Local field potentials survey to guide DBS programming in Parkinson’s disease: a clinical-neurophysiological longitudinal study

Abstract

Deep Brain Stimulation (DBS) programming in Parkinson’s disease relies on clinical evaluation, yet beta-band activity in local field potentials (LFPs) may offer objective guidance. We evaluated LFP-guided contact selection against clinical programming (CP) at initial monopolar review and long-term follow-up. Bipolar LFPs were recorded in patients with sensing-enabled DBS systems, where central levels were clinically chosen. Three methods based on beta peak amplitude were tested: Broad-Bipolar (non-adjacent rings), Narrow-Bipolar (contiguous rings), and Segment-Bipolar (horizontal segments). Performance was assessed using correlation, agreement, and selection similarity with CP and compared to random selection. Narrow-Bipolar showed the strongest correlation and agreement with CP at both timepoints, outperforming other methods and random selection. It showed the greatest selection stability over time and the highest prediction. Its performance was comparable to imaging-guided programming and did not differ between STN and GPi. These results support Narrow-Bipolar as a valid, easy-to-perform beta-based method for guiding DBS programming.

Introduction

Deep brain stimulation (DBS) of the subthalamic nucleus (STN) and globus pallidus internus (GPi) is an effective therapeutic option for Parkinson’s disease (PD) patients whose symptoms are not adequately controlled by pharmacological therapy^1,2.

A critical step for DBS optimization and tailoring is programming, which currently relies on clinical evaluation of motor symptoms and adverse effects during the titration of stimulation parameters (i.e., monopolar review)^3,4. Although this conventional clinical programming (CP) is considered the gold standard, it is time- and resource-consuming, prone to subjective errors, and requires specialized training, making it largely restricted to experienced centers^3,5–7. Additionally, CP often follows heterogeneous in-house procedural pipelines and is not standardized across centers⁸.

A possible strategy to overcome these limitations is using imaging-guided programming^7,9. However, the identification of the dorsolateral (sensorimotor) area of target nuclei (STN and GPi) with conventional Magnetic Resonance Imaging (MRI) remains challenging and requires advanced image post-processing^9,10. Several factors can affect lead localization accuracy and the overall reliability of these approaches, including movement-related artifacts during MRI acquisition, susceptibility artifacts around electrodes, limited spatial resolution and distortions of low-field (i.e., 1.5 and 3T) MRI in deep brain structures, and potential errors introduced by co-registration pipelines¹¹. Furthermore, there is evidence that the structurally defined sensorimotor regions do not necessarily coincide with the functional domain in the individual patient^12–14.

Local Field Potential (LFP) recordings from the DBS target structure in PD typically show hypersynchronized activity in the beta frequency band (13–30 Hz), which correlates with motor symptoms and is suppressed by L-Dopa and DBS^12,15–17. Consequently, beta oscillations evolved as a promising biomarker specifically for real-time applications like adaptive closed-loop DBS^18–20. Concerning DBS programming, previous studies using intraoperative and early postoperative LFP recordings showed that lead contacts exhibiting prominent beta activity frequently showed the largest therapeutic window and may therefore be used for chronic stimulation^21–25. Sensing-enabled DBS systems, allowing fast neurophysiological assessments in chronically implanted patients, can now facilitate LFP-based contact selection^26,27. To date, only a few studies have investigated the efficacy of beta-guided programming in PD^14,28,29, with one randomized controlled trial demonstrating similar motor outcomes to conventional CP and imaging-guided programming³⁰. However, these studies did not include both short- and long-term neurophysiological assessments. Also, the method used to identify the best stimulating contact remains a key challenge in implementing LFP-guided programming in clinical practice.

Here, we present a longitudinal study to assess and compare three different neurophysiological methods in selecting optimal stimulation contact between central ring electrode levels, applicable in real-time without signal post-processing, with conventional clinical programming (CP) in the STN and GPi. This selection may halve the contacts needed for clinical testing by determining whether beta sources fall more ventrally or dorsally across the electrode. The overall assessment period spans from initial programming (IP) to long-term clinical follow-ups (cFU) (i.e., after determining the optimal chronic stimulation setting). The study objectives were to (i) evaluate the performance of different LFP-based methods relative to CP (gold standard) using three complementary evaluation metrics (correlation strength, agreement level, and selection similarity) and validate it against a random selection (RS) model; (ii) assess the stability of contact selection and beta activity over time; (iii) determine the ability of each method to predict the clinical stimulating contact at follow-up, and (iv) to compare the best-performing neurophysiological approach against imaging-guided programming.

