Applying Biorthogonal wavelets and a Novel QuickLearn Algorithm for an Intelligent Ballistocardographic chair

• Published in: SMCals/06 : proceedings of the 2006 IEEE Mountain Workshop on Adaptive and Learning Systems, Utah State University, Logan, Utah, U.S.A., July 24-26, 2006 pp. 42 - 47
• Main Authors: Akhbardeh, A., Junnila, S., Varri, A.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.07.2006
• Subjects: Biomedical signal processing; Blood; Computer networks; Force measurement; Force sensors; Heart; Multilayer perceptrons; Sensor systems; Signal processing algorithms; Testing
• ISBN: 9781424401666; 1424401666
• DOI: 10.1109/SMCALS.2006.250690

In this paper, we classified ballistocardiogram (BCG) signals for healthy and unhealthy persons using QuickLearn (QL), a novel supervised on-line learning algorithm, and the biorthogonal spline wavelets. At the first stage, the mapping level, the input data are represented to the multi input-single output mapping function (MF) with fixed weights during a training phase. We can select any kind of mathematical function for this map and its complexity depends on input data complexity. By representing input data to MF, it gives us a scalar value. After shifting and scaling that value to the range [0,T], we can round it to have y, an integer value. The second stage, matching level, only includes an array with T cells called affine look-up table (ALT). Training phase for QL includes only one step and no learning cycles. In this single step, the integer value y is used as a reference address to call and upload label for corresponding input samples in N cells of ALT (copying label from cell [y-N/2] till cell [y+ N/2-1], data leakages to N-1 neighbor cells). In testing phase, we need only to recall and introduce the value of the cell with index y as the final output. Initial tests with BCG from six subjects (both healthy and unhealthy people) indicate that the method can classify the subjects into three classes with a high accuracy, high learning speed (elapsed time for learning around ten milliseconds), and very low computational load compared with the well-known neural networks such as multilayer perceptrons (learning time above five minutes)