Code for Paper:

Moving Beyond Functional Connectivity: Time-Series Modeling for fMRI-Based Brain Disorder Classification (IEEE TMI)

[Paper] [中文解读]

Introduction

🌟 Although FC remains the dominant fMRI input, we find that GCNs/Transformers perform better when fed tokenized raw BOLD time series. Shuffling BOLD temporal order while preserving FC reduces performance to the FC baseline.

🏆 We benchmark several recent time-series models (e.g., Leddam, iTransformer; the red boxes below indicate time-series–based methods) and find that they generally outperform traditional FC/dFC methods.

🌟 This motivates us to explore how recent advances in time-series analysis can benefit fMRI modeling. Building on two key principles in modern time-series research, Channel-Independence (CI) and Seasonal–Trend Decomposition, we propose DeCI, which performs deep cycle (seasonal) and drift (trend) decomposition via progressive residual extraction, models each ROI time series in a CI manner, and fuses predictions at the logit level.

🏆 DeCI (Channel-Independent) is more noise-robust than other Channel-Dependent baselines (e.g., iTransformer).

🌟 Seasonal–Trend Decomposition (or Cycle–Drift Decomposition for fMRI) substantially enhances the discriminability of raw features.

🏆 DeCI achieves strong performance with low computational overhead.

Usage

1. Install requirements. pip install -r requirements.txt

2. Download data. You can download all the datasets from datasets. All the datasets are well pre-processed and can be used easily. Then place them under a folder ./dataset.

3. Train the model. We provide the experiment scripts of all benchmarks under the folder ./scripts.

4. You can use bash commands to individually run scripts in the 'scripts' folder from the command line to obtain results for individual datasets. For example, you can use the command below to obtain the result of DeCI on TaoWu:

   bash scripts/DeCI/Taowu.sh

You can find the training history and results under the './logs' folder.

Meanwhile, the scripts folder contains all the execution scripts for our DeCI model, as well as scripts for FC-based methods (under the FC folder), dFC-based methods (under the dFC folder), general time-series models (under the GeneralTS folder), Multi-View-based methods (under the Multi_View folder), and Attention-based methods (under the Attn folder).

To reproduce the full set of DeCI results reported in the paper, you can run:

python hrun.py --opt 1

To run all the FC-based baselines, use:

python hrun.py --opt 2

To run all the dFC-based baselines, use:

python hrun.py --opt 3

For general time-series methods, use:

python hrun.py --opt 4

For Multi-view based methods, use:

python hrun.py --opt 5

For Attention-based methods, use:

python hrun.py --opt 6

Once the experiments are complete, you can run:

python extract_re.py

This script will automatically aggregate and organize the logs, generating the final performance tables based on the best hyperparameter configurations.

Acknowledgment

We appreciate the following GitHub repos a lot for their valuable code and efforts:

• Time-Series-Library (https://github.com/thuml/Time-Series-Library)
• ModernTCN (https://github.com/luodhhh/ModernTCN)
• Leddam (https://github.com/Levi-Ackman/Leddam)
• Jiaxing Xu et al. (https://github.com/brainnetuoa/data_driven_network_neuroscience)

Citation

If you find this repo helpful, please cite our paper.

@article{Yu2026DeCI,
  title        = {Moving Beyond Functional Connectivity: Time-Series Modeling for fMRI-Based Brain Disorder Classification},
  author       = {Yu, Guoqi and Hu, Xiaowei and Aviles-Rivero, Angelica I. and Qiu, Anqi and Wang, Shujun},
  journal      = {arXiv preprint arXiv:2602.08262},
  year         = {2026},
  url          = {https://arxiv.org/abs/2602.08262},
  doi          = {10.48550/arXiv.2602.08262}
}