Postural Sway Classification Using Triaxial Accelerometer Signals

Elderly individuals and patients with neurodegenerative diseases often face problems with balance, making them susceptible to falls. Monitoring postural sway is crucial to prevent falls and reduce mortality and morbidity rates. Wearable sensors such as triaxial accelerometers can be used for continuous monitoring. The acceleration measurements (
$XYZ$
) and the signal magnitude vector (
$SMV$
) are commonly used to classify postural sways and falls. Although
$XYZ$
and
$SMV$
provide satisfactory results when acceleration measurements are not noisy, their performance degrades significantly in the presence of noise. To address this issue, in this letter, four signals (
$DDVD$
) are extracted from
$XYZ$
and used for the postural sway classification with a newly proposed lightweight convolutional neural networks (CNNs) classifier model. Extensive experiments conducted using clean and noisy signals demonstrate that the use of
$DDVD$
signals, along with
$XYZ$
or all signals (
$SMV$
and
$XYZ$
), provides the best performance for both binary classification (stable
versus
unstable) or multiclass classification (standing
versus
AP
versus
ML
versus
unstable). In summary, utilizing
$DDVD$
signals improves the performance results even with noisy signals, making them valuable signals for classifying postural sway.