Towards an objective, standardized beta frequency peak detection algorithm to inform adaptive deep brain stimulation programming in Parkinson’s disease

doi:10.21203/rs.3.rs-3994762/v1

Download PDF

Article

Towards an objective, standardized beta frequency peak detection algorithm to inform adaptive deep brain stimulation programming in Parkinson’s disease

https://doi.org/10.21203/rs.3.rs-3994762/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 09 Aug, 2024

Read the published version in npj Parkinson's Disease →

You are reading this latest preprint version

Oscillatory activity within the beta frequency range (13-30Hz) serves as a Parkinson’s disease biomarker for tailoring deep brain stimulation (DBS) treatments. Currently, identifying clinically relevant beta signals, specifically frequencies of peak amplitudes within the beta spectral band, is a subjective process. To inform potential strategies for objective clinical decision making, we assessed algorithms for identifying beta peaks and devised a standardized approach for both research and clinical applications. Employing a novel monopolar referencing strategy, we utilized a brain sensing device to measure beta peak power across distinct contacts along each DBS electrode implanted in the subthalamic nucleus. We then evaluated the accuracy of ten beta peak detection algorithms, both existing and new, against a benchmark established by expert consensus. The most accurate algorithms matched the expert consensus in performance and reliably predicted the clinical stimulation parameters during follow-up visits. These findings highlight the potential of algorithmic solutions to overcome the subjective bias in beta peak identification, presenting viable options for standardizing this process. Such advancements could lead to significant improvements in the efficiency and accuracy of patient-specific DBS therapy parameterization.

Health sciences/Biomarkers/Diagnostic markers

Health sciences/Diseases/Neurological disorders/Parkinson's disease

Subthalamic nucleus (STN) deep brain stimulation (DBS) for Parkinson’s disease (PD) has demonstrated favorable outcomes in ameliorating motor and nonmotor symptoms ^1–3. With advancements in imaging and connectomics, treatment efficacy and enhanced therapeutic benefit has become dependent on precise localization and selection of the optimal DBS contact configuration ^4–6. A myriad of studies has broadly identified the dorsolateral STN as the “electrophysiological sweet spot” for targeting symptoms of PD ^6,7. Proximity of stimulation to this region exhibits a positive correlation with improvement of motor performance ^{6–8 9,10}. The importance of optimized targeting and stimulation parameterization is demonstrated by the effects of conventional non-adaptive, continuous high-frequency DBS.

Within the dorsolateral STN of patients with PD, local field potentials (LFPs), frequently exhibit increased oscillatory synchrony and elevated peak power in the beta frequency range of 13-30Hz ^6,11–17. Clinical evidence from human studies supports the growing consensus that LFP activity within this beta range is implicated in tonic movement ^15,18,19, is strongly correlated with rigidity and bradykinesia ^20–22, and that attenuation of beta power by dopaminergic medication and DBS results in symptom improvement ^{6,12–15,19,23–25}. Associations with symptom improvement are further supported by targeting the lower beta band (~ 13-20Hz) ^{15,23,24,26–29}, although there are reports of upper beta band (~ 20-30Hz) activity correlating with freezing of gait ^27,30,31 and governing lower limb activity ³². As such, specific relationships between the upper/lower beta band and clinical outcomes are debated but nonetheless implicative for aDBS therapies. Relatedly, adaptive DBS (aDBS) relies on biomarkers, predominantly the beta band, to adjust stimulation parameters based upon the degree beta frequency modulation. Fluctuations in the beta frequency range are currently being implemented to guide aDBS ^11,12,14,30. As such, identifying optimal aDBS settings is essential to maximizing therapeutic efficacy.

Maximal amplitude beta peak (MBP) presence in LFP power spectral densities (PSD) were strongly related to the “electrophysiological sweet spot” for lead placement in 92% of patients undergoing DBS ^6,33. As such, several studies have corroborated the utility of PSDs to identify the MBP to guide clinical decisions on DBS contact selection and parameterization to suppress beta power, which generally elicits clinical benefit ^{6,8,12,18,24,34}. While current clinical-commercial tools, including those from Medtronic, Newronika ³⁵, and PINS ³⁶ permit spectral analysis, there remains a need to systematically quantify LFP peaks from the displayed PSD [Figure 1].

Clinical decision-making is thus augmented by efficiency in beta peak identification for accurate contact selection and parameterization. Currently, the standard practice uses bipolar sensing montages to determine optimal DBS contact placement ³⁷. This conventional practice details meticulous comparison of MBP between contact pairs, denoted by the largest difference in pairs, which is largely reliant on trial-and-error and requires expertise in interpreting computed results ^12,24. To relieve this clinical burden and optimize the clinical workflow, an efficient peak detection algorithm could be integrated. Current efforts aim to standardize this process, and several algorithms have been proposed for detecting beta peaks, with varying degrees of complexity and generalizability. However, most of these studies perform peak detection based on subjective review ^{11,18,25,34,38}.

In this study, we evaluated objective algorithms for beta peak identification from implanted DBS electrodes. We observed acceptable concordance across observers, and despite subtle differences in algorithm success in classifying consensus peaks, many perform with low error. Importantly, algorithm performance varies by participant PSD, and we introduce a novel monopolar sensing montage that improves LFP resolution of beta peak classification and clinical interpretability. In light of global and socioeconomic disparities in DBS access ^39,40, our work could advance towards a streamlined, objective beta peak detection algorithm for identifying optimal contact configurations and stimulation parameters with greater feasibility and with less exhaustive programming guided intervention.

We collected LFP data from SenSight™ directional DBS leads (Model B33005, Medtronic; Minneapolis, MN USA) implanted in the STN in five participants with PD (N = 10 hemispheres). All participants were recruited from the University of Colorado Anschutz Medical Campus Advanced Therapies in Movement Disorders Program. LFP data can be collected using Percept™ PC BrainSense™ features using a bipolar sensing configuration. In addition, the investigational-use Percept SP Advanced Sense research system was used to collect LFP data to enable a monopolar sensing configuration through a temporary software unlock. The implanted neurostimulator maintained standard commercial operation, and therapy was not affected while using the research system. This temporary unlocking of additional sensing configurations was used only in the clinic under movement disorder neurologist supervision. LFP data were collected from one hemisphere per session. A sensing montage that iterated through predetermined configurations was performed to wirelessly stream LFP data at 250Hz for 55 seconds per recording (both bipolar and monopolar contact configurations). The following configurations were collected: Bipolar: all possible segment pair configurations (n = 9), all possible level pair configurations (n = 6); Monopolar: all individual segments (n = 6), all individual levels (n = 4). Three trials of each montage were collected to account for variability. Our study was carried out in accordance with the Colorado Multiple Institution Review Board (COMIRB # 21-4938) and the Declaration of Helsinki, with written informed consent obtained from all study participants.

