Postural Sway Detection using Kolmogorov-Arnold Network as Siamese Model

Continuous monitoring of postural sway is crucial for safeguarding elderly individuals and patients with neurological conditions, such as Parkinson’s disease, as their balance is significantly impacted. This continuous monitoring is achieved using triaxial accelerometer sensors that provide time series signals and are used to train models to detect the postural sway of elderly individuals or patients with neurological conditions. However, the available postural sway dataset has a limited number of samples, and the performance of trained models deteriorates when the accelerometer signal is noisy. A Siamese network Kolmogorov-Arnold Network (Siamese-KAN) is pro-posed to address this issue. This network can be trained with a few samples from each class. Training of the Siamese network is conducted by flattening the bispectrum of each accelerometer channel and then fusing the magnitudes to construct a single input vector. Various similarity functions are employed along with contrastive loss to train the Siamese network. Additionally, a Siamese network with multi-layer perceptron (MLP) is also similarly trained for comparison purposes. Preliminary results show that the Siamese-KAN model achieves better accuracy with clean signals than the Siamese-MLP. However, when the signal is noisy, the Siamese-KAN model outperforms the Siamese-MLP using Euclidean and Manhattan similarity functions, while the Siamese-MLP performs better with Cosine and Bray-Curtis similarity functions.