From beta thresholds to learned control: machine learning in adaptive deep brain stimulation

Deep brain stimulation (DBS) has transformed the treatment of Parkinson's disease by modulating pathological brain circuits, yet conventional continuous stimulation represents a one-size-fits-all approach that cannot adapt to fluctuating symptoms or individual variations in disease expression. The February 2025 FDA approval of adaptive DBS (aDBS) with beta-band sensing marks a pivotal regulatory milestone, supported by the ADAPT-PD clinical trial program and now entering routine neurosurgical practice. However, single-biomarker threshold-based control, while clinically validated, addresses only part of the symptom landscape and relies on hand-set parameters. Machine learning (ML) offers a bridge to next-generation closed-loop systems that integrate multi-modal biomarkers and learn individualized control policies. This narrative review synthesizes the physiological rationale for aDBS, the landmark trials and regulatory pathway that brought it to clinical approval, the diverse ML architectures being developed for biomarker discovery and policy optimization, and the neurosurgical dimensions that determine whether these systems can succeed in practice. We highlight that successful translation of ML-driven aDBS depends critically on surgical factors including lead placement precision, signal quality and patient selection, as well as on overcoming barriers in generalisability, safety validation, explainability and equity. The neurosurgeon sits at the center of this translational pipeline, and understanding both the promise and the practical limits of ML-augmented neuromodulation is essential for informed patient care and research design.