A Hybrid Factor Graph Model for Biomedical Activity Detection

For activity detection on biomedical time-series data, biomedical signals are modeled as a switching linear dynamical system with random variables, including discrete and continuous dynamics. We present a formalism for representing a system's joint probability density function as a hybrid factor graph. Solving inference problems is based on belief propagation using message passing. Inference results yield the activity estimations in terms of probability distributions instead of binary decisions. This work builds on previous efforts to consolidate factor graphs as unifying representations for signal processing algorithms. We show that the formalism can be successfully applied to detect activities in surface electromyography data acquired during walking. The modularity of factor graphs enables the straightforward adoption and extension of the formalism expanding its scope of application.