RISE-iEEG: Robust to Inter-Subject Electrodes Implantation Variability iEEG Classifier

TL;DR

RISE-iEEG tackles inter-subject electrode variability in iEEG decoding by introducing a participant-specific projection network that maps each subject's data to a shared latent space, followed by a common EEGNet-like discriminative network. The model is trained with per-participant updates and L2-regularized projection weights, enabling cross-subject generalization without requiring electrode coordinates. Across Music Reconstruction and AJILE12 datasets, RISE-iEEG outperforms HTNet, EEGNet, Random Forest, and Minimum Distance in both same-subject and unseen-subject settings, achieving higher F1 scores and demonstrating robust, interpretable neural encoding via Integrated Gradients. The approach reduces preprocessing burdens, improves generalization, and provides physiologically plausible insights into task-relevant brain regions, with code and data for reproducibility.

Abstract

Intracranial electroencephalography (iEEG) is increasingly used for clinical and brain-computer interface applications due to its high spatial and temporal resolution. However, inter-subject variability in electrode implantation poses a challenge for developing generalized neural decoders. To address this, we introduce a novel decoder model that is robust to inter-subject electrode implantation variability. We call this model RISE-iEEG, which stands for Robust to Inter-Subject Electrode Implantation Variability iEEG Classifier. RISE-iEEG employs a deep neural network structure preceded by a participant-specific projection network. The projection network maps the neural data of individual participants onto a common low-dimensional space, compensating for the implantation variability. In other words, we developed an iEEG decoder model that can be applied across multiple participants' data without requiring the coordinates of electrode for each participant. The performance of RISE-iEEG across multiple datasets, including the Music Reconstruction dataset, and AJILE12 dataset, surpasses that of advanced iEEG decoder models such as HTNet and EEGNet. Our analysis shows that the performance of RISE-iEEG is about 7\% higher than that of HTNet and EEGNet in terms of F1 score, with an average F1 score of 0.83, which is the highest result among the evaluation methods defined. Furthermore, Our analysis of the projection network weights reveals that the Superior Temporal and Postcentral lobes are key encoding nodes for the Music Reconstruction and AJILE12 datasets, which aligns with the primary physiological principles governing these regions. This model improves decoding accuracy while maintaining interpretability and generalization.

Figures (6)

• Figure 1: The RISE-iEEG model consists of two main networks: the projection network and the discriminative network. The projection network includes a participant-specific dense layer for each individual, which performs a linear transformation to map electrode data onto a common space. The discriminative network, based on the EEGNet architecture ref7, extracts meaningful features from the projected data using temporal and spatial convolutional neural networks. In this figure, the notation $P_i$ represents participant $i$, and $S_i$ represents the switch control for participant $i$. The total number of participants is denoted by $N$, while $T$ indicates the number of sample times. The dimension of the common space is represented by $R$, and $D_i$ specifies the number of electrodes for participant $i$.
• Figure 2: Performance comparison of the RISE-iEEG model (pink) with HTNet (blue), EEGNet (light blue), Random Forest (RF, purple), and Minimum Distance (MD, green) decoders, presented separately for each model in the 'same participant' setting for two tasks: (A) the Move vs. Rest task and (B) the Singing vs. Music task. Each point represents the mean F1 score of each fold across participants.
• Figure 3: (A, B) Performance comparison of RISE-iEEG (pink) with HTNet (blue), EEGNet (light blue), Random Forest (purple), and Minimum Distance (green) decoders for each participant in the 'unseen participant' setting for the (A) Move vs. Rest and (B) Singing vs. Music tasks. (C, D) Impact of training data amount on model performance in the second training step for the (C) Move vs. Rest and (D) Singing vs. Music tasks.
• Figure 4: Integrated Gradients (IG) weights for a single participant, showing the significance of each electrode at 600 ms intervals.(A) Move vs. Rest task and (B) Singing vs. Music task. The color scale indicates IG magnitude, with red representing higher importance and blue representing lower importance. This visualization reveals dynamic changes in electrode importance over time, offering insights into engaged brain regions during each task.
• Figure 5: Distribution of significant brain lobes across participants. (A) For the Move vs. Rest task, the significant lobes are as follows: Superior Temporal (ST), Rostral Middle Frontal (RMF), Middle Temporal (MT), Postcentral (PC), Precentral (PrC), Caudal Middle Frontal (CMF), Pars Opercularis (PO), Inferior Temporal (IT), SupraMarginal (SM), and Superior Frontal (SF). (B) For the Singing vs. Music task: Postcentral (PC), Superior Parietal (SP), Middle Frontal (MF), Lateral Occipital (LO), Precentral (PrC), Temporal Fusiform (TF), Central Opercular (CO), Temporal Occipital Fusiform (TOF), SupraMarginal (SM), and Occipital (O).