LFP postprocessing

Postprocessing and statistical analyses were performed using MATLAB R2023a [MathWorks, 2023]. Periodograms were generated using pwelch with a Hanning window, 50% overlap, and 256 discrete Fourier transforms (0.9766Hz resolution). Differences between right and left STN PSDs were not statistically significant per one-way ANOVA and Tukey’s HSD (p_RSTN = 0.9294 ± 0.1380, p_LSTN = 0.9295 ± 0.1364). Thus, session-level PSDs were averaged per contact.

Expert visual inspection of the PSDs and peak identification, as is performed in standard practice, was used to compare algorithms. A GUI was developed that permitted each reviewer to select up to two peaks. The GUI collected the x-axis frequency values (Hz) that were selected by the reviewer. The following strategy was used to derive the consensus peak frequency for each PSD: First, the proportion of all selected frequencies was computed. If one selected peak frequency exceeded 50%, it was set as the consensus frequency. If no selected frequency exceeded 50%, the median (most commonly selected) frequency was computed and selected as the consensus frequency. This approach for consensus derivation yielded 98.4% congruence. Figure 2 depicts the process that was used to compare reviewer-derived consensus peaks for both bipolar and monopolar examples to highlight the challenges of beta peak identification.

Ten existing or exploratory beta peak detection algorithms were selected from recent literature (published within the past ~ 10 years). Algorithms were heuristically categorized as absolute, algebraic, or elemental decomposition.

Absolute algorithms. Absolute implied that the value(s) within the algorithm did not change based on PSD properties. The Medtronic Percept PC detection algorithm (I), as described in ⁴¹ and used in ^42,43, utilized a > 1.1µVp threshold to identify a singular MBP. Other clinically relevant peaks were traditionally determined subjectively. As this algorithm was modified to be objective, it was subject to a high false-negative rate. The second method (II), as employed in ²⁶, identified any local maxima within the lower beta band (13-20Hz) and upper beta band (21-35Hz), thus assuming bimodality. The only alteration made from its original methodology was the absence of PSD normalization, as done in ²⁶.

Algebraic algorithms. Algebraic implied that the algorithm fluctuated based on numeric properties of the PSD. For example, the third algorithm (III) was adapted from an approach described in ³³. Herein, using a normalized PSD, a dynamic threshold was set for the MinPeakProminence parameter within MATLAB’s findpeaks function, defined as \(\frac{median\left({PSD}_{norm}\right)}{value}\), where the a priori defined value conventionally ranged from 4 to 12. However, in our adapted algorithm, an optimized value was found by iteratively running 10,000 potential values ranging from 0 to 10 until the greatest overlap with the consensus ground truth was achieved. The fourth algorithm (IV) was similarly adapted from ³³ such that the aforementioned iterative optimization procedure was used, but instead on a dynamic MinPeakProminence parameter threshold drawing from ^44,45 defined as \(stdev\left({PSD}_{norm}\right)\times value\). Lastly, the fifth method (V) was described in ⁴⁶ and employed in ²⁷. In this unaltered method, a beta peak was defined as a local elevation in power, in which the 4–6 points defining a peak were at least 20% greater than the means of the 4–6 points beyond either tail (therefore 8–12 points) of this local elevation.

Elemental decomposition algorithms. This subtype considered the aperiodic and periodic components of PSDs. For example, in the sixth method (VI), as described in ⁴⁷ and used in ⁴⁸, PSDs were deconstructed into aperiodic (\(\frac{1}{f}\) trend) via a linearized Lorentzian function and periodic components using a series of Gaussian fits. Peaks and prominences were then further fit based on these Gaussian and Lorentzian models via several ancillary mathematical transformations. From these series of model fits, parameters were defined per PSD to detect peaks. The seventh method (VII) ⁴⁵ defined the aperiodic component as a linearized Lorentzian function \({10}^{b}\times \frac{1}{(k+ {F}^{x})}\), where b = offset, k = knee parameter (in this case, k = 0), F = frequency vector, and x = aperiodic exponent. This aperiodic component was subsequently subtracted from the modified PSD, and a threshold of \(mean\left({PSD}_{modified}\right)+(3\times stdev\left({PSD}_{modified}\right))\) was used to identify peaks via MATLAB’s findpeaks function. In the eighth method (VIII), elemental decomposition was performed similarly to method VII, though an alternative dynamic threshold was employed from ⁴⁴. Instead of using findpeaks, this Heaviside-like threshold was defined as \(mean\left({PSD}_{modified}\right)+ stdev\left({PSD}_{modified}\right)\), from which peak indices were defined. Following the theme of changing thresholding equations, the ninth method (IX) also employed the same elemental decomposition as method VII, but instead defined the findpeaks threshold for the modified PSD as that of method III ³³. Lastly, the tenth method (X) defined each periodic element in a different fashion. Herein, the logarithmic transform of the PSD was fit to a 5th order polynomial using the local minima of the lower frequency range (3-5Hz) and values beyond the beta range as bounds for a \(\frac{1}{f}\) aperiodic proxy. The difference between the \(\frac{1}{f}\) aperiodic proxy and the \({log}_{PSD}\) was fit to a Gaussian model, and peaks were identified thereafter using the default findpeaks function ¹⁷.