Results

Performance of neurophysiological methods at initial programming

At IP, the phi (φ) correlation coefficient and quadratic weighted kappa (QWK) analyses demonstrated varying levels of association strength and agreement levels in contact selection between the different LFP-based methods and clinical-based programming (Fig. 1). The Narrow-Bipolar method showed very strong correlation (ϕ = 0.71) and substantial agreement (QWK = 0.65, 95% CI: 0.40, 0.87, p < 0.001) with monopolar review. Segment-Bipolar showed a strong correlation (ϕ = 0.56) and moderate agreement (QWK = 0.54; 95% CI: 0.24, 0.76, p = 0.003), while Broad-Bipolar did not correlate (ϕ = 0.01) and had only slight agreement (QWK = 0.11; 95% CI: −0.25, 0.41, p = 0.38) with monopolar review. When assessing the observed mean similarity between the bootstrap distribution of levels selected by LFP-based strategies and monopolar review, we observed that Narrow-Bipolar had higher mean similarity (0.9) than Segment-Bipolar (0.86) and Broad-Bipolar (0.75). When testing the difference between mean similarity of the LFP-based strategies with CP and a random bootstrap distribution, the Narrow-Bipolar and Segment-Bipolar methods were significantly superior to RS (p < 0.001 for both comparisons), while Broad-Bipolar (p = 0.17) did not statistically differ from RS (Fig. 2 and Table 1). Performances of LFP-based methods at IP were similar when selecting beta peaks from raw and post-processed data (see Supplementary information). Finally, no significant differences were observed in the distribution of levels matching with CP when considering peaks in the low- or high-beta sub-bands (Narrow-Bipolar: low-beta 15/22, high-beta 11/14, p = 0.7; Segment-Bipolar: low-beta 13/22, high-beta 9/14, p = 1.00; Broad-Bipolar: low-beta 12/22, high-beta 7/14, p = 0.73) and between STN and GPi (Narrow-Bipolar: STN 12/22, GPi 7/14, p = 0.70; Segment-Bipolar: STN 13/22, GPi 9/14, p = 1.00; Broad-Bipolar: STN 12/22, GPi 7/14, p = 1.00).

Fig. 1. Correlation and agreement between beta-based contact selection methods and clinical programming.

Phi correlation coefficient (A) and Quadratic Weighted Kappa (QWK, B) between LFP-based methods and clinical selection at initial programming (IP), last clinical follow-up (cFU). The figure also includes the same measures computed between contacts selected using LFP-based methods at IP and those selected clinically at cFU (IP/cFU). * indicates p < 0.05; ** indicates p < 0.01; ** indicates p < 0.001.

Fig. 2. Similarity of LFP-based methods with clinical programming.

A Bootstrap distributions illustrating the mean similarity with clinical programming of each method and random selection (RS) at initial monopolar review. B Bootstrap distributions of mean similarity of each method at initial programming and clinical selection at the last follow-up. Dashed lines represent the mean similarity of each method and random distribution.

Table 1.

Similarity between LFP-based and imaging-based methods and clinical programming and difference from random selection

	Selection method vs. CP at IP		Selection method at IP vs. CP at cFU
	Mean similarity	Bootstrap similarity vs. RS (p value)	Mean similarity	Bootstrap similarity vs. RS (p value)
Broad-Bipolar	0.75	0.2	0.75	0.34
Narrow-Bipolar	0.9	<0.001	0.93	<0.001
Segment-Bipolar	0.86	0.001	0.87	0.02
Imaging-based	0.9	0.006	0.83	0.43
RS	0.7	-	0.67	-

Left columns: mean similarity values of each selection method with clinical programming (CP) at initial programming (IP) and differences between the bootstrap similarity and the random selection (RS). Right columns: mean similarity values of each selection method performed at IP with CP at the last clinical follow-up (cFU), and differences between the bootstrap similarity and RS. Significant values (p < 0.05) are displayed in bold.

Stability of contact selection over time and performance of neurophysiological methods at clinical follow-up

When assessing the stability of clinical contact selection over time, we observed that the level activated at IP corresponded to that chronically used at cFU in 75% of cases (27/36 hemispheres, Fig. 3). Among the different LFP-based methods tested, Narrow-Bipolar demonstrated the highest stability between IP and cFU (88.8%, 32/36 hemispheres). Furthermore, the distribution of contact changes over time was similar to CP (p = 0.56). In contrast, Segment-Bipolar (69.4%, 25/36 hemispheres) and Broad-Bipolar (58.3%, 21/36 hemispheres) showed lower stability over time, and the distribution of contact changes from IP to cFU differed from CP (Broad-Bipolar vs. CP: p < 0.01; Segment-Bipolar vs. CP: p = 0.01). When evaluating the stability of beta activity over time, we found that beta peak amplitude (p = 0.18), frequency (p = 0.40), low-beta mean power (p = 0.07), and high-beta mean power (p = 0.62) did not change between IP and cFU (Fig. 3).

Fig. 3. Longitudinal changes in beta activity.

Upper panel Donut charts represent the stability of contact selection between initial programming (IP) and the last clinical follow-up (cFU) for clinical programming and each LFP-based method. Colored segments indicate the number and percentage of cases (36 in total), in which the same contact was retained at cFU. Lower panel Violin plots show changes in beta activity (amplitude and frequency of the beta peak, and beta power) from IP to cFU. Gray markers and lines indicate a decrease at cFU compared to IP, while green indicates an increase.

