Scientists Debate How AI Should Pick the Best Brain Stimulation Contacts in Parkinson’s Disease

Deep brain stimulation has transformed the lives of thousands of people with Parkinson’s disease, but the process of fine-tuning it remains a slow, labor-intensive art. Now, a scientific exchange published in npj Parkinson’s Disease has brought into the open one of the field’s most pressing questions: how should predictive algorithms that promise to identify the optimal stimulation contact be tested, adapted, and ultimately trusted by clinicians at the bedside? In a formal reply to an external evaluation of their work, a team of researchers led by Marjolein Muller and M. Fiorella Contarino of Leiden University Medical Center and the Haga Teaching Hospital in the Netherlands, together with Ioannis U. Isaias and Eduardo M. Moraud, has defended the methodology behind their machine learning algorithm while acknowledging the very real challenges of moving such tools from the research lab into routine clinical programming.

At the heart of the debate is an algorithm designed to predict which of the four contacts on a deep brain stimulation electrode is likely to deliver the greatest therapeutic benefit. The original study, published by Muller and colleagues in 2025, showed that signals recorded from the electrode itself, known as local field potentials, could be used to rank candidate stimulation contacts online, potentially sparing patients and clinicians the trial-and-error sessions that currently dominate post-operative programming. The new Matters Arising article responds to a team led by van der Weide, who took that algorithm, implemented it on an independent dataset from their own center, and reported on its practical utility and the variability they encountered when raters interpreted the underlying electrophysiological signals.

The Leiden-led authors open their reply with genuine appreciation. They describe the external evaluation as highly relevant and timely, noting that it provides valuable insight into real-world deep brain stimulation programming and highlights important considerations for interpreting local field potential based predictions. Two of those considerations stand out in particular: inter-rater variability, meaning the differences that arise when different human experts judge the same signal, and the influence of electrode placement, which can shift what the recording electrode actually captures from the surrounding neural tissue. Both factors, the authors note, matter enormously when a prediction tool is expected to perform consistently across patients, centers, and continents.

But the reply also raises a series of pointed methodological objections. The most fundamental concerns what is actually being predicted. In their original study, Muller and colleagues deliberately designed the algorithm to predict the top two contacts per hemisphere rather than a single best contact. This choice reflects two clinical realities. First, anatomically adjacent contacts may be similarly effective, and second, clinical evaluation of stimulation-induced side effects may not be fully captured by the local field potential signals on which the prediction is based. The evaluating team, by contrast, adapted the algorithm to predict a single optimal contact, and the responding authors argue that it remains unclear how that single best contact was determined. That methodological detail, they insist, is critical when interpreting the reported inter-rater variability and ensuring the reproducibility of the adapted algorithm.

The technical crux of the disagreement lies in how the beta-band signal should be read. Deep brain stimulation targeting the subthalamic region has long been associated with beta-band oscillations, rhythmic electrical activity in roughly the 13 to 30 hertz range that often peaks near the most therapeutically effective contact. A seemingly natural approach is to have a human rater identify the most prominent beta peak and select the corresponding contact. But the responding authors argue that this introduces subjectivity. In their experience, selecting channels based on the absolute highest beta-band value, regardless of whether a clear peak is present, is far more objective and consistent, typically producing high agreement among raters because no subjective judgment about what counts as a genuine peak is required. Crucially, they demonstrated in their original paper that the algorithm functions even when no distinct beta peak is present, suggesting that the aperiodic component of the local field potential, the non-oscillatory background of the signal, may contribute meaningful information about contact location.

The reply also raises questions about the clinical context of the validation dataset. The evaluating authors noted that their center predominantly selects the middle two contacts, the ventromedial and dorsomedial contacts, for therapy, a practice that aligns with their anatomical implantation strategy. The responding authors point out that while this local heuristic performs well at that center, other centers may employ different strategies or register wider variation in their targets. An algorithm whose job is to generalize, they argue, is probably best assessed on datasets with broader variability in contact selection rather than on a dataset closely resembling the original cohort. Such validation, they contend, would more comprehensively demonstrate the generalizability and clinical impact of predictive local field potential based methods.

