Postural Sway Classification using Bispectrum

To reduce mortality and morbidity rates among elderly individuals, continuous monitoring of their postural sway is necessary. This monitoring is implemented through the use of accelerometer sensors, which provide time series signals. The bispectrum, an example of high order spectral (HOS) analysis, is employed to analyze these time series, leveraging the relationship between the various spectral components of the sensor signal for postural sway classifications. From the bispectrum magnitude, features including mean, standard deviation, and entropy are extracted and utilized to train various traditional classifiers. Abstract features are also generated using pre-trained models (MobileNet, Inception, DenseNet, and ResNet) trained with bispectrum magnitude as input and fine-tuned last five layers and the fully connected layers. The classification performance obtained using traditional features and abstract features are presented and compared with two state-of-the-art methods. In addition, the superior performance of the proposed bispectrum is demonstrated by comparing the results of using abstract features from bispectrum magnitude against the abstract features from the traditionally used spectrogram magnitude. The impact of measurement noise on the accelerometer signals on stability classification is also assessed. Under noisy conditions, abstract features extracted from bispectrum magnitude yield the best results compared to state-of-the-art methods and abstract features from spectrogram magnitude. These findings underscore the efficiency of utilizing features derived from the bispectrum for accurate postural sway classifications, even in the presence of significant additive measurement noise.