Assessing algorithm accuracy for predicting 3-month follow-up therapeutic contact level of stimulation. Algorithms that passed the following criteria: 1) Bland-Altman comparison was non-significant, 2) MSE was lower than the median MSE across algorithms (MSE = 6), 3) at least one peak was identified for greater than 75% of the PSDs. This subtype considered the aperiodic and periodic components of PSDs. For example, in the sixth method (VI),

Statistical analyses

We evaluated the performance of individual reviewers and algorithms against the consensus peak distribution for the first peak using a 2-sample Kolmogorov-Smirnov test. ANOVA with post-hoc tests were used to highlight between-subjects performance differences, and one-sided paired t-tests elucidated algorithm performance differences between electrode referencing configurations. A Bland-Altman test compared methods to the ground truth consensus across subject-level and reference strategy-level strata, wherein nonsignificant p-values, roughly equal to mean-squared-error (MSE) at or smaller than five (median MSE of all algorithms), compared to expert inspection determined acceptable peak localization accuracy. Histograms and UpSet plots ⁴⁹ were employed to visualize the range of MBPs and the degree of concordance between expert reviews and objective algorithms.

Characterization of reviewer peak selection performance across all PSDs

For this study, we evaluated the performance of seven independent expert reviewers in their subjective selection of the first and second (if identified) beta peak frequency from fifteen PSD plots each across five participants (both left and right hemispheres). As a group, the distribution of consensus peaks derived from all 250 PSDs was non-significantly different from the peak selection distribution of each individual reviewer, determined using a 2-sample Kolmogorov-Smirnov test with the following p-values for each reviewer: R1: p = 0.97, R2: p = 1, R3: p = 1, R4: p = 0.78, R5: p = 0.49, R6: p = 1, R7: p = 0.87 (Fig. 3A). At a minimum, there are two factors that could affect the derivation of the consensus peak quality: 1) a subset of reviewers performed substantially worse on accurate classification of the consensus peak, or 2) a subset of PSDs was more challenging to classify across reviewers. To assess these two factors, we used an UpSet plot (Fig. 3B)⁴⁹ which is a method to visualize and quantify the number of intersections between sets. Both of these issues are not mutually exclusive and could contribute to the consensus peak accuracy. Summarizing across reviewers, regarding the number of PSD sets in which an individual reviewer correctly identified the consensus beta peak, highlights two reviewers that made accurate selections for less than 70% of the PSDs (R7 61% and R4 69%) while the remaining reviewers made accurate selections on greater than 80% of the PSDs. In the main summary across reviewer and PSD set intersections, all reviewers correctly identified the consensus beta peak on 35% of the PSDs. To determine if there was a subset of PSDs that was difficult to accurately assess across reviewers, we identified the participant ID of PSDs for which fewer than four reviewers accurately identified the consensus beta peak (Fig. 3C). We found that two participants contributed approximately 25% of their PSDs to this low-consensus group (Participant 3 and Participant 5). Fewer problematic PSDs were observed for Participant 1 (2%), Participant 2 (14%) and Participant 4 (0%).

Characterization of algorithm peak selection performance across all PSDs

Following derivation of expert defined consensus beta peaks for all 250 PSDs, we sought to evaluate the performance accuracy for ten algorithms used in the literature for detecting PSD peaks. We did not impose any a priori restrictions on the maximum number of peaks that algorithms could detect. Furthermore, the algorithms were not configured to guarantee the detection of at least one peak, meaning it was possible for an algorithm to fail to detect any peaks. Comparing the consensus-derived peak distribution (for the first peak detected) with that of each individual algorithm revealed that four algorithms (I, II, VI, X) produced peaks with distributions significantly different from the consensus (2-sample Kolmogorov-Smirnov test, p < 0.01). In contrast, six algorithms (III, IV, V, VII, VIII, IX) yielded peak distributions similar to the consensus, as shown in Fig. 4A. To evaluate algorithm accuracy in detecting the consensus peak, we adopted an approach similar to our assessment of reviewer performance. We visualized the performance of individual algorithms across PSDs and analyzed the overlap among algorithms that successfully identified the peak. This analysis is shown in Fig. 4B. We observed that algorithms exhibited substantial variability in accuracy, shown in the lower horizontal bar chart in Fig. 4B, which summarizes individual algorithm accuracy across all PSDs. Three algorithms performed below 50% accuracy (I = 30%, VI = 47%, VIII = 48%), three performed below 75% accuracy (X = 64%, VII = 64%, V = 66%) and four algorithms performed above 75% accuracy (II = 75%, III = 75%, IV = 76%, and IX = 75%). The vertical bar chart of the UpSet visualization in Fig. 4B depicts the main summary across algorithm and PSD set intersections. The greater number of possible intersection sets resulted in greater diversity of set relationships in the plot. In general, all ten algorithms correctly identified the consensus beta peak on 11% of the PSDs. Following the approach we used in the reviewer performance assessment, we evaluated whether a subset of PSDs was difficult to accurately assess across algorithms. For this analysis, we quantified the participant ID of PSDs for which fewer than five algorithms accurately identified the consensus beta peak (Fig. 3C). We found that all participants contributed a subset of PSDs that were challenging for a half or more of the algorithms. The following participant PSD contributions were observed: participant 5 was the most challenging (62% PSDs, where < 4 algorithms match the consensus peak), followed by participant 3 (38%); participants 1 and 2 exhibited moderate difficulty with 20% and 24% respectively, and participant 4 again exhibited the least difficulty across algorithms (4%). All PSDs with successfully detected peaks by less than four algorithms were represented within the center of the donut chart (Fig. 4C). Algorithms also exhibited substantial variability in the number of peaks detected across PSDs (Fig. 4D); most algorithms predominantly detected two peaks per PSD.

