Simultaneous Identification of Changepoints and Model Parameters in Switching Dynamical Systems

Many complex natural systems undergo shifts in dynamics at particular points in time. Examples include phase transitions in gene expression during the cell cycle, introduced species affecting predator-prey interactions, and disease outbreaks responding to intervention measures. Such changepoints partition timeseries into different dynamical regimes characterized by distinct parameter sets, and inference on both the changepoints and regime-specific dynamical parameters is of primary interest. Conventional approaches to analyzing switching dynamical systems first estimate changepoints, and then estimate dynamical parameters assuming the changepoints are fixed and known. Such two-stage approaches are ad-hoc, can introduce biases in the analysis, and do not fully account for uncertainty. Here, we introduce a rigorous, simulation-based inference framework that simultaneously estimates changepoints and model parameters from noisy data while admitting full uncertainty. We use simulation studies of oscillatory predator-prey dynamics and stochastic gene expression to demonstrate that our method yields accurate estimates of changepoints and model parameters together with appropriate uncertainty bounds. We then apply our approach to a real-world case study of COVID-19 intervention effects, and show that our inferred changepoints aligned closely with the actual dates of intervention implementation. Taken together, these results suggest that our framework will have broad utility in diverse scientific domains.