Reinforcement Learning-Based Focality Optimization for Multi-Electrode Temporal Interference Stimulation

OBJECTIVE: Multi-electrode Temporal Interference Stimulation (TIS) is a promising noninvasive technology for deep brain stimulation, but its clinical translation is hindered by the challenge of optimizing its high-dimensional and hybrid parameter space. Existing algorithms are often limited by high computational cost or a lack of practical flexibility. This paper introduces and validates a novel framework to overcome these barriers. METHODS: We developed a novel reinforcement learning (RL) framework that simultaneously optimizes the discrete positions and continuous intensities of stimulation electrodes. We systematically evaluated the framework's performance using six realistic finite element head models and benchmarked it against genetic algorithms (GA) and unsupervised neural networks (USNN). RESULTS: Our RL-based approach significantly enhanced stimulation focality, outperforming GA-based optimization. While achieving performance comparable to USNN, our framework offered the critical advantage of explicit control over the number of active electrodes. Furthermore, our findings reveal that focality improves with an increasing number of electrodes up to 16, with diminishing returns beyond this point. CONCLUSION: This work establishes a powerful and flexible computational paradigm for TIS optimization. SIGNIFICANCE: Our framework provides a crucial guideline for optimal system design and overcomes key limitations of previous methods. This paves the way for more precise and clinically robust noninvasive neuromodulation, accelerating the clinical translation of TIS.