The level of correlation and agreement with CP improved for all the LFP-based strategies at cFU compared to IP (Fig. 1). Narrow-Bipolar showed a very strong correlation (ϕ = 0.86) and almost perfect agreement (QWK = 0.81; 95% CI: 0.59, 0.99, p < 0.001), Segment-Bipolar showed very strong correlation (ϕ = 0.80) and substantial agreement (QWK = 0.74; 95% CI: 0.49, 0.93, p < 0.001), while the correlation between Broad-Bipolar and CP was very weak (ϕ = 0.06) and the agreement was slight (QWK = 0.12; 95% CI: −0.23, 0.43, p = 0.37). Finally, the distribution of levels matching with CP was similar when considering peaks in the low- or high-beta sub-bands (Narrow-Bipolar: low-beta 20/24, high-beta 9/12, p = 0.66; Segment-Bipolar: low-beta 19/24, high-beta 9/12, p = 0.54; Broad-Bipolar: low-beta 9/24, high-beta 9/12, p = 0.08) and between STN and GPi (Narrow-Bipolar: STN 17/22, GPi 12/14, p = 0.68; Segment-Bipolar: STN 16/22, GPi 12/14, p = 0.44; Broad-Bipolar: STN 13/22, GPi 5/14, p = 0.30).

Ability of neurophysiological methods to predict contact selection at cFU

Narrow-Bipolar selection at IP showed a very strong correlation (ϕ = 0.79) and substantial agreement (QWK = 0.76; 95% CI: 0.53, 0.94, p < 0.001) with CP at cFU. Segment-Bipolar showed strong correlation (ϕ = 0.50) and moderate agreement (QWK = 0.51; 95% CI: 0.19, 0.76, p = 0.02), while Broad-Bipolar did not correlate (ϕ = 0.05) and had slight agreement (QWK = 0.13; 95% CI: −0.22, 0.42, p = 0.34) with CP at cFU (Fig. 1). When assessing the observed mean similarity between the bootstrap distribution of levels selected by LFP-based strategies at IP and CP at cFU, we found that the Narrow-Bipolar method had a very high similarity to CP (0.95). Segment-Bipolar and Broad-Bipolar methods showed lower similarity levels (0.93 and 0.76, respectively). The comparison between the observed mean similarity of the LFP-based strategies and the random bootstrap distribution disclosed that the Narrow-Bipolar and Segment-Bipolar methods were significantly superior to RS (p < 0.01) (Fig. 2 and Table 1). Overall, these results suggest that the Narrow-Bipolar method had good predictive ability in identifying the level clinically selected after multiple follow-up visits. Performances of LFP-based methods at cFU were comparable when selecting beta peaks from raw and post-processed data (see Supplementary information).

Performance of the imaging-based approach and comparison with the best neurophysiological method

At IP, imaging-based level selection showed a strong correlation (ϕ = 0.64) and substantial agreement (QWK = 0.65; 95% CI: 0.36, 0.86, p < 0.001) with CP. The observed mean similarity between the bootstrap distribution of contacts selected by the imaging-based approach and CP was 0.9, which was significantly greater than the corresponding value for RS (p < 0.01) (Table 1). At cFU, the correlation between imaging-based selection and CP was very strong (ϕ = 0.79) and the agreement substantial (QWK = 0.76; 95% CI: 0.53, 0.94, p < 0.001). The observed mean similarity to CP was 0.83, and, unlike the Narrow-Bipolar method, this value did not differ from that obtained with RS (p = 0.43) (Table 1). When directly comparing the bootstrap distribution of similarity (with CP) between the Narrow-Bipolar and imaging-based methods, we found no differences between the two approaches, both at IP (p = 0.99) and cFU (p = 0.82).

Discussion

This longitudinal clinical-neurophysiological study compares the performance of three different beta-based DBS programming methods for identifying the best stimulation lead level compared to the current gold standard (CP) at both IP and after several follow-up visits (cFU). Broad-Bipolar method compared beta activity recorded between non-adjacent rings, Narrow-Bipolar included adjacent rings, while Segment-Bipolar was based on the comparison of horizontally paired segments. We found that Narrow-Bipolar outperformed the other methods in identifying the clinically selected level at IP and cFU in all metrics tested (correlation strength, agreement, and similarity level) and proved able to predict the level used over time. The performance of Narrow-Bipolar was comparable to an imaging-based method at both time points and did not change when considering the DBS target (STN or GPi). Performance was also unaffected by whether the dominant beta peak fell in the low- or high-beta sub-band. Overall, these findings suggest that LFP fast survey using a Narrow-Bipolar method can reliably guide DBS programming in PD.

