Non-Invasive Brain-Computer Interfaces: EEG Architectures, Deep-Learning Decoders and Emerging Trends

A brain-computer interface (BCI) gives the brain a communication channel that does not pass throughperipheral nerves and muscles. Among all sensing technologies, non-invasive electroencephalography (EEG) isthe most practical: it is safe, portable, inexpensive and has millisecond temporal resolution. Its weaknesses are alow signal-to-noise ratio, non-stationarity and large differences between people and sessions. This paper presentsa structured study of non-invasive, EEG-based BCIs. It covers the fundamentals (taxonomy, electrode placement,brain rhythms, paradigms, system architecture) and the signal-processing chain from classical machine learningto end-to-end deep learning. It then analyses in detail three recent IEEE studies that represent different parts ofthe field: ACNN-STADP, an adaptive CNN with dynamic pathways and spatio-temporal attention formotor-imagery classification (90.77% on BCI Competition IV-2a); EEG-Deformer, a dense convolutionalTransformer for attention, fatigue and workload decoding (82.72%, 79.32% and 73.18% leave-one-subject-outaccuracy); and a continuous-trajectory decoder that reconstructs the movement of an observed robotic arm fromlow-frequency EEG. A systematic review of 68 studies (2024-2025) frames these results. Eleven architecturediagrams and five result charts are provided. We close with the open problems of generalisation, data scarcity,interpretability, reproducibility and ethics, and with a conceptual architecture that combines edge inference,federated learning and closed-loop adaptation.