Engineering principles of Machine Learning-Driven Adaptive Deep Brain Stimulation: A narrative review

Deep brain stimulation (DBS) is an established therapy for the motor symptoms of Parkinson's disease (PD), but conventional systems deliver stimulation continuously and at fixed parameters, independent of the patient's fluctuating neurological state. Closed-loop or adaptive DBS (A-DBS) addresses this limitation by sensing neural activity and adjusting stimulation in real time. This paper reviews the engineering foundations of closed-loop DBS and the role of machine learning (ML) in personalizing therapy. This review explains Parkinson's disease as a control problem, summarizes the neural and imaging biomarkers used as feedback signals, particularly subthalamic beta-band local field potentials (LFPs) and diffusion/functional imaging correlates of stimulation response, and examines supervised and reinforcement learning approaches to real-time parameter selection. We further discuss clinical evidence supporting A-DBS, the regulatory and hardware constraints specific to implantable devices, and open challenges related to personalization, safety, and generalization across patients. The review concludes that the integration of neural biomarkers, wearable sensing, and machine learning represents a methodological template that is likely to extend beyond Parkinson's disease to other conditions treated by invasive neuromodulation.