Learning cycle-linear hybrid automata for excitable cells

We show how to automatically learn the class of Hybrid Automata called Cycle-Linear Hybrid Automata (CLHA) in order to model the behavior of excitable cells. Such cells, whose main purpose is to amplify and propagate an electrical signal known as the action potential (AP), serve as the "biologic transistors" of living organisms. The learning algorithm we propose comprises the following three phases: (1) Geometric analysis of the APs in the training set is used to identify, for each AP, the modes and switching logic of the corresponding Linear Hybrid Automata. (2) For each mode, the modified Prony's method is used to learn the coefficients of the associated linear flows. (3) The modified Prony's method is used again to learn the functions that adjust, on a per-cycle basis, the mode dynamics and switching logic of the Linear Hybrid Automata obtained in the first two phases. Our results show that the learned CLHA is able to successfully capture AP morphology and other important excitable-cell properties, such as refractoriness and restitution, up to a prescribed approximation error. Our approach is fully implemented in MATLAB and, to the best of our knowledge, provides the most accurate approximation model for ECs to date.