Abstract
While functional Magnetic Resonance Imaging (fMRI) can provide detailed information regarding the functional activity of the whole brain, its cumbersome experimental setting and high cost prevent its application in ecological conditions. This represents a challenge in the context of neurofeedback therapy. To address this challenge, one possible approach is to train a predictor for the localized functional activity using multimodal EEG-fMRI records and utilize that predictor in real-time EEG sessions. Using publicly available multimodal real-time EEG-fMRI data, we present a detailed evaluation of regularized linear regressors both in the context of predicting localized fMRI activity and to characterize the properties of individual EEG-fMRI runs. Our results indicate that while it is possible to find clusters of similar time-series in which localized brain activity is highly predictable (Pearson’s r=0.43), the capacity to train regularized linear regressors able to generalize to new subjects remains limited (r=0.24), highlighting two distinct strategies in EEG-fMRI neurofeedback settings. We characterize the fundamental role for the clustering distance used to identify similar time-series, based both on theoretically-grounded considerations and experimental observations. Our results suggest a clear preference for the cosine metric, instead of the Pearson-based metric utilized in current literature. We further suggest that highly-predictable clusters in regular offline paradigms would correspond to high-performing subjects in the context of real-time neurofeedback, and evaluate a regressor-free clustering strategy based on the Log-Euclidean distance of covariance matrices.
Introduction: The Neurofeedback Challenge
Neurofeedback (NF) is a therapeutic tool where patients learn to regulate their own brain activity by observing a real-time representation of their brain function. While functional Magnetic Resonance Imaging (fMRI) provides superior spatial resolution for targeting specific brain structures, its high cost and restrictive environment limit clinical accessibility.
Electroencephalography (EEG) offers a portable, low-cost alternative but lacks the spatial precision of fMRI. To combine the strengths of both modalities, researchers develop "proxy" models: they record EEG and fMRI simultaneously, train a machine learning model to predict the fMRI signal from EEG features, and then deploy that model in standalone EEG sessions. This paper addresses the challenge of improving these models by refining how brain patterns are grouped (clustered) to create more accurate, specialized predictors.
Methodology: Refining the EEG Finger-Print (EFP)
The study focuses on optimizing the EEG Finger-Print (EFP) method through several key technical refinements:
- Feature Extraction: Raw EEG signals were converted into time-frequency features (power spectra) using the Stockwell transform.
- Regularized Regression: Ridge Regression (L2 regularization) was employed to map EEG features to fMRI signals in the Primary Motor Cortex (M1) and Supplementary Motor Area (SMA).
- Advanced Clustering: The authors replaced standard Pearson-based similarity with Cosine distance. They demonstrated that Pearson correlation is suboptimal because it ignores "translation shifts" in model weights, whereas Cosine distance is sensitive to these shifts while remaining scale-invariant.
- Dimensionality Optimization: Using intrinsic Principal Component Analysis (lPCA), the researchers determined that 12 frequency bands provided better performance than the standard 10.
Key Results and Insights
The evaluation revealed a significant performance gap between specialized models and those intended to generalize across subjects.
- Prediction Accuracy: Using Cosine or Canberra distance, the researchers achieved a Pearson correlation of r = 0.43 within specialized clusters. However, when generalizing to entirely new subjects (Leave-One-Subject-Out), accuracy dropped to r = 0.24 for M1 and r = 0.20 for SMA.
- Spatial Performance: Topographic maps illustrate how prediction performance varies across the scalp, identifying the most informative EEG channels for motor-related fMRI activity.
- The "Pearson Pitfall": Theoretical analysis confirmed that Pearson-based clustering can group dissimilar models due to its translation invariance. Switching to Cosine distance significantly improved the identification of similar time-series.
- Predictability as an Individual Trait: The clustering algorithm naturally grouped data from the same individuals, suggesting that "predictability" is an inherent trait of certain "high-performing" subjects who adhere more consistently to experimental protocols.
Conclusion and Future Directions
While specialized models show high accuracy, the "generalization gap" remains a significant hurdle for universal EEG-to-fMRI predictors. The study suggests a clear preference for Cosine metrics in clustering and highlights the potential of regressor-free strategies, such as Log-Euclidean distance of EEG covariance, to identify high-performing subjects. Future work may involve deep learning approaches and larger datasets to further bridge the gap between EEG and fMRI spatial resolution.
Summary updated