Ensemble Coupled Hidden Markov Models for Joint Characterisation of Dynamic Signals

Iead Rezek, Stephen J. Roberts, Peter SykacekAISTATS 2003 · OpenReview

TL;DR

This paper introduces Coupled Hidden Markov Models (CHMMs) within a Variational Bayesian framework to jointly model and characterize dynamic signals and their associated noisy, dynamic labels. By treating labels as noisy observables and allowing for flexible temporal dependencies between signal and label chains, CHMMs offer a more robust and interpretable approach to complex time-series analysis.

What Problem Does This Paper Address?

Modeling dynamic systems often involves dealing with multiple types of data, such as continuous signals and discrete labels. A significant challenge arises when these labels are not perfect but are instead noisy, subjective, and possess their own temporal dynamics, as is common with human-generated scores. Traditional methods often struggle to reconcile the different natures of these data streams and account for the inherent uncertainty and temporal relationships between them. For instance, human scoring of sleep stages might lag behind actual state changes, and labels might be generated at a different sampling rate than the primary signal. This paper aims to address this by developing a framework that can jointly model these diverse and potentially unreliable data sources, capturing their interactions and temporal dependencies.

Methods

The core contribution is the development and application of Coupled Hidden Markov Models (CHMMs) estimated using a Variational Bayesian framework.

  • Coupled Hidden Markov Models (CHMMs): Unlike standard HMMs, CHMMs allow for multiple "chains" that model different aspects of the data. In this work, one chain models the primary dynamic signal (e.g., physiological data), and another models the associated labels. Crucially, these chains can influence each other through "linking edges" in the network topology, which can be estimated to capture temporal lags or leads between them.
  • Variational Bayesian Framework: This approach provides a principled way to approximate the intractable posterior distribution of model parameters. It offers several advantages over maximum likelihood methods, including robustness to issues like singular covariance matrices and the ability to integrate out parameters, leading to more stable estimations. The framework also provides a "free energy" measure that can be used for model selection (e.g., choosing the optimal state-space dimension or lag).
  • Integrated Feature Extraction and Modeling: The CHMMs are equipped with linear observation models, allowing for integrated feature extraction and modeling. This means that features can be learned directly within the model rather than being pre-processed and fed in as fixed inputs.
  • Multivariate Linear Observation Models: For continuous signals, a multivariate autoregressive (AR) model is used. For discrete labels, a multinomial observation model is employed.

Datasets

The paper uses synthetic data for initial model evaluation and then applies the CHMMs to two real-world datasets:

  • Cheyne Stokes Data: This dataset consists of simultaneous EEG and respiration recordings, characterized by distinct states related to breathing disorders.
  • Brain Computer Interface (BCI) Data: This dataset involves surface EEG recordings from subjects performing cognitive tasks (auditory-imagination and left-right motor imagination).

No new benchmarks were introduced; the focus was on applying the developed CHMM framework to existing data types.

Key Insights and Ideas

  • Labels as Noisy Observables: A key conceptual shift is treating labels not as ground truth but as noisy manifestations of the underlying system, similar to the primary signals. This allows for uncertainty in the labels to be explicitly modeled.
  • Flexible Temporal Dependencies: The ability to estimate the lag and lead relationships between different signal chains is crucial for understanding how different dynamic processes interact over time. This goes beyond simple cross-correlations.
  • Robustness of Variational Bayesian Estimation: The variational approach proves to be more robust than maximum likelihood methods, particularly in handling noisy data and avoiding issues like singular matrices.
  • Model Selection via Free Energy: The free energy provides a principled criterion for selecting optimal model structures, including state-space dimensions and temporal lags, directly from the data.
  • Automatic State Pruning: The variational implementation naturally prunes states that are not supported by the data, simplifying the model without explicit regularization.

Key Results

  • Synthetic Data: The CHMMs successfully identified the correct state-space dimensions and lags when tested on synthetic data, demonstrating the model's ability to recover underlying structure. The free energy landscape clearly indicated the optimal model configuration.
  • Cheyne Stokes Data: The CHMMs provided a segmentation of the data that aligned with known physiological regimes, similar to results obtained with single HMMs but with the added benefit of automatic state pruning.
  • BCI Data: The CHMMs achieved comparable or slightly better generalization accuracies than previous sampling-based methods on classification tasks, particularly when dealing with corrupted labels. The variational approach was also significantly faster in converging.

Limitations and Future Work

  • Lag Estimation Sensitivity: For real-world data, the free energy landscape for lag estimation can be relatively flat, making it difficult to pinpoint a single "optimal" lag. This suggests that the choice of feature extraction window size can influence the perceived lag.
  • Outlier Handling: While the variational approach is more robust than sampling methods to certain data corruptions, sections with extreme saturation (leading to singular matrices) still needed to be removed. The authors suggest that an "outlier model" might further improve performance in such cases.
  • Future Work: The paper implicitly suggests exploring more sophisticated observation models and investigating the impact of different feature extraction strategies on lag estimation. Further research into handling extreme outliers and extending the framework to more complex network topologies could also be valuable.

Summary updated

Abstract

How does one model data with the aid of labels, when the labels themselves are noisy, unreliable and have their own dynamics? How does one measure interactions between variables that are so different in their nature that a direct comparison using, say cross-correlations, is meaningless? In this paper these problems are approached using Coupled Hidden Markov Models which are estimated in the Variational Bayesian framework. Signals can be diverse since each chain has its own observation model. Signals can have their own dynamics and may temporally lag or lead one another by allowing linking edges in the network topology to be estimated and chosen according to the most probable posterior model. Integrated feature extraction and modelling is accomplished by providing the Markov models models with linear observations models. We derive Coupled Hidden Markov Models estimators, apply and compare them with sampling based approaches found in the literature.
BibTeX
@inproceedings{rezek2003ensemble,
  author = {Iead Rezek and Stephen J. Roberts and Peter Sykacek},
  title = {{Ensemble Coupled Hidden Markov Models for Joint Characterisation of Dynamic Signals}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2003},
  url = {https://openreview.net/forum?id=cmJLcVEW3O},
}
Figure