Overall algorithm performance for detecting reviewer consensus beta power peaks

The main goal of the study was to assess algorithm performance in detecting expert-defined beta peaks in PSD representations of LFP data. We evaluated the performance of individual algorithms using two approaches. First, we used a Bland-Altman test to assess the agreement between the expert consensus peaks and the first peak identified by each algorithm. A t-test on the mean difference between the two measurement methods was used to determine if there was a significant difference from 0, indicating a systematic bias between the measures. In this case, a p-value less than alpha (0.05) would suggest a possible systematic bias, indicating that those algorithms may be less reliable. Second, we quantified algorithm accuracy by evaluating the mean squared error between the consensus peaks and the first peak derivations for each algorithm. Representative results showing the comparison of consensus and algorithm detected peaks are shown for all five participants in Fig. 5. (The top row depicts PSDs from segments on the left hemisphere that exhibited the highest peak beta, and the bottom row depicts PSDs from levels on the right hemisphere that exhibited the highest peak beta; all PSDs were derived from monopolar referenced recordings). Of greatest interest was the first peak detected, as all reviewers selected up to two peaks, with the first predicated as the main peak. For the first peak (Fig. 6A), we found that five of the ten algorithms passed the Bland-Altman test (I, V, IX, III, and IV) while five failed to pass (VII, VIII, II, X, VI). In general, there was a correlation between passing the Bland-Altman and having a lower MSE, with four algorithms that passed the Bland-Altman test exhibiting the lowest MSE. However, we found that MSE could be impacted by the number of PSDs in which any peak was detected. For example, algorithm I, which failed to identify any peak for 67% of the PSDs, had the lowest MSE; this algorithm was the most conservative algorithm in peak detection. All other algorithms successfully identified at least one peak in over 73% of the PSDs. We observed a similar phenomenon when assessing accuracy for the second peak (Fig. 6B). Only seven of the ten algorithms detected a second peak, and these detections occurred at a lower frequency across PSDs (average of 26% of PSDs per algorithm). The increased stringency in second peak detection resulted in lower MSEs, similar to the accuracy of algorithm I for first peak classification.

The cohort of PD participants (n = 5; 10 hemispheres) used for this study was representative of the heterogeneity in PSD morphology (e.g., peak number, slope, peak magnitude). Despite the inherent variability in overall PSD structure, both reviewers and algorithms were generally able to agree on the identified beta peak(s) in many cases. We directly evaluated the participant-based variability in algorithm performance to determine whether specific participants exhibited more challenging PSD morphologies for beta peak detection. In Fig. 6C, we arranged participants from highest to lowest accuracy based on the average MSE across algorithms, revealing notable differences that reinforce earlier results; for example, participant 4 exhibits less variability in MSE than participant 2. Applying an ANOVA to assess participant MSE variability demonstrated a significant difference in accuracy between participants (F(4, 45) = 6.04, p = 0.0006). Post-hoc tests revealed that participant 2 was significantly different from participants 3 (p = 0.04), 4 (p = 0.0002), and 5 (p = 0.01).

In this study, PSD plots were generated using two referencing strategies: bipolar, which references between neighboring contacts on the same lead, and monopolar, which involves referencing between contacts on different leads. We define the monopolar approach based on the concept that this strategy allows for the isolation of activity from a single contact or set of contacts (i.e., combined segments) on an implanted DBS lead. We sought to determine whether the monopolar approach resulted in greater isolation of peak activity and, thereby, greater accuracy in peak identification. In Fig. 6D, we directly compared the PSD data derived from monopolar and bipolar contact configurations using a paired t-test. We observed that PSDs generated by monopolar referenced contact configurations resulted in greater accuracy compared to bipolar referenced PSDs (monopolar mean MSE = 3.2; bipolar mean MSE = 7.8; p = 0.01).

Accuracy of top performing algorithms to predict contact level of stimulation at 3 months follow-up

Finally, we evaluated whether the top performing algorithms, ranked by accuracy in detecting expert reviewer-derived consensus beta peak(s), would have clinical utility. We replicated this approach by assessing how well the top three performing algorithms predicted the contact level of stimulation clinically determined and used at 3-months follow-up. We used three criteria to determine the top performing algorithms: 1) pass the Bland-Altman test while retaining the null hypothesis, 2) detect at least one peak from 75% of the PSDs, and 3) accuracy rating of MSE < 5, which was the median MSE across all ten algorithms. Per our evaluative criteria, three methods (III, V, and IX) were deemed most satisfactory. Of note, methods III and V were both algebraic methods such that a dynamic threshold was used to identify relevant peaks using properties of individual PSDs. Method IX employed the same algebraic thresholding approach as method III, albeit applied to a PSD whose 1/f component was subtracted; thus it was classified as elemental decomposition yet drew heavily from algebraic algorithms. As such, all three satisfactory methods overlapped in the use of individualized, dynamic peak amplitude thresholding procedure, pointing to a promising, reliable and standardizable peak detection strategy. At 3-month follow-up, seven of the ten hemispheres were programmed either at a single level or, if programmed across levels, had greater than 50% stimulation fractionated to one level. These seven hemispheres were used for the comparison. When comparing the level selected by the three top performing algorithms based on the max beta peak identification, we observed the following accuracies in predicting the contact level for each hemisphere at 3-months follow-up: V = 71%, IX = 86%, and III = 100%.

In this study, we evaluated the utility of a set of algorithms in matching the performance of experts in identifying therapeutic frequency peaks in PSD derived from LFP recordings in patients with PD. A primary goal was to assess objective clinical decision-making strategies for improved efficiency and efficacy in DBS programming. DBS therapy management relies on accurate evaluation of physiological responses to stimulation, including electrophysiological signals at the site of stimulation. In PD, real-time measures of LFP show that modulation of STN beta power correlates with various clinical responses ^18,23,50,51.

Clinicians have employed a variety of biomarkers to guide lead placement and stimulation titration for DBS. Neuroimaging-guided approaches for DBS programming have yielded non-inferior adverse motor symptom attenuation compared to conventional, monopolar review techniques^52,53. Other strategies have employed methods drawing from electric fields in tandem with functional MRI signatures along cortico-basal ganglia-thalamo-cortical circuitry to inform targeting and stimulation optimization ⁴. More recently, evoked potentials have also demonstrated utility as a viable biomarker, particularly with emerging advancements in stimulation parameterization ⁵⁴. However, by far, advances in therapeutic delivery have identified LFP features as robust, widely supported biomarkers for DBS programming and aDBS control system feedback ^{16,24,43,55–58}.

