BrainSF: A Foundation Model for Whole Brain Functional Signals Forecasting

1 - The paper lacks clear organization, particularly in the Methods section, which requires a more detailed description of the model architecture to enable thorough evaluation.

2 - Given the brain's dynamic complexity and inherent chaotic behavior, foundational models should include mechanisms for capturing uncertainty in time-series data, such as using variational encoder-decoder architectures for sequence-to-sequence modeling or incorporating partial differential equation (PDE) terms in the loss function for physics-informed neural networks (PINNs). The proposed approach, however, does not address this crucial aspect.

3 - The abstract claims that the authors propose a self-supervised model to embed whole-brain fMRI signals into vector representations for continuous future prediction. However, the Methods section indicates the use of 1,000 parcels, which suggests a misalignment between the stated objective and actual implementation.

4 - The paper lacks a dedicated interpretability section.

5 - The reported generalization metrics ($R^2 , R, MSE$) require contextualization; how do we know they are good?

6 - The computational complexity of the model is significant, as highlighted in Section 3.3.

• A primary concern is the lack of baseline comparisons in this study. Although the authors claim that this is the only work addressing the fMRI forecasting problem, prior fMRI foundation models, such as BrainLM (Ortega Caro et al., 2023), which employs masked-prediction training, can also perform forecasting by placing the mask over the prediction segment. Additionally, the forecasting horizon in this study (up to 30 TR) aligns with the sequence length used in BrainLM (200 TR, if recalled correctly). The authors' rationale for not including a comparison with BrainLM is unclear.

• This absence of baselines raises a second concern: the model architecture used in this paper appears somewhat antiquated techniques. Without comparisons to other models, there is no evidence that the current methodology is more effective or novel for fMRI forecasting compared to contemporary time-series forecasting architectures. Beyond the MAE architecture previously applied to fMRI data, numerous advanced time-series forecasting models could be potential baselines that could be evaluated on fMRI data, such as those available in the “Time-Series Library” repository (https://github.com/thuml/Time-Series-Library/tree/main).

• The study does not address the influence of fMRI auto-correlation, a significant factor in fMRI forecasting. Typically, fMRI signals exhibit auto-correlation across a time lag of approximately -10 to 10 seconds, potentially making short-term forecasting within this range easier. In this study, the forecasting horizon for Table 1 is set to 15 TR, which largely falls within this auto-correlation window. While this does not imply an issue with the methodology itself, it would be valuable to examine prediction performance when introducing a gap between the input and forecasting time windows to reduce the effects of auto-correlation. At a minimum, the authors should discuss this factor, as it plays an important role in fMRI forecasting task.

1. Lack of Discussion on Computational Time: The paper does not address the computational time required for processing large-scale fMRI datasets, which is a critical factor given the substantial number of time points involved. This omission makes it difficult to evaluate the practicality of the proposed model in real-world applications.

2. Limited Utility of Time Point Reconstruction: While the model focuses on reconstructing future time points, this approach alone may not provide sufficient insights into brain dynamics. The authors should incorporate downstream tasks to validate the model's performance more effectively, such as analyzing functional connectivity in resting-state fMRI or examining activation maps in task-based fMRI studies.

3. Insufficient Validation through Diverse Applications: The experimental results primarily demonstrate zero-shot forecasting, but the lack of validation through additional applications, such as functional connectivity analyses, limits the understanding of the model's practical utility in broader neuroscience research contexts. Incorporating such analyses would strengthen the case for the model's effectiveness and relevance.