Amyotrophic Lateral Sclerosis (ALS) is a progressive neurological disorder with no fully effective treatment currently available. In this study, a novel tensor-based feature reduction method, Higher Order Spectral Regression Discriminant Analysis (HOSRDA), is proposed to enhance Brain-Computer Interface (BCI) performance in individuals with ALS. HOSRDA extends the principles of Spectral Regression Discriminant Analysis (SRDA) to handle multi-dimensional EEG data, effectively addressing the challenges of high-dimensionality and ill-conditioned scatter matrices in the analysis of P300 speller data. This method reduces the dimensionality of EEG signals while preserving class separability, enabling efficient classification using Linear Discriminant Analysis (LDA). Furthermore, HOSRDA leverages a regression framework to address the computationally expensive eigenvalue decomposition of scatter matrices, a challenge faced by traditional methods like HODA, significantly improving computational efficiency. Experiments conducted on EEG data from five ALS patients show that the HOSRDA-LDA model achieves an average character detection accuracy of 84.04%, demonstrating its potential for real-time BCI applications. Compared to traditional methods such as LDA without feature reduction and Support Vector Machine (SVM), HOSRDA outperforms in terms of classification accuracy and computational efficiency, with significantly reduced training times. The HOSRDA method converges in an average of 2.04 seconds over three repetitions, making it highly suitable for online BCI systems. These findings suggest that HOSRDA can improve the accessibility and usability of BCIs for ALS patients, with potential applications extending to broader clinical and real-world settings, without the need for time-consuming training sessions or considering factors like literacy.