The current approach for evaluating beta power to guide DBS therapy in PD relies on spectral analysis. While current clinical-commercial tools facilitate spectral analysis and can be used to identify and monitor clinically relevant signals of interest, identification of validated frequency-based biomarkers (e.g., MBP, which may provide the most efficacious therapeutic outcome) relies on the subjective experience of a clinician [31], which lacks objectivity, relies on specialized knowledge, and may be subject to bias. Our focus on the primary peak within the low beta frequency range for guiding STN DBS parameter tuning for PD is predicated on the high reliability of this biomarker for predicting therapeutic response ^19,57,59. We assessed 250 PSDs from 5 participants and 10 hemispheres and observed high reviewer agreement (Fig. 3), with all reviewers exhibiting peak selections that did not significantly differ from the consensus peak distribution. Although reviewers exhibited high agreement for a significant proportion of PSDs, a subset of PSDs was challenging to assess across reviewers (Fig. 3C). In contrast, algorithm performance varied significantly, either in overall performance across PSDs or with respect to specific subsets of PSDs that were challenging to match with the consensus. For example, Participant 2 resulted in many mismatches across algorithms (Fig. 6C). To assess for systematic bias, we used the Bland-Altman analysis using a threshold of p > 0.05 as the first method to identify accurate algorithms that had acceptable peak identification accuracy when compared to expert-defined (Table 1, Fig. 6A B). We observed that more conservative algorithms (i.e., algorithms that failed to detect any peak in specific PSDs) exhibited high agreement with the consensus when peaks were identified. An algorithm that performed well on only a subset of PSDs may not be generalizable; therefore, we eliminated algorithms that failed to render a peak selection for greater than 30% of PSDs used in the study. In addition, we set the accuracy threshold to the median performance across the ten algorithms (MSE = 5). Notably, the three algorithms (III, V, and IX) that met performance criteria successfully predicted the therapeutic level of stimulation clinically determined for 7/10 hemispheres at 3-months follow-up. These methods exhibited shared qualities; methods III and V were algebraic methods such that a dynamic prominence threshold using PSD properties was employed to identify relevant peaks. Of note, method IX employed the same algebraic thresholding approach as method III, albeit applied to a PSD whose 1/f component was subtracted; thus, it was classified as elemental decomposition yet drew heavily from algebraic methodologies (Table 1). As such, all three algorithms overlapped in the use of dynamic peak amplitude thresholding, pointing to a promising, reliable mathematical mechanism undergirding a potentially standardizable, universal peak detection strategy.

Our exploration reveals that monopolar referencing may significantly improve the resolution of beta peak identification (Fig. 6D, [t(9) = -2.7687, p = 0.0109]). Findings spotlight practical candidates for objective beta peak detection using a high-resolution monopolar configuration with potential utility in empirically and efficiently determining optimal starting points for DBS therapy. Further refinement of a method utilizing this monopolar referencing strategy could offer a quantitative, generalizable peak detection approach that is independent of subjective review. Variable peak determination presents important implications for clinical decision-making. Of note, certain methods suggest peak bimodality in a subset of patients, while others do not, as illustrated by distribution differences (Fig. 4D). Our approach predominantly identified a more prominent peak within the low beta range. Given that low beta (~ 13-20Hz) power, in comparison to high beta (~ 21-30Hz) power, is more significantly correlated with bradykinesia severity ¹⁴ and more systematically suppressed in response to STN DBS therapy ²³, peak detection algorithms tuned for low beta peak identification could augment clinical programming efficacy. However, the algorithms we tested could have been tuned for peak detection in the high beta range, which has been implicated in PD phenotyping; high beta activity is more pronounced in akinetic-rigid compared to tremor-dominant motor subtypes ^31,60. Recent tractography work suggests that high beta may also reflect a regionally segregated informational pathway for lower limb cortical-subcortical coherence, observed in freezing of gait ^{11,27,29,31,32}, and others ascribe high beta activity to cognitive features of PD ¹³ or normal physiology ²⁶. If high beta is associated with PD pathophysiology, uncovering bimodality may be crucial for symptom management. Thus, it is imperative that clinicians converge to a consensus on how to identify clinically relevant beta-band peaks in LFP power spectra.

Limitations of this study include small sample sizes, heterogeneity of different LFP montages (which is also a strength), and a narrow focus on the beta frequency band. Several studies have noted the clinical utility of LFP features within the theta ^30,61,62, alpha ^29,58, and gamma ^12,29,32 frequency ranges, and the algorithms tested in this study may or may not prove useful for these spectral bands. Nevertheless, this study compared theoretical results to applied clinical outcomes to evaluate our methods’ practical clinical utility. Notably, in our assessment of objective, algorithmic peak detection approaches, results revealed that top-performing algorithms predicted the optimal contact level clinically determined and programmed by expert neurologists in > 70% of participant hemispheres. Selection of specific algorithms based on study design and available resources in future work may improve data collection consistency and interpretation across studies.

By elucidating the ambiguity in current approaches and highlighting promising objective beta peak identification methods, this study points toward practical avenues for the advancement of standardized, efficient, and data-driven clinical decision-making strategies., Resulting insight may contribute to better-targeted stimulation delivery and optimized clinical benefit. Future studies may choose to derive approaches from methods III, IX, or V, given their translational utility in DBS programming, as demonstrated by acceptable peak detection and contact level prediction accuracy in our sample. Insights produced from this study also inform LFP biomarker-driven control strategies for advanced, patient-specific aDBS therapy. In addition, pioneering work towards a streamlined, standardized objective peak detection algorithm could be of benefit for under-resourced clinics across the world or telemedicine interventions by providing access to efficient, computerized peak detection algorithms without oversight from specialized movement disorders neurologists or other members of the care team.

DISCLOSURES: JAT, DSK received research funds from Medtronic. AHB, CZ, MC, RSR are employees of Medtronic.

CONTRIBUTORSHIP:

FUNDING: Investigator-initiated funding from Medtronic.

