TL;DR
This paper challenges the common assumption that the brain's macroscopic dynamics must be nonlinear for accurate modeling. Using system identification on fMRI and iEEG data, the authors demonstrate that simple linear models outperform complex nonlinear ones in predicting brain activity, suggesting that averaging, noise, and limited data effectively linearize neural signals at the macroscale.
What Problem Does This Paper Address?
Computational neuroscience faces a persistent trade-off between model accuracy and simplicity. While individual neurons exhibit complex nonlinear dynamics, it's often assumed that whole-brain models must also be nonlinear to be accurate. This assumption, however, has rarely been rigorously tested by directly comparing the predictive power of linear and nonlinear models on real neuroimaging data. The authors aim to settle this question by applying system identification techniques to resting-state fMRI and iEEG data.
Methods
The core methodology is system identification, a data-driven approach to building and evaluating dynamical models. The authors compare various linear and nonlinear model families using a prediction error (PE) framework. Key metrics for model comparison include:
- Predictive Power (R²): How well a model predicts future time series values from past data.
- Whiteness of Residuals: How much unmodeled dynamics remain after fitting.
- Computational Complexity: The time required to fit the models.
They also investigate the linearizing effects of four macroscopic properties: spatial averaging, temporal averaging, observation noise, and limited data samples, using both simulated and real data.
Datasets
- Human Connectome Project (HCP): Resting-state fMRI data from 700 subjects.
- Restoring Active Memory (RAM) project: Resting-state iEEG data from 122 subjects.
Key Insights and Ideas
- Linearity Reigns Supreme (at the Macroscale): Contrary to intuition, linear autoregressive models consistently outperformed a range of sophisticated nonlinear models in predicting resting-state brain activity across both fMRI and iEEG.
- The "Linearizing" Power of Aggregation: The authors identify four key factors that can mask or counteract microscopic nonlinearities at the macroscopic level:
- Spatial Averaging: Averaging signals across large neuronal populations smooths out nonlinearities.
- Temporal Averaging: Low-pass filtering inherent in neuroimaging signals (like the HRF in fMRI) also acts as temporal averaging.
- Observation Noise: Scanner or electrode noise can blur nonlinear relationships.
- Limited Data Samples: Insufficient data can make it difficult to detect true nonlinearities, especially in high-dimensional systems.
- Interpretability Wins: The superior performance of linear models, combined with their inherent interpretability, makes them a highly attractive choice for understanding macroscopic brain dynamics.
- HRF Nuances: Explicitly modeling the hemodynamic response function (HRF) did not improve the performance of linear models and had minimal impact on nonlinear ones, suggesting that linear autoregressive models can implicitly capture some HRF-like dynamics.
Key Results
- fMRI: Linear models achieved the highest R² values and the whitest residuals, while also being the most computationally efficient.
- iEEG: Similar trends were observed, with linear autoregressive models outperforming nonlinear ones. iEEG data showed higher R² values overall, likely due to its higher sampling rate and smoother signals.
- Computational Efficiency: Linear models were at least an order of magnitude faster to train than nonlinear models.
- Model Comparison: The study provided a comprehensive, data-driven comparison of various linear and nonlinear model families, offering a valuable benchmark for future modeling efforts.
Limitations and Future Work
- Space of Nonlinear Models: The authors acknowledge that their tested nonlinear models do not represent the entire vast space of possible nonlinear models, and thus their results do not definitively prove that no nonlinear model can outperform linear ones.
- Task-Based Dynamics: The current study focused on resting-state data. Future work could explore whether nonlinearities become more apparent during task-based conditions, where external inputs and more pronounced signal changes are present.
- Input Models for Task fMRI: Extending the framework to task fMRI would require developing accurate data-driven "input models" that capture how experimental stimuli influence brain signals.
- Physiological Mechanisms: While the study identifies factors that linearize dynamics, further research is needed to fully understand the mechanistic physiological reasons behind these effects and how they interact.
- Model Interpretability vs. Causality: While linear models are interpretable, understanding the causal influence of these macroscopic linear signals on microscopic activity remains an open question.
Summary updated