Brain-computer interfaces (BCIs) are becoming closed-loop medical cyber-physical systems in which neural activity is decoded, transformed into a control or therapeutic signal, and returned to the patient through speech synthesis, prosthetic actuation, neurofeedback, electrical stimulation, or clinician-supervised rehabilitation. This survey synthesizes machine-learning trends from January 2024 through May 2026 with emphasis on medical and psychiatric translation. The dominant methodological transition is from small supervised decoders toward multimodal representation learning, self-supervised pretraining, domain adaptation, latent-dynamics stabilization, uncertainty-aware control, and clinically constrained online learning. The biomedical use cases are heterogeneous: non-invasive EEG and fNIRS support neurorehabilitation, cognitive-state monitoring, depression phenotyping, and neurofeedback; ECoG, sEEG, and intracortical arrays support high-performance communication and dexterous motor neuroprostheses; closed-loop DBS and responsive neurostimulation motivate psychiatric biomarkers for treatment-resistant depression, obsessive-compulsive disorder, post-traumatic stress disorder, addiction, and affective dysregulation. A unifying technical problem is generalization under severe distribution shift caused by subject variability, electrode impedance, medication state, vigilance, neuropsychiatric symptom fluctuation, and tissue-electrode nonstationarity.
Keywords: Foundation Models, Adaptive Decoding, Neuroprostheses, Closed-Loop Therapeutics, and Neural Data Governance.
Citation: Easttom, C. (2026). Machine Learning and Brain-Computer Interfaces for Medical and Psychiatric Translation, 2024-2026. J Psychol Neurosci; 8(4):1-11. DOI : https://doi.org/10.47485/2693-2490.1172