Classification-Based Prediction of Effective Connectivity Between Timeseries With a Realistic Cortical Network Model

Emanuele Olivetti, Danilo Benozzo, Jan Bím, Stefano Panzeri, Paolo AvesaniFrontiers Comput. Neurosci 2018 · DOI · OpenReview

TL;DR

This paper introduces a novel machine learning approach to predict effective connectivity (causal relationships) between brain regions using time-series data. By training a classifier on simulated data from a biophysically realistic neural network model, they achieve significantly higher accuracy than traditional methods, offering a promising way to infer brain network dynamics.

What Problem Does This Paper Address?

Understanding how different brain regions interact is crucial for deciphering brain function. This interaction pattern is known as "effective connectivity." Traditionally, inferring this connectivity from neural recordings involves either model-free or model-based approaches. Model-based methods, especially those using biophysically plausible neural network models, offer deeper insights into underlying neural mechanisms. However, "inverting" these complex models to extract causal relationships from observed data is a significant challenge. Existing machine learning approaches for causality detection often struggle with limited data or are not tailored for complex, biophysical neural models. This paper aims to bridge this gap by developing a method that leverages the strengths of both realistic neural models and machine learning for accurate and robust effective connectivity inference.

Methods

The core of this work lies in a two-pronged approach:

  1. A Realistic Cortical Network Model: The authors extend a previously developed biophysically plausible neural network model (Mazzoni et al., 2008) to simulate interactions between multiple neural circuits (three in this study). This model generates realistic time-series data (Local Field Potentials or LFPs) that mimic actual neural activity. Crucially, the model allows for the ground truth of causal interactions to be known by design.
  2. A Classification-Based Causality Prediction Method: Instead of directly inverting the complex neural model, they train a supervised machine learning classifier. This classifier learns to predict the pattern of causal interactions (represented as a binary matrix) given a multivariate time-series as input. The classifier is trained on a large dataset of simulated time-series and their corresponding known causal configurations generated by their neural network model. This approach bypasses the need for direct model inversion and can be trained on simulated data, addressing the common issue of limited real-world data in neuroscience.

The feature space used for the classifier is based on Granger causality principles, quantifying the predictability of one time-series from the past of others.

Datasets

The paper utilizes two primary datasets:

  1. NN Dataset: This dataset is generated by the authors' novel, biophysically plausible neural network model. It consists of multivariate time-series (LFPs) and their corresponding ground-truth causal configuration matrices. This dataset is used for training and evaluating the proposed supervised method and for comparison with traditional methods.
  2. MAR Dataset: This dataset is generated using a traditional Multivariate Autoregressive (MAR) model. It serves as a benchmark and is used to test the adaptability of the proposed supervised method to different generative models.

Key Insights and Ideas

  • Bridging Biophysical Realism and Machine Learning: The paper demonstrates that a machine learning classifier can effectively learn to infer causal relationships from the complex, non-linear dynamics generated by a biophysically detailed neural network model, overcoming the difficulty of direct model inversion.
  • Simulated Data as a Powerful Training Tool: By generating a large dataset of simulated time-series with known causality, the authors circumvent the common problem of data scarcity in neuroscience, enabling robust training of supervised learning models.
  • Adaptability of the Supervised Method: The proposed method shows a remarkable ability to adapt to the underlying generative model. When trained on data from the realistic neural network model, it significantly outperforms traditional methods. When trained on simpler MAR model data, its performance becomes comparable to traditional methods, highlighting its flexibility.
  • Realistic Neural Models Produce Granger-Causal Dynamics: Despite not being explicitly designed with Granger causality principles, the generated time-series from the neural network model exhibit dynamics consistent with Granger causality, validating the model's ability to capture essential aspects of neural interactions.

Key Results

  • Superior Performance: The proposed classification-based method, when trained on the NN dataset, significantly outperforms Granger Causality Analysis (GCA) in detecting causal structures. The Area Under the Curve (AUC) for the Receiver Operating Characteristic (ROC) curves increased from 0.816 (GCA) to 0.914 (Supervised NN).
  • Adaptability Demonstrated: When the supervised method was trained on MAR data and tested on NN data, its performance dropped to levels comparable to GCA (AUC ~0.818), indicating that its superior performance on NN data is due to its adaptation to that specific generative model.
  • Validation of the Neural Network Model: Traditional analyses (cross-correlation and Geweke index) confirmed that the simulated LFPs from the neural network model accurately reflect the designed causal architecture, with expected time lags and statistical properties.

Limitations and Future Work

  • Scalability to More Nodes: The current study is limited to networks of up to three nodes. Scaling to larger networks presents a significant challenge due to the super-exponential growth in the number of possible causality configurations (classes for the classifier). The authors suggest exploring cell-based classification (multiple classifiers) or avoiding feature engineering to address this.
  • Observational Assumption: Like many causality frameworks, this method relies on observational data. Future work could explore causality frameworks that incorporate interventional data or account for confounding variables.
  • Non-linear Classification: While the current classifier uses logistic regression (linear), exploring non-linear classification algorithms could further improve accuracy.
  • Exploring Other Generative Models: Extending the approach to other types of realistic neural models and comparing performance would be valuable.

Summary updated

Abstract

Effective connectivity measures the pattern of causal interactions between brain regions. Traditionally, these patterns of causality are inferred from brain recordings using either non-parametric, i.e., model-free, or parametric, i.e., model-based, approaches. The latter approaches, when based on biophysically plausible models, have the advantage that they may facilitate the interpretation of causality in terms of underlying neural mechanisms. Recent biophysically plausible neural network models of recurrent microcircuits have shown the ability to reproduce well the characteristics of real neural activity and can be applied to model interacting cortical circuits. Unfortunately, however, it is challenging to invert these models in order to estimate effective connectivity from observed data. Here, we propose to use a classification-based method to approximate the result of such complex model inversion. The classifier predicts the pattern of causal interactions given a multivariate timeseries as input. The classifier is trained on a large number of pairs of multivariate timeseries and the respective pattern of causal interactions, which are generated by simulation from the neural network model. In simulated experiments, we show that the proposed method is much more accurate in detecting the causal structure of timeseries than current best practice methods. Additionally, we present further results to characterize the validity of the neural network model and the ability of the ...
BibTeX
@article{olivetti2018classification,
  author = {Emanuele Olivetti and Danilo Benozzo and Jan Bím and Stefano Panzeri and Paolo Avesani},
  title = {{Classification-Based Prediction of Effective Connectivity Between Timeseries With a Realistic Cortical Network Model}},
  journal = {Frontiers Comput. Neurosci.},
  year = {2018},
  doi = {10.3389/fncom.2018.00038},
}
Figure