Supervised Contrastive Learning Enables High Performance P300 Spelling with Minimal Calibration

The P300 speller is a widely adopted brain computer interface (BCI) paradigm that enables hands free character selection based on event-related potentials elicited through an oddball stimulus paradigm. Despite its utility, the system's performance is often constrained by the low signal-to-noise ratio and complex spatiotemporal characteristics of EEG signals, especially when only a limited number of repetitions or labeled samples are available. Moreover, substantial within-session calibration is typically required to achieve reliable decoding before online spelling, posing a major practical barrier. To tackle these challenges, we propose SCL-EEGMixer, a lightweight, end to-end neural architecture that combines a convolutional mixer network with supervised contrastive learning. The model extracts discriminative spatiotemporal representations via the convolutional mixer and enhances learning with a hybrid loss that fuses cross-entropy and supervised contrastive objectives. This design promotes intra-class compactness and inter-class separability, enabling robust learning from scarce labeled data. Extensive evaluations on both a public benchmark and a self-collected dataset demonstrate that SCL-EEGMixer consistently outperforms representative baselines in both binary P300 classification and character recognition tasks under a within-session protocol. Notably, it maintains high accuracy and information transfer rate even when trained with as few as one or two calibration characters, highlighting its potential for reducing within-session calibration burden in P300 spelling.