Answers are grounded in this paper and cite their sources.
Background: The advent of deep learning-based EEG BCIs in stroke rehabilitation applications is abundant; however, their clinical translation is still at a standstill. Accuracy on benchmark datasets of healthy subjects often conceals important shortcomings in terms of generalizability, physiological plausibility, and real-time safety standards. To date, there is no standard methodology to evaluate these BCIs in accordance with regulatory requirements (FDA GMLP, EU AI Act) and physiological limitations (Hebbian plasticity window). Methods: A forensic audit of 24 studies of state-of-the-art (SOTA) techniques that appeared in high-impact publications between 2018 and 2025 was performed. Each paper was scored for six non-compensatory criteria for Translational Safety Thresholds (TST): subject independent validation, physiological XAI, feedback latency (<300 ms), closed-loop control, standardized clinical outcomes (FMA-UE/WMFT), and proper sample size (N>20). Four categories of studies were created based on technical safety and clinical translatability. Statistical significance was determined by Chi-Square Goodness-of-Fit tests. Results: The audit showed that there was a system failure of translational safety. None of the studies (0.0%) demonstrated high technical rigor along with low clinical readiness (Quadrant 1), while only one study (4.2%) satisfied both the technical and clinical criteria (Quadrant 2), but it failed to be validated in stroke patients. Most of them (75.0%) belonged to Quadrant 4 (Low Technical/Low Clinical Readiness). Some of the critical failure rates were 95.8% for splitting at the subject level (95% CI: 78.9–99.9%), 91.7% for physiological XAI, and 87.5% for latency compliance. The distribution was significantly different from the expected uniform distribution (χ 2 (3) = 34.33, p < 0.0001). Conclusion: Recent SOTA EEG-BCI studies suffer from a highly dangerous mismatch between the engineering performance measures and patient safety standards. The almost complete lack of compliance with regulatory and neurophysiological standards indicates that the majority of models described in the literature are not suitable for practical applications. It is proposed that the implementation of Translational Safety Thresholds be made compulsory by academic journals and regulatory bodies in order to close the gap between them.
No takes yet. Share an insight, caveat, or question.
Waqas Ahmad (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: