Subject-Independent Imagined Speech Detection via Cross-Subject Generalization and Calibration

TL;DR

The paper tackles subject-independent imagined-speech decoding from EEG by evaluating cross-subject generalization under strict LOSO and a subject-calibrated LOSO (SC-LOSO) framework. It adopts a cyclic inter-subject training strategy with short per-subject updates and compares it to longer, deeper updates, using the MRF-EEGNet architecture to learn invariant speech-related features. Additionally, it demonstrates that lightweight few-shot calibration on the target subject can substantially close the gap to subject-dependent performance, with notable gains at calibration levels as low as 5–10%. These findings suggest a practical path toward scalable, user-adaptive imagined-speech BCIs that balance generalization and personalization, with future work focusing on larger datasets and augmentation strategies to further enhance cross-subject robustness.

Abstract

Achieving robust generalization across individuals remains a major challenge in electroencephalogram based imagined speech decoding due to substantial variability in neural activity patterns. This study examined how training dynamics and lightweight subject specific adaptation influence cross subject performance in a neural decoding framework. A cyclic inter subject training approach, involving shorter per subject training segments and frequent alternation among subjects, led to modest yet consistent improvements in decoding performance across unseen target data. Furthermore, under the subject calibrated leave one subject out scheme, incorporating only 10 % of the target subjects data for calibration achieved an accuracy of 0.781 and an AUC of 0.801, demonstrating the effectiveness of few shot adaptation. These findings suggest that integrating cyclic training with minimal calibration provides a simple and effective strategy for developing scalable, user adaptive brain computer interface systems that balance generalization and personalization.

Figures (2)

• Figure 1: Overall experimental pipeline comprising the data acquisition, training, and test phases under two evaluation schemes: strict LOSO and subject-calibrated LOSO. In the strict LOSO setup, the model is trained using data from all subjects except $S_{n-1}$and directly evaluated on the unseen subject. In the subject-calibrated LOSO scheme, a small portion of $S_{n-1}$'s data is used for fine-tuning before testing, leading to final classification outputs between speech and idle states.
• Figure 2: Comparison of MCC and AUC scores under the SC-LOSO scheme across different calibration ratios (5 %, 10 %, and 15 %). The blue solid line represents the training configuration of ($N{=}5$ and $R{=}6$), and the orange dashed line corresponds to ($N{=}5$ and $R{=}6$).