Another gap concerns how predictive accuracy was measured in the adaptation. The responding authors note that the current study does not report how the accuracy of the adapted decision tree algorithm was evaluated, and they warn that the timing and conditions of comparison matter. Comparing local field potentials recorded during monopolar review with contacts chosen during that same monopolar review session, or instead with contacts chosen at a six-month follow-up, could lead to different outcomes. Post-operative programming is complicated by the stun effect, a temporary improvement or fluctuation in symptoms immediately after electrode implantation that can distort early clinical assessments. The choice of reference standard, and when it is measured, can therefore determine whether an algorithm appears accurate or inaccurate.

Rather than simply defending their work, the authors lay out a concrete roadmap for future validation. They call for a structured, multi-center approach that ensures all possible contact choices are represented in the validation data. Local field potential recordings and clinical choices, they argue, should share the same timing and clinical state, ideally recorded after the stun effect period, with the patient off medication and off stimulation. For evaluation, they suggest assessing the agreement between the primary clinical choice and the top two algorithmic recommendations, which mirrors the design of the original algorithm. If researchers do adapt the tool to predict a single contact, they should use objective selection measures to prevent inter-rater differences from contaminating the results. And in a move that gets at the deepest conceptual question in the field, they propose that studies add an objective evaluation of the contact producing the best therapeutic effect, which may differ from the clinically chosen contact, since clinical selection also weighs side effects and clinician preference.

Perhaps the most important point of agreement in the exchange concerns the proper role of algorithms in medicine. Both sides accept that predictive tools should complement rather than replace clinical judgment. The responding authors emphasize that their algorithm is intended as a decision-support tool to reduce the programming burden, not as a standalone determinant of contact selection, and they note that the external findings reinforce the importance of combining electrophysiological biomarkers with imaging and clinical evaluation. This framing reflects a broader consensus in neuromodulation research: electrophysiological signals such as beta-band power are informative biomarkers of electrode position, but they are one stream of evidence among several, and the clinician in the room must ultimately integrate physiology, anatomy, patient-reported effects, and observed side effects.

The exchange, published as an open-access Matters Arising article on 27 August 2026, is a window into how a young technology matures. Deep brain stimulation programming currently depends heavily on expert time, and any tool that reliably narrows the search for the right contact could shorten clinic visits, reduce patient discomfort, and make the therapy more accessible in centers without specialist programming expertise. But the path from a promising algorithm to a trusted clinical instrument runs through exactly the kind of scrutiny this reply exemplifies: independent testing on external data, transparent methodology, objective outcome measures, and validation across diverse centers with different implantation strategies and clinical cultures. The authors close by expressing their hope that clarifying the methodology for applying their protocol, for deriving single-contact selection, and for evaluating predictive accuracy will help interpret the observations, and that expanding validation across diverse centers will advance the integration of predictive tools into deep brain stimulation workflows. The conversation, they make clear, is not over, and continued discussion and collaborative efforts are precisely what will bridge predictive algorithms and clinical practice.

Subject of Research: AI-guided selection of optimal deep brain stimulation contacts using local field potentials in Parkinson's disease

Article Title: Reply to: Bridging predictive algorithms and clinical practice in DBS contact selection

Article References: Muller, M., Isaias, I. U., Moraud, E. M., & Contarino, M. F. (2026). Reply to: Bridging predictive algorithms and clinical practice in DBS contact selection. npj Parkinson's Disease, 12(1), Article 205. https://doi.org/10.1038/s41531-026-01534-1

Image Credits: AI Generated

DOI: 10.1038/s41531-026-01534-1

Keywords: deep brain stimulation, Parkinson's disease, local field potentials, predictive algorithms, contact selection, beta-band oscillations, decision-support tools, neuromodulation, algorithm validation, npj Parkinson's Disease, subthalamic stimulation, clinical programming