Neural State-Space Modeling with Latent Causal-Effect Disentanglement

AbstractDespite substantial progress in deep learning approaches to time-series reconstruction, no existing methods are designed uncover local activities with minute signal strength due their negligible contribution the optimization loss. Such however can signify important abnormal events physiological systems, such as an extra foci triggering propagation of electrical waves heart. We discuss a novel technique for reconstructing activity that, while small strength, is cause subsequent global that have larger strength. Our central innovation approach this by explicitly modeling and disentangling how latent state system influenced potential hidden internal interventions. In neural formulation state-space models (SSMs), we first introduce causal-effect dynamics via interacting ODEs separately describes 1) continuous-time intervention, 2) its effect on trajectory system’s native state. Because intervention not be directly observed but disentangled from effect, integrate knowledge intervention-free system, infer assuming it responsible differences between actual hypothetical dynamics. demonstrated proof-of-concept presented framework ectopic disrupting course normal cardiac remote observations.KeywordsNeural ODEIntervention modellingCardiac EP