HGFM

HGFM (Hypergraph Foundation Model) is a self-supervised foundation model for brain disease diagnosis from resting-state functional MRI (fMRI). Introduced by researchers at Tsinghua University's iMoonLab, Xi'an Jiaotong University, and Shanghai University in IEEE Transactions on Neural Networks and Learning Systems (early access, April 2025), it reframes the analysis of brain connectivity around high-order correlations rather than the pairwise edges that dominate conventional graph-based neuroimaging models.

Most connectome models represent the brain as a graph in which each edge links exactly two regions of interest (ROIs). This pairwise abstraction discards the group-level interactions—where three or more regions co-activate as a functional unit—that are increasingly understood to be diagnostically informative. HGFM instead builds hypergraphs, where a single hyperedge can connect an arbitrary set of ROIs, and learns an encoder over this richer structure through self-supervised pretraining. The pretrained encoder is then adapted to specific diagnostic tasks via few-shot fine-tuning.

By decoupling representation learning from labels, HGFM addresses a chronic obstacle in clinical neuroimaging: disease-specific labeled fMRI cohorts are small, heterogeneous, and expensive to assemble. A shared encoder pretrained across pooled cohorts can transfer to individual diseases with limited supervision, positioning HGFM within the broader move toward foundation models for connectomics.

#Key Features

• High-order hypergraph representation: Models the brain with hyperedges that connect multiple ROIs simultaneously, capturing group-level functional interactions that pairwise graphs cannot express.
• Dual-scale self-supervised pretraining: Combines a brain functional-network link prediction task on individual subject networks with a group interaction-network link prediction task on population-level networks, learning both individual and cohort structure without diagnostic labels.
• Few-shot adaptation: A single pretrained encoder is fine-tuned with few-shot learning to each downstream diagnostic task, reducing reliance on large per-disease labeled datasets.
• Cross-disease generalization: Evaluated across four distinct brain diseases, demonstrating that the learned high-order representations transfer rather than overfitting to one cohort.
• Clinically oriented design: Built on resting-state fMRI, the modality most widely available in psychiatric and neurological research, easing integration with existing pipelines.

#Technical Details

HGFM is built on the hypergraph computation paradigm, using hypergraph neural network encoders implemented with the DHG (DeepHypergraph) toolkit from iMoonLab. Pretraining proceeds in two self-supervised stages: link prediction on individual subject-level brain functional networks, and link prediction on group-level brain interaction networks that aggregate population structure. This dual objective lets the encoder learn representations spanning both within-subject connectivity and across-subject commonality. Downstream diagnosis is performed by attaching a classification head to the pretrained encoder and fine-tuning with few-shot learning. The model was evaluated on fMRI data from 4,409 subjects spanning four brain diseases, and the authors report that it outperforms existing state-of-the-art methods across all four diagnostic tasks.

#Applications

HGFM targets computational neuroscientists and clinical researchers who diagnose or stratify brain disorders—such as psychiatric and neurodegenerative conditions—from functional connectivity. Because the encoder is pretrained once and adapted with only a handful of labeled examples, it is well suited to rare-disease cohorts, multi-site studies with limited per-site annotation, and exploratory biomarker work where assembling large supervised datasets is impractical. Its high-order formulation is also useful for studying group-level functional modules that pairwise connectivity analyses overlook.

#Impact

HGFM extends the foundation-model paradigm into hypergraph-based connectomics, an area where most prior work relied on task-specific, pairwise-graph models trained from scratch on individual cohorts. By showing that high-order, label-free pretraining transfers across four diseases on a 4,409-subject benchmark, it offers evidence that shared brain-connectivity representations can ease the data-scarcity bottleneck in clinical neuroimaging. The main limitation to broad adoption is openness: at publication the authors did not release pretrained weights, training code, or a model/data card for HGFM specifically, which constrains immediate reproducibility despite the availability of the general-purpose DHG library the work builds upon.