Bove, F. et al. Long-term Outcomes (15 Years) After Subthalamic Nucleus Deep Brain Stimulation in Patients With Parkinson Disease. Neurology (2021). https://doi.org:10.1212/wnl.0000000000012246
Kühn, A. A. et al. High-frequency stimulation of the subthalamic nucleus suppresses oscillatory beta activity in patients with Parkinson's disease in parallel with improvement in motor performance. J Neurosci 28, 6165–6173 (2008). https://doi.org:10.1523/jneurosci.0282-08.2008
Little, S. et al. Adaptive deep brain stimulation in advanced Parkinson disease. Ann Neurol 74, 449–457 (2013). https://doi.org:10.1002/ana.23951
Boutet, A. et al. Predicting optimal deep brain stimulation parameters for Parkinson’s disease using functional MRI and machine learning. Nature Communications 12, 3043 (2021). https://doi.org:10.1038/s41467-021-23311-9
Chen, P.-L. et al. Subthalamic high-beta oscillation informs the outcome of deep brain stimulation in patients with Parkinson's disease. Frontiers in Human Neuroscience 16 (2022). https://doi.org:10.3389/fnhum.2022.958521
Horn, A., Neumann, W.-J., Degen, K., Schneider, G.-H. & Kühn, A. A. Toward an electrophysiological “sweet spot” for deep brain stimulation in the subthalamic nucleus. Human Brain Mapping 38, 3377–3390 (2017). https://doi.org:https://doi.org/10.1002/hbm.23594
Roediger, J. et al. Automated deep brain stimulation programming based on electrode location: a randomised, crossover trial using a data-driven algorithm. Lancet Digit Health 5, e59-e70 (2023). https://doi.org:10.1016/s2589-7500(22)00214-x
Lu, C. W., Chou, K. L. & Patil, P. G. Correspondence of optimal stimulation and beta power varies regionally in STN DBS for Parkinson disease. Parkinsonism & Related Disorders 78, 124–128 (2020). https://doi.org:https://doi.org/10.1016/j.parkreldis.2020.08.008
Akram, H. et al. Subthalamic deep brain stimulation sweet spots and hyperdirect cortical connectivity in Parkinson's disease. Neuroimage 158, 332–345 (2017). https://doi.org:10.1016/j.neuroimage.2017.07.012
Bot, M. et al. Deep brain stimulation for Parkinson's disease: defining the optimal location within the subthalamic nucleus. J Neurol Neurosurg Psychiatry 89, 493–498 (2018). https://doi.org:10.1136/jnnp-2017-316907
Anderson, R. W. et al. Lack of progression of beta dynamics after long-term subthalamic neurostimulation. Ann Clin Transl Neurol 8, 2110–2120 (2021). https://doi.org:10.1002/acn3.51463
Cagnan, H., Denison, T., McIntyre, C. & Brown, P. Emerging technologies for improved deep brain stimulation. Nat Biotechnol 37, 1024–1033 (2019). https://doi.org:10.1038/s41587-019-0244-6
Chen, Y. et al. Neuromodulation effects of deep brain stimulation on beta rhythm: A longitudinal local field potential study. Brain Stimul 13, 1784–1792 (2020). https://doi.org:10.1016/j.brs.2020.09.027
Feldmann, L. K. et al. Subthalamic beta band suppression reflects effective neuromodulation in chronic recordings. Eur J Neurol 28, 2372–2377 (2021). https://doi.org:10.1111/ene.14801
Jenkinson, N. & Brown, P. New insights into the relationship between dopamine, beta oscillations and motor function. Trends in Neurosciences 34, 611–618 (2011). https://doi.org:https://doi.org/10.1016/j.tins.2011.09.003
Shah, A. et al. Combining Multimodal Biomarkers to Guide Deep Brain Stimulation Programming in Parkinson Disease. Neuromodulation 26, 320–332 (2023). https://doi.org:10.1016/j.neurom.2022.01.017
Wang, D. D. et al. Subthalamic local field potentials in Parkinson's disease and isolated dystonia: An evaluation of potential biomarkers. Neurobiol Dis 89, 213–222 (2016). https://doi.org:10.1016/j.nbd.2016.02.015
Lofredi, R. et al. Pallidal Beta Activity Is Linked to Stimulation-Induced Slowness in Dystonia. Mov Disord 38, 894–899 (2023). https://doi.org:10.1002/mds.29347
Neumann, W. J. et al. Long term correlation of subthalamic beta band activity with motor impairment in patients with Parkinson's disease. Clin Neurophysiol 128, 2286–2291 (2017). https://doi.org:10.1016/j.clinph.2017.08.028
Kühn, A. A., Kupsch, A., Schneider, G. H. & Brown, P. Reduction in subthalamic 8–35 Hz oscillatory activity correlates with clinical improvement in Parkinson's disease. Eur J Neurosci 23, 1956–1960 (2006). https://doi.org:10.1111/j.1460-9568.2006.04717.x
Little, S., Pogosyan, A., Kuhn, A. A. & Brown, P. β band stability over time correlates with Parkinsonian rigidity and bradykinesia. Exp Neurol 236, 383–388 (2012). https://doi.org:10.1016/j.expneurol.2012.04.024
Ray, N. J. et al. Local field potential beta activity in the subthalamic nucleus of patients with Parkinson's disease is associated with improvements in bradykinesia after dopamine and deep brain stimulation. Exp Neurol 213, 108–113 (2008). https://doi.org:10.1016/j.expneurol.2008.05.008
Feldmann, L. K. et al. Toward therapeutic electrophysiology: beta-band suppression as a biomarker in chronic local field potential recordings. NPJ Parkinsons Dis 8, 44 (2022). https://doi.org:10.1038/s41531-022-00301-2
Lewis, S. et al. Pilot Study to Investigate the Use of In-Clinic Sensing to Identify Optimal Stimulation Parameters for Deep Brain Stimulation Therapy in Parkinson's Disease. Neuromodulation (2023). https://doi.org:10.1016/j.neurom.2023.01.006
Tinkhauser, G. et al. Beta burst dynamics in Parkinson's disease OFF and ON dopaminergic medication. Brain 140, 2968–2981 (2017). https://doi.org:10.1093/brain/awx252
Giannini, G. et al. Case report: Bilateral double beta peak activity is influenced by stimulation, levodopa concentrations, and motor tasks, in a Parkinson's disease patient on chronic deep brain stimulation. Front Neurol 14, 1163811 (2023). https://doi.org:10.3389/fneur.2023.1163811
Plate, A. et al. Peaks in the beta band of the human subthalamic nucleus: a case for low beta and high beta activity. J Neurosurg 136, 672–680 (2022). https://doi.org:10.3171/2021.3.Jns204113
Tinkhauser, G. et al. Beta burst coupling across the motor circuit in Parkinson's disease. Neurobiol Dis 117, 217–225 (2018). https://doi.org:10.1016/j.nbd.2018.06.007
van Wijk, B. C. et al. Subthalamic nucleus phase-amplitude coupling correlates with motor impairment in Parkinson's disease. Clin Neurophysiol 127, 2010–2019 (2016). https://doi.org:10.1016/j.clinph.2016.01.015
Chen, C. C. et al. Subthalamic nucleus oscillations correlate with vulnerability to freezing of gait in patients with Parkinson's disease. Neurobiol Dis 132, 104605 (2019). https://doi.org:10.1016/j.nbd.2019.104605
Toledo, J. B. et al. High beta activity in the subthalamic nucleus and freezing of gait in Parkinson's disease. Neurobiol Dis 64, 60–65 (2014). https://doi.org:10.1016/j.nbd.2013.12.005
