Joint signal and symbol processing for a bio-sensing application

A Joint Signal and Symbol Processing (JSSP) framework is presented for a bio-sensing signal analysis application in which intramuscular electrodes are used to identify and classify action potentials from muscle fibers. The signal analysis objective in this application is to detect and track sequences of action potentials from several simultaneous sources. This task is challenging because the 3-channel sensor signals are the superposition of action potential sequences from as many as 10 or more significant sources. Also, action potentials within a sequence have shape variations that sometimes can be quite dramatic because of sensor to muscle displacements. Furthermore, the action potentials from different muscle fibers often span a large dynamic range of amplitudes, making the detection and tracking of low-amplitude action potentials significantly more difficult. Our JSSP solution for addressing these complexities utilizes the IPUS architecture for the Integrated Processing and Understanding of Signals. This architecture broadly specifies a computational organization for various application-independent aspects of a JSSP system. In view of the fact that complex signal behaviors and complex signal interactions create a myriad of special situations, the IPUS architecture also facilitates the creation of a software environment for progressively improving upon previously known JSSP solutions to address new types of signal complexities. It does this by limiting the applicability of previous solutions to an appropriate subclass of input signals to be identified via application-dependent heuristics embedded in IPUS control. Using this IPUS-based approach, our most current JSSP solution for the bio-sensing application consists of four interdependent computational receivers. Each receiver's signal processing stage produces data that in turn is used to support the processing of symbol structures representing signal analysis results. Statistical principles such as a-posteriori probability maximization, entropy minimization, likelihood maximization, and utility maximization are used to organize receiver computations. The invocation of these algorithms is under the knowledge-based control of IPUS mechanisms. On a database of challenging neuromuscular signals, the resulting JSSP system is found to achieve a classification accuracy of 85.6%, a significant improvement over the 66.0% accuracy of a system it was developed to replace. Practically speaking, the improved classification accuracy implies an order of magnitude reduction in the amount of user-interactive editing needed to make an automatic system's decomposition results amenable to scientific and clinical analysis.