Introduction: Advancements in brain-computer interfaces (BCIs) have improved real-time neural signal decoding, enabling adaptive closed-loop neuromodulation. These systems dynamically adjust stimulation parameters based on neural biomarkers, enhancing treatment precision and adaptability. However, existing neuromodulation frameworks often depend on high-power computational platforms, limiting their feasibility for portable, real-time applications. Methods: We propose RONDO (Recursive Online Neural DecOding), a resource-efficient neural decoding framework that employs dynamic updating schemes in online learning with recurrent neural networks (RNNs). RONDO supports simple RNNs, long short-term memory networks, and gated recurrent units, allowing flexible adaptation to different signal type, accuracy, and real-time constraints. Results: Experimental results show that RONDO's adaptive model updating improves neural decoding accuracy by 35% to 45% compared to offline learning. Additionally, RONDO operates within real-time constraints of neuroimaging devices without requiring cloud-based or high-performance computing. Its dynamic updating scheme ensures high accuracy with minimal updates, improving energy efficiency and robustness in resource-limited settings. Conclusions: RONDO presents a scalable, adaptive, and energy-efficient solution for real-time closed-loop neuromodulation, eliminating reliance on cloud computing. Its flexibility makes it a promising tool for clinical and research applications, advancing personalized neurostimulation and adaptive BCIs.
Lee et al. (Fri,) studied this question.