Machine Learning for Interpretable Decoding of High-Dimensional Time Series: Toward the Discovery of a Neural Biomarker for Anxiety

Abstract

This dissertation advances the field of interpretable machine learning for high-dimensional neural time series by developing new methodologies for decoding latent structure from brain-wide neural recordings. Motivated by the challenge of identifying robust, biologically interpretable, brain networks descriptive and predictive of internal emotional states, particularly anxiety, this work presents a series of methodological innovations that address critical limitations in current machine learning approaches for structured neural data. Beyond developing novel methodology for brain network discovery, this work also presents an anxiety brain network which is validated broadly via numerical, qualitative, and behavioral strategies.Our brain networks consist of spectral features derived from local field potentials (LFPs), which operate as a neighborhood voltage in the brain around a given brain region. By recording LFPs while mice are simultaneously engaged in behaviors relevant to an emotional state of interest, machine learning methodologies can be applied to learn latent networks of LFP activity that explain behaviors. Previous works have made use of an interpretable supervised factor model known as discriminative cross-spectral factor analysis with nonnegative matrix factorization (dCSFA-NMF) to identify brain networks descriptive of stress, aggression, and social behavior. Traditionally, this framework has been applied by training on a carefully chosen behavioral context and then subsequently evaluating on new subjects in the same or related contexts. While this framework has been successful before, we demonstrate that the traditional modeling framework fails for our anxiety contexts – inspiring developments both in experimental training strategies and optimization frameworks. Key machine learning contributions include: (1) the development of a multi-assay supervision framework that enables the model to learn shared representations across heterogeneous behavioral paradigms; (2) the introduction of a cosine distance-based stability criterion for brain network representations across training iterations to assess and enforce the reproducibility of learned latent factors; and (3) a staged training procedure that integrates unsupervised NMF-based factor initialization, encoder pretraining for representational alignment, and domain-aware optimization with sample rebalancing. These methods are explicitly designed to identify brain networks that are independent of confounding behavioral states and more likely to generalize broadly in the population for predicting behavior and emotion. Scientific utility of the novel methodologies is validated in the context of decoding internal anxiety states in mice. Using neural recordings from eight brain regions across three distinct anxiety-related experimental paradigms – elevated plus maze, bright open field, and acute fluoxetine administration – we demonstrate that single-assay models fail to generalize beyond their training context. In contrast, the proposed multi-assay dCSFA-NMF framework identifies a brain-wide electome network that consistently and specifically encode an internal anxious state and a brain-wide electome network that may be more broadly predictive of negative affect. These networks generalize to held-out subjects and conditions, remain interpretable in terms of neuroanatomical structure and frequency content, and are validated through behavioral assays, disease models, and optogenetic manipulation. Further extending the generality of the modeling framework, I present preliminary work toward a "Universal Mouse Code of Emotion" - a multi-behavior, multi-dataset modeling architecture capable of decoding multiple internal states (anxiety and social behavior) across datasets with missing data due to non-overlapping recording configurations or sensor failure. This is achieved through the integration of a transformer-based encoder, variational inference, and a reconstruction-prediction joint training objective. The resulting model is shown to effectively impute missing data, maintain predictive accuracy across behaviors, and preserve interpretability through linearly constrained decoder weights and network sparsity. Collectively, this dissertation provides a comprehensive framework for interpretable, domain-general machine learning in the analysis of high-dimensional neural time series. The contributions made here extend the theoretical and practical utility of structured factor models in neuroscience, while also laying foundational tools applicable to other domains involving complex temporal signals with missing data, multi-task structure, and interpretability constraints. The anxiety biomarker serves as a case study that rigorously tests and demonstrates the efficacy of these methods, but the modeling principles and algorithms developed herein are broadly extensible to challenges in physiological signal analysis, wearable computing, behavioral sensing, and adaptive closed-loop systems. These contributions thus represent a significant advancement at the intersection of machine learning, neuroscience, and biomedical signal processing.