Back to search

Boston University

Contributions to state-space modeling with applications in neural data analysis

Abstract

dc:description.abstract

The state-space model is a powerful mathematical framework for describing and analyzing latent dynamic processes. It is particularly well-suited for neural data analysis, where observed signals are shaped by underlying, unobserved processes. By applying this model to complex neural phenomena, we gain a natural approach to integrating information across multiple neural sources and spatiotemporal scales. Moreover, it enables the discovery of low-dimensional representations of behavioral and cognitive states while providing a robust means of expressing confidence in estimates and inferences. With the remarkable technological advances in neuroscience over the past few decades, our ability to record neural activity has expanded dramatically. These break-throughs have empowered neuroscientists to explore increasingly intricate aspects of neural coding and communication. Due to this fact, we now have access to neural signals from diverse sources, different regions, and various types of neural activity. This abundance of information presents several crucial challenges. First, it is important to determine how to effectively integrate signal information from multiple origins to make inference about the latent neural dynamics. Moreover, as we construct increasingly complex models for neural signal data, assessing their goodness-of-fit becomes essential. This challenge is further magnified when we employ knowingly misspecified models for the sake of computational efficiency and interpretability. In such cases, identifying periods of model misspecification is not only a technical necessity but also a fundamental step toward deeper scientific understanding. In this work, we extend neural state-space modeling methods in two key directions. First, we introduce a novel hierarchical state-space model framework that integrates information from two fundamental classes of neural data: action potentials, commonly known as spikes, and rhythmic local field potentials (LFPs). We apply this framework to the challenge of identifying hippocampal ripple-replay events—periods characterized by prominent high-frequency oscillations in the hippocampus, during which spiking activity reflects non-local spatial structure. Our model incorporates three latent processes: 1) a continuous state process representing continuous trajectories through an environment, 2) oscillatory dynamics shaping LFPs, and 3) a discrete state governing transitions between local representation and ripple-replay events. These latent states drive two observed processes: 1) marked point processes capturing spatially tuned spikes and 2) LFPs modulated by oscillatory dynamics. Building on this model structure, we develop a causal filter algorithm using only past information and a noncausal smoother algorithm using all the data from a fixed interval to detect ripple-replay events and decode their content. We validate our approach on both simulated and real neural data, offering new insights into the strengths and limitations of our model and the underlying mechanisms of replay events. The second extension of our neural state-space modeling framework focuses on evaluating its goodness-of-fit over short time intervals. In many neural data analysis problems, models are intentionally misspecified, as constructing fully accurate models would be computationally prohibitive while offering only marginal inferential gains. To address this challenge, we developed three classes of metrics designed to identify localized periods where model misspecification leads to inaccurate estimates. Each metric assesses the consistency between the information contained in individual observations and the information extracted by the state-space model from a local window of data. These metrics fall into three categories: 1) a set of divergence measures, specifically F-divergences, which compare the distributions of the latent state process directly; 2) measures of overlap between estimated credible regions obtained from the state distributions; and 3) predictive checks for individual observations, based on samples from the estimated state process. We applied these metrics to the problem of estimating nonlocal spatial representations in both simulated and real hippocampal spiking data, demonstrating the strengths and limitations of our approach. Together, these advancements pave the way for more precise data-driven decision-making and a deeper understanding of latent processes across neuroscience and even other disciplines.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Mathematics & Statistics
Grantor
Boston University
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Zeng, Sirui
Advisor dc:contributor.advisor
  • Eden, Uri T.

Subjects

dc:subject × 1

Rights

Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/2144/53395

Chain of custody

source
Harvested from
Boston University
Base URL
open.bu.edu/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Zeng, Sirui. Contributions to state-space modeling with applications in neural data analysis. doctoral thesis, Boston University, 2025. https://hdl.handle.net/2144/53395