A Factor Graph-Based Change Point Detection Algorithm With an Application to sEMG-Onset and Activity Detection

Change point detection (CPD) algorithms are relevant tools to achieve triggering of various functions, e.g., in medical support devices. In the context of mechanical ventilation, one such application exists in detecting muscle activity based on electromyographic (EMG) measurements of the diaphragm. Change point detection algorithms to be applied in this setting are required to reliably detect the onset of the EMG signal in real-time and are usually desired to operate on the raw signal, thus minimizing the required effort for prior signal processing. In turn, information about the periods of muscular activity facilitates a wide range of subsequent signal processing and estimation algorithms. A novel algorithm for EMG-onset and activity detection is proposed based on a probabilistic graphical model, formulated as a factor graph. Factor graphs form a class of probabilistic graphical models representing the factorization of probability density functions as bipartite graphs. They can be used to exploit the conditional independence structure of the underlying model to effciently solve inference problems by message passing on graphs. Hence, the factor graph framework is capable of recovering a wide range of classical results in signal processing, estimation and control in a unified framework. The present work advocates the use of this class of models in the field of change point detection and activity estimation. Based on a combined factor graph representation of both the Kalman filter and the expectation maximization algorithm, regularized signal estimation is achieved. In conjunction with a simple dynamic model, the sparsity of the estimated input results in a filtered state estimate denoting the estimated activity level. Thresholding on this signal yields the desired detection of the onset and activity of the EMG signal. Possible extensions are outlined for automated adaptation of the threshold levels. The presented example highlights the efficacy of the proposed method on clinical data