Fischer, P. et al. Alternating Modulation of Subthalamic Nucleus Beta Oscillations during Stepping. J Neurosci 38, 5111–5121 (2018). https://doi.org:10.1523/JNEUROSCI.3596-17.2018
Darcy, N. et al. Spectral and spatial distribution of subthalamic beta peak activity in Parkinson's disease patients. Exp Neurol 356, 114150 (2022). https://doi.org:10.1016/j.expneurol.2022.114150
Cagle, J. N. et al. Suppression and Rebound of Pallidal Beta Power: Observation Using a Chronic Sensing DBS Device. Front Hum Neurosci 15, 749567 (2021). https://doi.org:10.3389/fnhum.2021.749567
Marceglia, S. et al. Double-blind cross-over pilot trial protocol to evaluate the safety and preliminary efficacy of long-term adaptive deep brain stimulation in patients with Parkinson's disease. BMJ Open 12, e049955 (2022). https://doi.org:10.1136/bmjopen-2021-049955
Swinnen, B. E. K. S., Buijink, A. W., Piña-Fuentes, D., de Bie, R. M. A. & Beudel, M. Diving into the subcortex: The potential of chronic subcortical sensing for unravelling basal ganglia function and optimization of deep brain stimulation. NeuroImage 254, 119147 (2022). https://doi.org:https://doi.org/10.1016/j.neuroimage.2022.119147
Strelow, J. N. et al. Local Field Potential-Guided Contact Selection Using Chronically Implanted Sensing Devices for Deep Brain Stimulation in Parkinson's Disease. Brain Sci 12 (2022). https://doi.org:10.3390/brainsci12121726
Thenaisie, Y. et al. Towards adaptive deep brain stimulation: clinical and technical notes on a novel commercial device for chronic brain sensing. J Neural Eng 18 (2021). https://doi.org:10.1088/1741-2552/ac1d5b
Jamora, R. D. G. & Miyasaki, J. M. Treatment gaps in Parkinson's disease care in the Philippines. Neurodegener Dis Manag 7, 245–251 (2017). https://doi.org:10.2217/nmt-2017-0014
Sarica, C. et al. Trends and disparities in deep brain stimulation utilization in the United States: a Nationwide Inpatient Sample analysis from 1993 to 2017. Lancet Reg Health Am 26, 100599 (2023). https://doi.org:10.1016/j.lana.2023.100599
Goyal, A. et al. The development of an implantable deep brain stimulation device with simultaneous chronic electrophysiological recording and stimulation in humans. Biosensors and Bioelectronics 176, 112888 (2021). https://doi.org:https://doi.org/10.1016/j.bios.2020.112888
Jimenez-Shahed, J. Device profile of the percept PC deep brain stimulation system for the treatment of Parkinson's disease and related disorders. Expert Rev Med Devices 18, 319–332 (2021). https://doi.org:10.1080/17434440.2021.1909471
Vaou, O. E. et al. Symptom optimization through sensing local field potentials: Balancing beta and gamma in Parkinson's disease. Deep Brain Stimulation 1, 5–14 (2023). https://doi.org:https://doi.org/10.1016/j.jdbs.2023.01.001
Parameshwaran, D. & Thiagarajan, T. C. Characterizing peaks in the EEG power spectrum. Biomedical Physics & Engineering Express 5, 045023 (2019). https://doi.org:10.1088/2057-1976/ab29d0
Weber, J., Klein, T. & Abeln, V. Shifts in broadband power and alpha peak frequency observed during long-term isolation. Sci Rep 10, 17987 (2020). https://doi.org:10.1038/s41598-020-75127-0
de Solages, C., Hill, B. C., Koop, M. M., Henderson, J. M. & Bronte-Stewart, H. Bilateral symmetry and coherence of subthalamic nuclei beta band activity in Parkinson's disease. Exp Neurol 221, 260–266 (2010). https://doi.org:10.1016/j.expneurol.2009.11.012
Donoghue, T. et al. Parameterizing neural power spectra into periodic and aperiodic components. Nature Neuroscience 23, 1655–1665 (2020). https://doi.org:10.1038/s41593-020-00744-x
Wiest, C. et al. The aperiodic exponent of subthalamic field potentials reflects excitation/inhibition balance in Parkinsonism. Elife 12 (2023). https://doi.org:10.7554/eLife.82467
Conway, J. R., Lex, A. & Gehlenborg, N. UpSetR: an R package for the visualization of intersecting sets and their properties. Bioinformatics 33, 2938–2940 (2017). https://doi.org:10.1093/bioinformatics/btx364
Brittain, J. S. & Brown, P. Oscillations and the basal ganglia: motor control and beyond. Neuroimage 85 Pt 2, 637–647 (2014). https://doi.org:10.1016/j.neuroimage.2013.05.084
He, S. et al. Subthalamic beta-targeted neurofeedback speeds up movement initiation but increases tremor in Parkinsonian patients. eLife 9, e60979 (2020). https://doi.org:10.7554/eLife.60979
Lange, F. et al. Reduced Programming Time and Strong Symptom Control Even in Chronic Course Through Imaging-Based DBS Programming. Front Neurol 12, 785529 (2021). https://doi.org:10.3389/fneur.2021.785529
Peeters, J. et al. EEG-based biomarkers for optimizing deep brain stimulation contact configuration in Parkinson’s disease. Frontiers in Neuroscience 17 (2023). https://doi.org:10.3389/fnins.2023.1275728
Dale, J., Schmidt, S. L., Mitchell, K., Turner, D. A. & Grill, W. M. Evoked potentials generated by deep brain stimulation for Parkinson's disease. Brain Stimul 15, 1040–1047 (2022). https://doi.org:10.1016/j.brs.2022.07.048
Binder, T. et al. Feasibility of local field potential-guided programming for deep brain stimulation in Parkinson's disease: A comparison with clinical and neuro-imaging guided approaches in a randomized, controlled pilot trial. Brain Stimulation 16, 1243–1251 (2023). https://doi.org:https://doi.org/10.1016/j.brs.2023.08.017
Fasano, A. et al. Local Field Potential-Based Programming: A Proof-of-Concept Pilot Study. Neuromodulation: Technology at the Neural Interface 25, 271–275 (2022). https://doi.org:https://doi.org/10.1111/ner.13520
Radcliffe, E. M. et al. Oscillatory beta dynamics inform biomarker-driven treatment optimization for Parkinson's disease. J Neurophysiol 129, 1492–1504 (2023). https://doi.org:10.1152/jn.00055.2023
Yin, Z. et al. Local field potentials in Parkinson's disease: A frequency-based review. Neurobiology of Disease 155, 105372 (2021). https://doi.org:https://doi.org/10.1016/j.nbd.2021.105372
Neumann, W. J. et al. Toward Electrophysiology-Based Intelligent Adaptive Deep Brain Stimulation for Movement Disorders. Neurotherapeutics 16, 105–118 (2019). https://doi.org:10.1007/s13311-018-00705-0
Neuville, R. S. et al. Differential Effects of Pathological Beta Burst Dynamics Between Parkinson’s Disease Phenotypes Across Different Movements. Frontiers in Neuroscience 15 (2021). https://doi.org:10.3389/fnins.2021.733203
Hirschmann, J. et al. A direct relationship between oscillatory subthalamic nucleus-cortex coupling and rest tremor in Parkinson's disease. Brain 136, 3659–3670 (2013). https://doi.org:10.1093/brain/awt271
Timmermann, L. et al. The cerebral oscillatory network of parkinsonian resting tremor. Brain 126, 199–212 (2003). https://doi.org:10.1093/brain/awg022

