Processed fNIRS Graph Representations for Explainable Graph Neural Network Modeling of Cochlear Implant Outcomes

Dataset description This dataset contains processed graph representations derived from functional near-infrared spectroscopy (fNIRS) recordings acquired from 31 postlingually deafened adult cochlear implant recipients within the first month following device activation. The data were developed for the study “Beyond Single Biomarkers: An Explainable Graph Neural Network Framework for Modeling Brain Imaging Data.” Each participant is represented by a hierarchical graph designed to capture multiple aspects of brain organization. Channel-level nodes correspond to fNIRS measurement channels grouped into predefined cortical regions of interest, including the left inferior frontal region, bilateral superior temporal and angular regions, and the occipital region. Channels within each region are connected locally, while an aggregation node represents each region and provides a higher-level representation of inter-regional interactions. Node features contain task-related information derived from audio-only and visual-only speech conditions. These include average oxyhemoglobin response waveforms, temporal features describing trial-to-trial changes and response consistency, and region-of-interest encoding. Aggregation-node features summarize the corresponding channel-level information within each cortical region. Multiple graph instances are provided for each participant following trial- and channel-level augmentation used in the study. The dataset also contains the corresponding one-year speech-understanding outcomes used as prediction targets and the cochlear implant side information incorporated into the model. The graphs can therefore be used to reproduce the graph neural network analyses, investigate alternative machine-learning approaches, or explore explainability and brain–behavior modeling methods. The shared files contain processed and derived graph representations rather than the original raw fNIRS recordings. They are intended to facilitate reproducibility of the published analysis while preserving the original participant-level imaging recordings. Source code for graph processing, model training, leave-one-subject-out evaluation, and perturbation-based explainability analyses is available at https://github.com/jesmaelpoor/Explainable-GNN-Brain-Imaging.