The methods for data collection allowed us to exclude several factors potentially affecting our results. First, monopolar review was always performed before neurophysiological recordings, and data were analyzed offline to identify the highest beta peak. Therefore, the clinician performing the initial programming was blind to the “best” LFP-based contact level. Also, the “stun” effect likely impacts LFP recordings by inducing a transient decrease in beta activity in the first weeks after surgery^8,31,32. However, we performed the monopolar review at a time when the “stun” effect was likely subsiding³¹, and our finding of beta power stability over months supports this idea. To minimize possible post-stimulation effects, DBS was turned off 2 min before LFP recordings, as beta-band power generally returns to baseline within this timeframe³³. Importantly, since the distance between recording sites is known to affect signal amplitude, each Method included only contact pairs with the same interelectrode distance (e.g., 1–3 vs. 0–2 and 0–1 vs. 2–3). Similarly, the dimension of the recording surface influences the amplitude of the recorded signal; thus, segments and rings were compared independently.

The first novel finding of our study concerns the observation of different performances of three distinct beta-based methods in identifying the lead level selected at initial clinical programming, as per the standard monopolar review. The Narrow-Bipolar method demonstrated a strong correlation and substantial agreement with CP, Segment-Bipolar showed a strong correlation and substantial agreement, while Broad-Bipolar exhibited a slight agreement. Narrow-Bipolar also reached the highest similarity with CP in contact selection, and it proved superior to a simulated random selection. The reason for the diverging performances among methods likely depends on the different combinations of bipolar recording pairs used to compare the beta peak amplitude. The different spacing of contact pairs considered in each method might affect spatial resolution, resulting in different abilities in discriminating the source of beta oscillations between central levels. Bipolar recordings capture the differences in electric potentials between recording sites. Therefore, the amplitude of the signal will be higher when only one of the two recording locations is close to the source of beta oscillations, and lower when both contacts are either close or distant from it³⁴. In this context, bipolar recordings between central and peripheral levels (Narrow-Bipolar) will provide information on the central contact closer to the source of beta oscillations, while sampling a relatively small target volume, thus retaining spatial resolution. Conversely, recordings from non-adjacent contacts (Broad-Bipolar) will achieve lower spatial resolution by capturing the activity of broader target volumes. Furthermore, since both levels 1 and 2 fall within the volume captured by 1–3 and 0–2 pairs, Broad-Bipolar might specifically have a reduced ability to discriminate the source of beta oscillations between central levels. Notably, while previous studies assumed that the source of beta activity might be localized between recording contact pairs^14,29, this hypothesis was originally based on multichannel EEG studies and may not necessarily be translated into LFP signals. Finally, consistent with the findings of a recent study³⁰, the performance of Segment-Bipolar was good. Indeed, recordings from horizontally paired segments span small target volumes, thus potentially providing an optimal spatial resolution³⁴. The lower performance than Narrow-Bipolar may be due to cases where the source of beta activity is symmetrically distributed around the electrode, resulting in small potential differences between recording segments and, in turn, low beta activity. Notably, Narrow-Bipolar showed comparable performances even with data post-processing, including more robust identification of beta peaks involving the removal of the aperiodic spectral component and data normalization. This finding is relevant since it suggests that the Narrow-Bipolar approach is reliable and robust.

An additional strength of our study is that clinical and neurophysiological information was assessed multiple times, allowing us to compare IP and the last cFU. Clinically, we found that the levels selected using monopolar review at IP were maintained after multiple follow-up visits (cFU) in 75% of cases. Further supporting its superiority among beta-based selection strategies, Narrow-Bipolar was the approach demonstrating the highest selection stability from IP to cFU. The longitudinal design of our study also allowed us to assess the temporal stability of beta activity. We observed no significant differences in beta peak frequency, peak amplitude, or mean power in the low- and high-beta bands when comparing IP and cFU recordings. This finding aligns with a recent study by Fasano et al.³⁵, which examined beta oscillations at various time points following STN-DBS surgery in PD patients. That study reported an initial increase in mean beta power, followed by sustained stability in subsequent assessments, whose timing mirrored the schedule used in our research. Altogether, these results underscore the importance of timing in LFP recordings for DBS programming. By performing monopolar review at >3 weeks after surgery, we aimed to minimize the influence of the post-operative “stun” (microlesional) effect on beta oscillations^31,35. The observed beta stability in space (i.e., lead level), frequency, and power over the following months suggests that beta activity may serve as a reliable support for level selection when measured beyond this post-surgical period.

Another key finding of our study relates to the ability of LFP-guided programming using Narrow-Bipolar to predict the clinically selected contact after multiple follow-up visits. This has relevant translational implications when considering the possibility of integrating LFPs into the earliest phases of DBS programming. The fact that beta activity at IP is a good predictor of the optimal level determined after multiple clinical assessments highlights the reliability and robustness of this approach. Furthermore, the concordance with initial monopolar review of the Narrow-Bipolar method was not inferior to imaging-guided level selection, which is currently considered another valuable alternative to conventional CP^7,9,10,30. However, Narrow-Bipolar achieved even greater performances in predicting clinically selected levels at cFU than the imaging-based approach. This result might reflect the different identification strategy between the two approaches. While imaging can inform lead placement by providing static guidance of sensorimotor anatomical subregion, LFPs offer a dynamic and physiology-based confirmation of the functional sweet spot within the target^36,37. Functional rather than structural guidance may thus be more sensitive in identifying the optimal target site^13,22. In this context, our results highlighting the potential of beta activity in DBS programming are overall consistent with previous studies on intraoperative and early postoperative LFP recordings^22–24,28.

