Interpreting neural and behavioral data using Hidden Markov Models

Measurements of neural activity and behavior are becoming increasingly detailed, producing ever-larger multidimensional datasets. Yet these measurements offer only glimpses into the workings of complex brains. Much remains hidden. Hidden Markov Models (HMMs) are versatile tools for inferring information about unobserved brain states from behavior, neural recordings, or their combination. HMMs and their extensions have been frequently and successfully used in neuroscience for this purpose. An HMM describes a Markov process in which hidden states transition between one another and generate observations that depend on the current state. HMMs are particularly well suited for naturalistic behavioral paradigms that lack a fixed trial structure, and they can provide interpretable insights into an animal's internal state. In this review, we introduce the general framework of HMMs and explain how to implement them. We comprehensively review how HMMs have been applied to neural and behavioral data, and how hidden states identified in behavior or neural activity can be used to understand neural processing. We highlight example studies in which HMMs have provided insight into the brain processes underlying behavior and perception. Our emphasis is on how specific data properties or research questions motivate different modeling choices. We also discuss the relationship between HMMs and alternative methods for latent state discovery.