Table 1

Descriptions of each method and thresholding approaches.
Method		Description	Alternative Threshold	Bland-Altman p-value (MSE) [95% Conf. Int]
Absolute	I [Jimenez-Shahed, 2021; Vaou et al., 2023]	Medtronic’s Percept PC BrainSense Survey (BSS) uses a static peak detection threshold of > 1.1µVp. Other peaks are subjectively determined.		I 0.566 (0.035) [0.000, 0.083]
Absolute	II [Giannini et al., 2023]	Local maxima are identified within the lower (13-20Hz) and upper (20-30Hz) beta ranges.		II 8.820^− 10 [3.608, 8.008] (38.0630)
Algebraic	III [Darcy et al., 2022]	With normalized PSD signatures, a dynamic threshold is set for findpeaks’ MinPeakProminence of median(PSD_norm) / divisor. Iteratively run until an optimal divisor value is found.	IV [Parameshwaran & Thiagarajan, 2019; Weber et al., 2020] Dynamic threshold of std(PSD_norm) * multiple is used.	III 0.379 (4.590) [2.184, 7.478]	IV 0.091 (6.061) [3.258, 9.278]
Algebraic	V [Plate et al., 2021; de Solages et al., 2010]	Beta peak is defined as a local elevation of power such that a peak’s 4–6 adjacent points must be 20% greater than the means of the outer 8–12 points.		V 0.617 (3.342) [1.656, 5.374]
Elemental Decomposition	VI [Donoghue et al., 2020; Wiest et al., 2023]	Power spectra are deconstructed into aperiodic (1/f trend) and periodic components using a series of Gaussian fits. Peaks and prominences are fit based on these fitted models.	[Wilson et al., 2022] The aperiodic component is treated as temporally dynamic instead of static, and averaged before peak identification.	VI 4.140^− 6 (8.685) [6.143, 11.667]
	VII [Weber et al., 2020]	The aperiodic component is defined by the linearized Lorentzian function \({10}^{b}\frac{1}{(k+ {F}^{x})}\) where b = offset; k = knee parameter (k = 0); F = frequency vector; x = aperiodic exponent. This is subtracted from the PSD signature. A threshold of mean(PSD) + 3SD(PSD) is used to identify peaks/	VIII [Parameshwaran & Thiagarajan, 2019] An alternate dynamic threshold of mean + 1SD is used to generate a Heaviside-like function with which peak indices are defined.	VII 0.002 (1.907) [0.541, 3.739]	VIII 0.010 (5.200) [2.822, 7.982]
	VII [Weber et al., 2020]		IX [Darcy et al., 2022] An alternate dynamic threshold approach from III is used on non-normalized-PSD.	VII 0.002 (1.907) [0.541, 3.739]	IX 0.178 (3.697) [1.434, 6.394]
	X [Wang et al., 2016]	log(PSD) is fit to a 5th order polynomial using the local minima in the lower frequency range (3-5Hz) and values above the beta range as a 1/f proxy. The difference between 1/f proxy and the log(PSD) is fit to a Gaussian and peaks identified.		X 3.470^− 8 (8.486) [5.132, 12.291]

There is a conflict of interest Funding to JAT and DK provided by Medtronic

Download PDF

Journal Publication

published 09 Aug, 2024

Read the published version in npj Parkinson's Disease →

Editorial decision: revise
16 Apr, 2024
Review #2 received at journal
04 Apr, 2024
Review #1 received at journal
27 Mar, 2024
Reviewer #2 agreed at journal
25 Mar, 2024
Reviewer #1 agreed at journal
22 Mar, 2024
Reviewers invited by journal
19 Mar, 2024
Editor assigned by journal
28 Feb, 2024
Submission checks completed at journal
28 Feb, 2024
First submitted to journal
27 Feb, 2024

You are reading this latest preprint version

Towards an objective, standardized beta frequency peak detection algorithm to inform adaptive deep brain stimulation programming in Parkinson’s disease

Status:

Journal Publication

Version 1

Abstract

Figures

1. INTRODUCTION

2. METHODS

3. RESULTS

4. DISCUSSION

Declarations

References

Tables

Additional Declarations

Status:

Journal Publication

Version 1