The commentary, authored by Anneke van der Weide, Y. Wiggerts, D. Hubers and colleagues, responds to work in the growing field of data-driven contact selection. In deep brain stimulation of the subthalamic nucleus, the surgical target most commonly used for Parkinson’s disease, a quadripolar or directional lead is implanted with multiple metal contacts along its shaft. Postoperatively, the clinical team must choose which contact or combination of contacts to activate, a decision traditionally made through painstaking trial and error during programming sessions, guided by the patient’s symptom response and the emergence of side effects such as contralateral motor symptoms, speech difficulties, dysarthria, autonomic changes or mood shifts.

For decades this selection process has been as much art as science. The volume of tissue activated, the spatial relationship between each contact and the patient’s individually delineated motor subregion of the subthalamic nucleus, and the distance from fiber tracts such as the hyperdirect and corticospinal pathways all influence the therapeutic window. Modern approaches therefore combine structural and diffusion-weighted magnetic resonance imaging, sometimes with intraoperative or postoperative computed tomography to localize the lead, to generate patient-specific models of electric field spread. Machine learning classifiers trained on retrospective programming outcomes have been proposed to rank contacts by their probability of producing benefit with minimal adverse effects, and several groups have reported accuracies that appear to approach or exceed inter-rater agreement among expert programmers.

The authors of the commentary do not dispute the promise of these methods. Rather, they argue that the reported performance of predictive models can obscure fundamental obstacles to clinical translation. One central concern is the definition of the ground truth itself. When an algorithm is trained on the contact that a programmer ultimately chose, or on the contact associated with the best clinical outcome at follow-up, the labels inherit all the inconsistencies of real-world practice. Programming decisions vary across centers, across devices and across the experience levels of the clinicians involved. Follow-up intervals differ, medication states confound motor assessments, and the contact that is optimal at three months may not be optimal at one year as the disease evolves, as tissue reaction around the lead matures and as stimulation parameters are adjusted.

A second concern involves the validation standard that new algorithms are held to. Comparing a model’s prediction against a single human expert’s choice can make the model look artificially good or artificially bad depending on that expert’s idiosyncrasies. Comparing against consensus ratings from multiple blinded experts is more rigorous but raises its own question: if experts themselves disagree, what exactly is the model being trained to reproduce? The commentary emphasizes that metrics such as accuracy, area under the receiver operating characteristic curve or F1 score, while useful for benchmarking, do not by themselves establish that a model will change clinical decisions or improve patient outcomes. What matters, the authors contend, is prospective demonstration that algorithm-assisted selection shortens the time to stable therapy, reduces the number of programming visits, or measurably improves motor and quality-of-life outcomes compared with standard care.

The commentary also addresses the practical heterogeneity of the implanted hardware, a problem that is easy to underestimate. Different manufacturers offer leads with different contact geometries, contact spacings, degrees of directional segmentation and current-shaping capabilities, and the field is moving toward adaptive and closed-loop devices that sense local field potentials and adjust delivery in real time. A predictive model trained exclusively on ring-mode contacts from one device generation may not transfer cleanly to segmented directional leads from another, where the effective anatomical coverage of a contact changes with rotation and with the use of multiple simultaneous current fractions. Imaging pipelines add further variability: direct lead localization from postoperative CT, registration errors between CT and preoperative MRI, and the choice of atlas or patient-specific segmentation of the subthalamic nucleus can each shift the computed relationship between a contact and the motor territory by a fraction of a millimeter to a millimeter or more, which is on the order of the distances that differentiate a good contact from a poor one.

Data quantity and quality present a related bottleneck. High-quality labeled datasets are scarce because they require patients who have undergone detailed imaging, careful postoperative programming and structured long-term follow-up. Multicenter pooling is the obvious solution, but pooling introduces harmonization problems across scanner platforms, surgical techniques and outcome measures. The commentary suggests that the field needs agreed reporting standards, shared benchmarks and ideally openly available datasets with common definitions of what constitutes a successful contact selection, so that competing algorithms can be compared on equal footing rather than on private, incomparable cohorts.

Perhaps the most clinically resonant part of the argument concerns workflow integration. Even a well-validated model provides only a ranked list of candidate contacts with associated confidence levels. The clinician in the programming room must reconcile that ranking with information the model may not see: the patient’s reported sensations during test stimulation, medication timing, cognitive and psychiatric history, the patient’s priorities between mobility and speech, and the practical constraints of the device’s battery and safety limits. The authors argue that tools framed as decision support rather than decision replacement are far more likely to be adopted. If the algorithm presents its top candidates together with the anatomical reasoning, for example the estimated overlap with the motor territory and proximity to internal capsule fibers, the clinician can interrogate the recommendation, override it when the clinical picture demands it, and learn from the interaction. A black-box ranking delivered without explanation invites either blind trust or justified skepticism, and neither serves the patient.

The commentary also raises the question of when in the therapeutic trajectory such tools should be applied. Early programming, in the first weeks after lead implantation, is arguably where prediction offers the greatest payoff, because the therapeutic window is often narrow and empirical exploration is slowest and most burdensome for the patient. But it is also when the perioperative state, including microlesion effects from electrode insertion and residual swelling, can distort the relationship between anatomy and stimulation response. Later, after months of chronic stimulation, the picture stabilizes but many patients have already reached a satisfactory configuration, reducing the marginal value of prediction. Striking the right moment for algorithmic input, and designing studies that measure outcomes at that moment, is part of the bridging work the authors call for.

Equity and generalizability form another layer of the argument. If training cohorts are drawn disproportionately from high-volume academic centers in a small number of countries, the resulting models may encode narrow surgical and programming cultures. Contact selection reflects local conventions, such as preferred current settings, typical amplitudes and the aggressiveness of medication reduction, all of which differ internationally. A model that silently learns those conventions will perform differently when deployed elsewhere. The commentary implies that algorithm developers should document the provenance of their training data as carefully as they document their model architecture, and that prospective multi-center trials are the only way to establish that performance generalizes beyond the development environment.

None of these criticisms amount to a rejection of computational contact selection. On the contrary, the commentary’s tone is constructive: the field has generated genuinely exciting predictive tools, and the authors’ argument is that the next phase of progress depends on methodological discipline rather than additional architectural novelty. They call for standardized outcome definitions, blinded expert consensus labels, external validation on fully independent cohorts, transparent reporting of failure cases, and ultimately randomized prospective studies in which algorithm-guided programming is compared with conventional care on patient-centered endpoints such as time to therapeutic benefit, number of programming sessions and quality-of-life measures.

For patients, the stakes are concrete. Every programming visit represents time away from work and family, and every suboptimal contact configuration can mean months of avoidable tremor, rigidity, slowness or stimulation-induced side effects. An algorithm that reliably narrowed the search from four or more candidate contacts to one or two could meaningfully compress the journey to stable therapy. The commentary’s authors make clear that they believe this goal is achievable, but only if developers and clinicians build the bridge together, with algorithms designed around the realities of the programming room rather than the metrics of the machine learning benchmark.

As deep brain stimulation expands beyond Parkinson’s disease into dystonia, essential tremor, epilepsy, obsessive-compulsive disorder and treatment-resistant depression, and as sensing-enabled adaptive devices become standard, the number of degrees of freedom in stimulation delivery will only grow. Manual exploration of that expanded parameter space will become progressively less feasible, which makes trustworthy predictive tools not a luxury but a necessity. The contribution of this Matters Arising piece is to define, with clinical precision, what “trustworthy” will have to mean: models validated against consensus truth, tested prospectively, explained transparently and integrated respectfully into the judgment of the clinicians who remain accountable for the person attached to the lead.