Importantly, from a practical standpoint, our overall findings point to the potential usefulness of the Narrow-Bipolar approach in guiding DBS clinical programming. Specifically, it may allow halving the number of contacts requiring clinical testing. Indeed, when beta activity at contacts 0–1 exceeds 2–3, a more ventral beta source is suggested, and thus only contacts 0, 1A, 1B, and 1C could be tested. Conversely, the opposite pattern would suggest a more dorsal source, and clinical testing may be limited to contacts 2A, 2B, 2C, and 3. Figure 4 shows a proposal algorithm for a possible neurophysiologically-oriented fast DBS programming based on our data.

Fig. 4. Algorithm for a possible neurophysiologically-oriented fast DBS programming.

Flow-chart for application of LFP-based level selection using Narrow-Bipolar method, comparing beta peak amplitude in bipolar recordings from 0–1 and 2–3 contact pairs.

Our study has some limitations. First, because of the technological limitations of available sensing-enabled DBS systems, LFP signals were recorded using a bipolar montage, which yields intrinsic ambiguity and limitations in localizing the source of the beta oscillations, especially for extreme levels (0 and 3)³⁴. We tried to overcome this issue by enrolling only patients in whom one of the two central levels had the widest therapeutic window at monopolar review, but this inclusion criterion limits the generalizability of our findings for cases where upper- or lowermost levels are clinically optimal. It should be noted, however, that surgical placement of directional leads is typically planned so that central segmented contacts (enabling more precise current steering) are positioned within the optimal functional target region, as determined by intraoperative neurophysiological assessment^38,39. Consequently, the a priori likelihood that the central levels (1 and 2) fall within the optimal target location is higher compared to the peripheral contacts (0 and 3). Another potential confounding factor of the study is that patients were evaluated >1 h beyond their usual inter-dose interval (clinical OFF state). Within this timeframe, some patients might not have reached a fully stable OFF condition. However, all patients were clinically assessed to confirm the OFF state, and longer withdrawal periods were not feasible in some patients due to severe OFF symptoms, such as OFF dystonia and severe gait disturbances. In addition, we defined as best contacts those displaying the widest therapeutic window, as in previous studies, even though those may not necessarily yield the greatest motor symptoms improvement^40,41. Importantly, recently updated DBS systems allow measuring LFPs from individual contacts independently by referencing to the contralateral DBS lead. However, such monopolar LFP recordings need to be validated for application in clinical practice, and the method presented here might be useful for this aim. Furthermore, monopolar recordings are not applicable in specific cases (i.e., unilateral DBS lead or altered impedances on one electrode) where bipolar LFP signals might still be informative and are more sensitive to artifacts. Therefore, it is likely that combining monopolar and bipolar recordings will provide the most comprehensive and accurate neurophysiological information, leveraging the spatial resolution of monopolar LFPs and the robustness of bipolar signals to noise and artifacts. Finally, a further limitation is that directional and multipolar stimulation were not leveraged in this study. Among the 19 patients, only five (3 STN and 2 GPi) were receiving directional stimulation, and none were receiving multipolar stimulation during clinical programming. Therefore, our analysis focused on ring-level contact selection, which was the most relevant and consistent approach for our cohort. Future studies should also specifically investigate the potential of segment-based recordings to guide directional stimulation programming and clinical outcomes of LFP-guided reprogramming over conventional CP on large patient cohorts.

In conclusion, our study provides real-world evidence supporting the use of LFPs to assist DBS contact selection in PD for both STN and GPi. Specifically, evaluating beta peak amplitude using the Narrow-Bipolar beta-based method might offer a valid and practical approach to orientate clinical programming, potentially halving the number of levels to test. Conventional CP can be time-consuming, subjective, and highly dependent on clinician expertise and patient feedback, which might be confounded by fatigue, cognitive impairment, speech disturbances, or delayed onset of clinical improvement^4,42. Using beta-based approaches would reduce programming times and follow-up visits, thus decreasing patient discomfort and healthcare system burden, while providing a simplified and more broadly applicable strategy in resource-limited centers.

