Fairness is an important factor to consider in graph neural networks (GNNs) as biases in the data can be amplified by the link structure. Despite ongoing research, existing fairness-aware GNN methods often assume that the sensitive attribute values for all demographic groups are available during training. This assumption restricts their practical applicability, especially in scenarios where training examples for certain demographic groups are unavailable. To address this limitation, we propose FairGRUNT, a novel GNN framework designed to handle training scenarios with unseen demographic groups. FairGRUNT employs disentangled representation learning to separate node embeddings for class prediction from those encoding demographic information, thereby reducing dependency between them. A pretraining stage assigns soft labels to a subset of the unlabeled nodes indicating seen or unseen group membership based on their prediction confidence. These soft labels are then used to train a demographic classifier and guide a statistical parity-based fairness regularizer, which is integrated into the training objective to mitigate bias in the GNN predictions. Experimental results on real-world datasets show that FairGRUNT outperforms traditional fairness-aware GNN methods in reducing biases in node classification, particularly for sensitive attributes with unseen groups.