, 7.5 and 10 s to generate four iFOG waveforms (the ratio of freeze to locomotor band power); fourteen binary traces from each of the four iFOG waveforms were then formed using an iFOG threshold of 0.5 to 7 in steps of 0.5, resulting in a total of 56 binary FOG traces per sensor, and 392 per TUG trial. Each binary waveform provided a measure of number of FOG events and the percent time frozen.

Figure 4

Agreement between clinical assessment and accelerometry (as measured by the intraclass correlation coefficient, ICC) for number of FOG events (A-D) and percent time frozen (E-H), as a function of sensor location, sampling window width, and threshold.

Figure 5

Sensitivity (A-D) and specificity (E-H) of accelerometer-based detection of FOG, as a function of sensor location, sampling window width, and threshold.