Methods

Participants and clinical evaluations

Nineteen PD patients (9 females, age 55.8 ± 7.5 years) treated with bilateral STN- (12 subjects) or GPi-DBS (7 subjects) were included. All were implanted with the sensing-enabled Medtronic Percept™ PC neurostimulator and quadripolar segmented leads (model B33005, 0.5 mm intercontact spacing) identical for both STN and GPi. Monopolar review was performed after L-DOPA withdrawal (>1 h beyond the usual inter-dose interval, in the clinical OFF condition), and all four lead levels were tested in a monopolar configuration (i.e., using the pulse generator as anode and the tested level as cathode), separately for each hemisphere^4,43. The therapeutic window was assessed by increasing stimulation amplitude in 0.2 mA steps and evaluating the effect on Parkinsonian motor symptoms (bradykinesia, rigidity, and tremor), with bradykinesia severity most frequently used as the target symptom given its consistent presence across patients. The level with the widest therapeutic window was selected for chronic stimulation. Importantly, CP was solely based on clinical assessment, blinded to LFP data or imaging guidance. Due to technical constraints of bipolar recordings of the Percept™ PC, it is currently impossible to obtain a real-time, reliable estimate of beta power sources located at the most cranial and caudal contact levels (i.e., 0 and 3). Therefore, since the aim was to compare neurophysiology-based vs. clinical-based programming, only patients showing the widest therapeutic window at one of the two central levels were included. All patients underwent multiple on-demand follow-up visits (mean: 5.7, range 4–8), and stimulation was optimized if needed (i.e., side effects or unsatisfactory clinical improvement). The last cFU was always performed ≥6 months after surgery and >1 month after stable stimulation parameters and contacts (mean time from IP to cFU: 9 months, range 6–13 months). Clinical and demographic characteristics of patients and the timing of assessments are reported in Table 2. Research was completed in accordance with the Helsinki Declaration and approved by the Local Ethics Committee of Padua Province. All patients gave their written informed consent.

Table 2.

Demographic and clinical characteristics of patients

Subject	Target	Sex	Age	Disease duration (years)	Age at onset (years)	MDS-UPDRS III pre-DBSa	MDS-UPDRS III post-DBSb	LEDD pre-DBS	LEDD post-DBS	N visits between IP and cFU	cFU: time after implant (months)
1	STN	F	50	17	32	54	18	1186	80	6	9
2	GPi	F	63	17	45	51	42	1185	735	7	13
3	STN	F	58	6	51	42	33	720	270	5	8
4	STN	M	54	9	43	29	17	1520	120	8	16
5	STN	M	62	12	50	46	27	1290	595	5	6
6	GPi	F	70	9	60	38	26	690	450	6	8
7	STN	M	52	14	37	36	23	661	505	7	11
8	GPi	M	45	9	35	84	62	875	325	8	12
9	GPi	F	52	19	12	55	40	550	400	7	12
10	STN	M	56	11	44	61	32	1740	740	7	13
11	GPi	F	64	17	47	56	47	1180	170	4	8
12	STN	F	66	19	46	74	45	917.5	575	6	9
13	STN	M	58	13	45	56	45	1055	400	4	7
14	STN	M	60	12	48	45	31	1327	570	5	7
15	STN	F	53	4	48	29	26	1305	400	5	6
16	STN	F	65	9	56	41	22	700	325	5	6
17	GPi	M	69	23	45	62	43	1113	1008	6	6
18	GPi	M	54	13	41	47	37	850	400	5	8
19	STN	M	47	7	40	40	26	1502	550	4	6
Mean ± SD			55.8 ± 7.5	12.6 ± 5	43.4 ± 10.1	49.8 ± 14.2	33.8 ± 11.6	1071.9 ± 333.3	453.6 ± 228.3	5.7 ± 1.2	9 ± 2.9

^aPerformed before surgery, in the OFF-medication condition.

^bPerformed 1 year after surgery, in the OFF-medication, ON-stimulation condition.

cFU clinical follow-up, LEDD Levodopa Equivalent Daily Dose, IP initial programming, SD standard deviation.

Data acquisition and analysis

Immediately after clinical assessment at IP and cFU, LFPs were acquired with patients at rest with their eyes open, 2 min after turning off the stimulation. Recordings from each hemisphere (38 nuclei) were obtained using the BrainSense™ Survey (BSS) feature separately between level pairs (0–3, 0–2, 1–3, 0–1, 1–2, and 2–3), and individual segment pairs (1A–1B, 1B–1C, 1A–1C, 2A–2B, 2B–2C and 2A-2C) (Fig. 5). Briefly, BSS captures LFPs sampled at 250 Hz for ≈40 s with stimulation turned off and displays power spectra of each bipolar recording^26,27. Data were stored locally on the Clinician Programmer and exported as JSON files for offline analyses, which were performed using MATLAB v.2022b (The MathWorks Inc., Natick, Massachusetts). A 5th-order Butterworth filter was applied (5–98 Hz). Power spectra were computed with Welch’s method (1 s window length, 50% overlap) and beta peaks (13–30 Hz) were detected with the findpeaks MATLAB function, visually inspected and manually adjusted if needed. Two hemispheres (both STN) were excluded from subsequent analyses since no clear beta peak could be detected. Consequently, a total of 36 hemispheres were included in the final analyses. In cases where two distinct beta peaks were present (observed in seven hemispheres), the peak with the highest power was selected for subsequent analysis. Each identified peak was then classified into the low-beta (13–20 Hz) or high-beta (21–30 Hz) sub-band. Then, bipolar recordings from levels and horizontally paired segments were ranked in descending order based on the power spectral density (PSD) of the identified beta peaks. In separate analyses, LFPs were further post-processed to increase the robustness of peak identification and further validate the results on the performances of LFP-based methods. We used two complementary strategies: (i) estimation and subtraction of the aperiodic (1/f) component from each bipolar recording using the FOOOF (Fitting Oscillations & One-Over-F) open-source software, and (ii) z-scoring of power spectra across frequencies (5–49 Hz) within each recording, followed by computation of the mean power within ±1 Hz of the peak frequency to rank bipolar recordings^44,45 (see also Supplementary Information). To evaluate possible longitudinal changes of beta activity, power spectra were z-scored across frequencies (5–49 Hz) within each recording. Beta peak amplitude, mean PSD in the beta, low and high beta ranges were then computed and compared between IP and cFU.

