abide-i

Multi-modal CNN-BERT with leave-one-site-out cross-validation. Combines brain MRI and meta-features.

Autoencoder + FCN combining fMRI and sMRI data (Rakic et al., 2020).

Multi-atlas deep neural network with hinge loss. 79.13% on augmented data.

Shared-weight feature extraction network. Precision: 76.15%, Recall: 80.65%.

Multi-attention CNN. AUC: 0.79 on ABIDE-I.

Adversarial learning-based node-edge graph attention network. 1,007 subjects across 17 sites.

ROI-aware graph convolutional network. Interpretable biomarker discovery. 1,035 subjects.

Graph Convolutional Network combining fMRI and sMRI with max voting.

Multi-task transformer neural network on UM dataset. Attention mechanism for feature extraction.

Graph contrastive learning with graph transformer. 74.8% sensitivity.

Support Vector Machine with functional connectivity features. Classic baseline comparison.

Deep learning approach by Heinsfeld et al. (2017). Anterior-posterior brain connectivity disruption.

Multi-view Site Graph Convolutional Network handling multi-site variability.

Participant-specific functional connectivity matrices (Abraham et al., 2017).

Random Forest baseline. Sensitivity: 69%, Specificity: 58%.

Graph Transformer for brain disorder diagnosis. Significantly outperforms BrainNetTF (73.2%).

Best performing model in comprehensive comparison study (2024).

Traditional ML baseline with functional connectivity matrices.

Multi-view approach addressing site heterogeneity.