Introduction:
Deep Brain Stimulation (DBS) is a major breakthrough at the intersection of medicine and technology and has revolutionised the treatment of severe neurological, psychiatric, and movement disorders. It has significantly evolved through the years, yet the conventional open-loop paradigms remain constrained by inconsistent symptom control and the inefficient use of energy. The paradigm shift comes from switching to adaptive closed-loop DBS enabled by artificial intelligence and machine learning capable of adjustable real-time stimulation based on neural and behavioural biomarkers. This systematic comparative review evaluates ML-based closed-loop DBS systems across disorders such as Parkinson’s disease, chronic pain, dystonia, essential tremor, epilepsy, and psychiatric conditions.
Methods:
Thirty studies were analysed based on five key parameters: clinical indication and DBS target, biomarker modality, ML approach, clinical outcomes, and energy efficiency. The strategies include static optimisation (pre-surgical planning) and dynamic optimisation (real-time adjustment) to identify the most effective computational approaches used in functional neurosurgery.
Results:
Closed-loop strategies ranged from threshold-based adaptive control to reinforcement learning, CNN classifiers, and MPC. Static optimisation demonstrated statistically superior performance over programming, including improved target coverage, reduced electric field leakage, an increased therapeutic window, and reduced programming time. For dynamic optimisation, outcomes depend on biomarker fidelity, particularly local field potential beta oscillations (13–30 Hz). For multi-parameter optimisation tasks, reinforcement learning (RL) offered an effective model-free approach, autonomously learning stimulation policies and improving focality. Model Predictive Control (MPC) achieved more than 20 percentage reduction in tracking error and provided superior real-time regulation.
Conclusion:
Across studies, adaptive DBS improved symptom control. Stimulation time and energy consumption were reduced by up to 50 per cent. Reinforcement learning and MPC showed promise in epilepsy, while amplitude-adaptive cortical control was effective in ET. Network-guided targeting and sleep-state adaptation expanded potential beyond PD. Overall, AI-driven closed-loop DBS enhances precision, clinical efficacy, and energy efficiency.