Fig. 5. Experimental protocol.

Before surgery, all patients underwent a brain MRI and clinical motor evaluation using MDS-UPDRS-III. Bipolar LFPs were recorded 3–4 weeks after surgery at initial programming (IP) and after ≥6 months of clinical follow-up (cFU). The beta peak amplitude, computed on LFPs recorded from contact pairs, was compared according to the three different methods: Broad-Bipolar (1–3 vs. 0–2), Narrow-Bipolar (2–3 vs. 0–1), and Segment-Bipolar (1A–1B, 1B–1C, 1A–1C, 2A–2B, 2B–2C and 2A–2C). Power spectra from a PD patient are shown in the bottom: in this case, level 2 was selected in all methods.

Neurophysiological methods for beta-based level selection

LFP-based contact selection between middle segmented levels (1 vs. 2) was performed by evaluating beta peak amplitude using three methods comparing LFPs recorded by contact pairs with the same distance (i.e., 0–2 vs. 1–3, 0–1 vs. 2–3, 1A/B/C-1B/C/A vs. 2A/B/C-2B/C/A) (Fig. 5). Importantly, all programming strategies were specifically designed for direct application as a bedside tool solely using the Clinician Programmer without requiring offline data processing.

Three different contact selection methods were evaluated and compared in this work (Fig. 5). The Broad-Bipolar and Narrow-Bipolar methods include both central and peripheral contacts (non-adjacent 0-2/1–3 pairs and adjacent 0-1/2–3 pairs, respectively) and are both based on the assumption that beta activity is higher when one contact is inside the source of beta activity and the other outside²⁹. Therefore, they allow to discriminate whether the source of beta activity is located more ventrally (beta peak 0–2 > 1–3 or 0–1 > 2–3) or dorsally (beta peak 1–3 > 0–2 or 2–3 > 1–2). For Broad-Bipolar, if the beta peak from the 0–2 recording was higher, level 1 (more ventral level between central contacts) was selected as the “best”; if the peak from the 1–3 recording was more prominent, level 2 (more dorsal level between central contacts) was chosen. In the Narrow-Bipolar method, 1 or 2 was considered as the “best” level depending on whether the greatest beta peak was from 0–1 or 2–3 bipolar recording, respectively. Notably, despite the same general assumption, the key difference between Broad-Bipolar and Narrow-Bipolar methods was the width of the bipolar recording.

Finally, in the Segment-Bipolar method, the optimal stimulation level was identified by comparing the beta peak power in recordings from horizontally paired segments (i.e., A–B, B–C, A–C) between levels 1 and 2 (1A–1B vs. 2A–2B; 1B–1C vs. 2B–2C; 1A–1C vs. 2A–2C)³⁰. Here, the level showing higher beta peak amplitude in at least two of the three comparisons was considered the “best” level.

Neuroimaging data acquisition and contact selection

All patients underwent a pre-surgical brain MRI (including 0.9 mm iso 3D-MPRAGE, 0.9 mm iso 3D-T2-weighted and 2D-SWI sequences) within 1 month of the implant on a 3T Philips Ingenia (Philips Healthcare, Amsterdam, Netherlands). Post-operative head CT scans (slice thickness ≤1 mm, no gap between slices; reconstruction slice thickness ≤0.625 mm; soft tissue reconstruction kernel) were obtained >3 months after surgery to minimize pneumocephalus and brain shift artifacts. DBS electrodes were localized using Lead-DBS v3.1, an open-source MATLAB toolbox that integrates pre-operative MRI with post-operative CT scans⁴⁶.

Electrodes and subcortical targets were visualized using the DISTAL atlas⁴⁷. An experienced rater (LR), blinded to the patients’ clinical, neurophysiological and programming information, identified the two “best” levels^9,48 based on spatial proximity to the sensorimotor subregion of the target and distance from other neighboring structures (e.g., internal capsule and substantia nigra). This approach accounts for the common clinical scenario where multiple lead levels are near the anatomical target and ensured that at least one of the two middle-segmented contacts was selected as “best” level. 3D reconstructions of electrode localizations in STN and GPi are reported in Supplementary Fig. 2.

