Brain–Computer Interfaces (BCIs) offer promising communication pathways be tween the human brain and external devices, yet their deployment in real-world settings remains limited by noise susceptibility, subject variability, and practical constraints. This thesis addresses these challenges using Rapid Serial Visual Presentation (RSVP)-based P300 paradigms, with a focus on enhancing robustness, generalisability, and real-world applicability. To begin, I systematically investigated the impact of real-world behavioral artifacts, body movement, head movement, and talking, on EEG signal quality and classification performance. Results reveal that such noise substantially degrades system accuracy and often leads to significant data loss during artifact rejection, underscoring the limitations of traditional denoising methods and motivating adaptive solutions. Next, I explored subject-independent classification using various deep learning models. Transformer-based architectures, especially when combined with EEG-specific convolutional neural networks (i.e., EEGNet), demonstrate superior generalisation under the Leave-One-Subject-Out (LOSO) framework. In particular, the hybrid Transformer with a multiband input strategy consistently outperformed other architectures, highlighting the effectiveness of integrating temporal attention mechanisms with frequency-specific spatial filtering. This underscores robust generalisation across subjects, a crucial step toward the practical and scalable deployment of BCIs. While subject-independent models offer strong baseline performance, their real world adaptability is further enhanced through lightweight subject-specific calibration, where incorporating a small amount of personalised data yields noticeable performance gains. To reduce system complexity without compromising accuracy, I identified a 16-channel EEG configuration that retains over 95% of baseline performance, with 10–12 channel setups also viable for certain applications. However, channel selection remains inherently subject- and task-dependent, necessitating application-specific validation. Additionally, I evaluated the impact of display modality, finding that both head-mounted displays (HMDs) and traditional monitors support comparable classification performance. With their added portability, HMDs represent a promising avenue for mobile and ecologically valid BCI applications. Practical recommendations are offered on model choices, data volume, channel optimisation, and calibration, aiming to bridge the gap between lab-based BCI research and real-world deployment. While the work primarily focuses on RSVP-based P300 detection, the insights contribute broadly to the development of more noise-tolerant and effective BCI systems.