Statistical analysis

The performance of each beta-based method was evaluated using three different complementary metrics: correlation, agreement and selection similarity, to provide a comprehensive assessment of each selection strategy. Correlation captures the strength and direction of relationships between each method and CP, agreement quantifies the degree of concordance in selected levels, reflecting how closely each method replicates the gold standard, while similarity provides a proportion of exact matches in contact selection. The correlation strength between LFP-based methods and the CP-based level selection was evaluated using the ϕ coefficient, a measure of binary correlation suitable for categorical agreement. Correlation was considered very weak for 0 ≤ |ϕ| < 0.2, weak for 0.2 ≤ |ϕ| 0.3, moderate for 0.3 ≤ |ϕ| < 0.4, strong for 0.4 ≤ |ϕ| < 0.7, and very strong for 0.7 ≤ |ϕ| ≤ 1⁴⁹. The level of agreement between LFP-based methods and CP-based selection was estimated with the QWK. The agreement was considered slight for 0 < QWK ≤ 0.2, fair for 0.2 < QWK ≤ 0.4, moderate for 0.4 < QWK ≤ 0.6, substantial for 0.6 < QWK ≤ 0.8, and almost perfect for 0.8 < QWK ≤ 1⁵⁰. QWK p values were derived using permutation testing (1000 permutations) implemented via the sklearn.metrics.cohen_kappa_score function. Confidence intervals were estimated using bootstrap resampling with 1000 iterations. Finally, the degree of similarity between contact selection performed by LFP-based methods and CP was quantified by computing a custom similarity function that assigns a score of 1 or 0 based on the presence of a match between selections. To assess the statistical significance of similarity measures, a bootstrap hypothesis testing approach was employed. For each similarity measure, 25,000 bootstrap samples were generated by resampling with replacement from the original dataset. Unlike permutation-based approaches, which are limited by the number of unique label rearrangements, the bootstrap allows for a far larger set of resamples, providing robust estimation even with relatively small sample sizes. The mean similarity was calculated for each bootstrap sample, forming a bootstrap distribution for each measure. A one-sided hypothesis t test was then conducted to compare the observed mean similarity of each measure against the bootstrap distribution of a RS. The null hypothesis assumed that the observed mean similarity was not significantly greater than the mean similarity from RS. The p value was calculated as the proportion of bootstrap samples from RS ≥ to the observed mean similarity for the given measure.

Correlation strength, agreement, and similarity levels between LFP-based methods at IP and CP at cFU were used to estimate the ability of each neurophysiological method to predict the level selected after multiple follow-up visits. These measures were also used to compare the overall performance of the best LFP-based method to the imaging-based approach we tested. Additionally, to evaluate possible differences between STN and GPi in the performance of each LFP-based method, we compared the proportion of cases with neurophysiology-CP concordant level selection between DBS targets using Fisher’s exact test. The same approach was used to assess possible differences when considering peaks in the low- or high-beta sub-bands. Importantly, all these analyses were performed on both LFP raw data (main text) and data obtained after various post-processing procedures, including normalization and removal of the aperiodic spectral component (see Supplementary Information).

To verify whether the changes in the stimulation level from IP to cFU (i.e., from 1 to 2 or vice versa) had a similar distribution to those occurring using the LFP-based methods, we applied the Chi-squared test. To evaluate the stability of beta-band activity over time, we assessed the recordings from contact pairs included in the best LFP-based method (Narrow-Bipolar). The bipolar recording displaying the highest beta peak at IP was selected, and the beta peak amplitude, frequency, and low- and high-beta mean PSD were compared between IP and cFU. The Shapiro-Wilk test for normality was used to assess data distribution in the various analyses. Parametric or non-parametric tests were applied accordingly, with the significance level set at p < 0.05. Statistical analyses were implemented using Python 3.13 libraries (scipy, sklearn) and MATLAB.

Author contributions

Conception of the work and project design: A.G., A.A. and V.D. Data collection: V.D., A.G., D.C. and L.L.G. Data analysis: V.D., L.W., L.R. and C.P. Interpretation of data, manuscript draft, revision and editing: V.D., G.T., A.A., A.L. and A.G. All authors read and approved the final submission.

Data availability

The data obtained in this research are available from the corresponding author upon reasonable request.

Competing interests

The authors declare the following competing interests: A.A. has received honoraria/consultation fees from AbbVie, Bial, Bayer, Stada, Covantec, Ever Pharma, Medscape, Roche, Theravance Biopharma, UCB, and Zambon. AA has also received funding from Horizon 2020—Ministry of Education, University and Research (MIUR), Ministry of Health (MOH). A.G. has received compensation for consultancy and speaker-related activities from Bial, Zambon, STADA; he receives funding from the European Union “Next Generation EU, Mission 4 Component 1, Project code: P20223HHZ8, CUP: C53D23008440001”. G.T. received financial support from Boston Scientific and Medtronic, not related to the present work. Research agreement with RuneLabs is not related to the present work. G.T. receives funding from the Swiss National Science Foundation (project number: PZ00P3_202166) and the Swiss Parkinson Association.

Supplementary information

The online version contains supplementary material available at 10.1038/s41531-025-01208-4.