Chronic adaptive deep brain stimulation versus conventional stimulation in Parkinson's disease: a blinded randomized feasibility trial

Abstract

Deep brain stimulation (DBS) is a widely used therapy for Parkinson's disease (PD) but lacks dynamic responsiveness to changing clinical and neural states. Feedback control might improve therapeutic effectiveness, but the optimal control strategy and additional benefits of 'adaptive' neurostimulation are unclear. Here we present the results of a blinded randomized cross-over pilot trial aimed at determining the neural correlates of specific motor signs in individuals with PD and the feasibility of using these signals to drive adaptive DBS. Four male patients with PD were recruited from a population undergoing DBS implantation for motor fluctuations, with each patient receiving adaptive DBS and continuous DBS. We identified stimulation-entrained gamma oscillations in the subthalamic nucleus or motor cortex as optimal markers of high versus low dopaminergic states and their associated residual motor signs in all four patients. We then demonstrated improved motor symptoms and quality of life with adaptive compared to clinically optimized standard stimulation. The results of this pilot trial highlight the promise of personalized adaptive neurostimulation in PD based on data-driven selection of neural signals. Furthermore, these findings provide the foundation for further larger clinical trials to evaluate the efficacy of personalized adaptive neurostimulation in PD and other neurological disorders. ClinicalTrials.gov registration: NCT03582891 .

Extended Data Fig. 1.. Localization of leads over sensorimotor cortex and within subthalamic nucleus in native space.

a-d, Example localization of cortical and subcortical leads in patient 2, generated by fusing postoperative CT with preoperative MRI scans. Contacts appear as white CT artifacts due to metal content and are labeled with red arrows. a, Cortical leads on axial T1-weighted MRI through the vertex. b, STN leads on axial T2-weighted MRI through the region of the dorsal STN, 3 mm inferior to the intercommissural plane. c-d, Cortical leads on oblique sagittal T1-weighted MRI passing through the long axis of the lead array in left (c) and right (d) hemispheres, respectively. e-h, Location of cortical leads for each patient overlayed on 3D reconstruction of cortex rendered using Locate Electrodes Graphical User Interface (LeGUI). Electrodes used in the anterior and posterior cortical montages are shown in cyan and yellow, respectively. For patient 1 (e), 2 (f) and 4 (h), anterior and posterior montages covered the pre- and postcentral gyrus, respectively. For patient 3, right side (g), the anterior montage included one electrode on the middle frontal and one on the precentral gyrus. The posterior montage comprised one pre- and one postcentral electrode. In all figures, red arrows indicate the location of the central sulcus.

Extended Data Fig. 2.. Initial and finalized adaptive stimulation parameters and example adaptive control policies.

a, Suggested initial parameters for algorithms developed for time scales of minutes to hours, as identified during steps 5 and 6 of the pipeline. An update rate of 10 s typically provided a signal to noise ratio that allowed for adequate discrimination between the presence and absence of the most bothersome symptom, and this could often be improved with a further increase in update rate. The ramp rate chosen for each patient depended on the results of step 5 (we chose an example of 1 mA/s). b, Detailed final adaptive stimulation parameters including control signals, thresholds, FFT interval, update rates, blanking periods, onset and termination duration, and ramp rates used for each patient and hemisphere. c-e, Examples of potential control policies that can be used for an adaptive algorithm, using artificial data. The upper subpanels of each subfigure illustrate an on-state biomarker (blue), as used in our study, along with thresholds (red). Lower subpanels demonstrate the adjustment of stimulation amplitude based on the relationship of the neural signal to the thresholds. c, A single threshold control policy with two stimulation amplitudes. When the biomarker is above the threshold, stimulation amplitude decreases and once below threshold, stimulation amplitude increases. d, A dual threshold control policy with three stimulation amplitudes (not used in this study), which may be applied to address three symptom states. When the neural signal is below both thresholds, the stimulation amplitude is high (e.g., 4 mA). When the biomarker is between the two thresholds, stimulation adjusts to a middle amplitude (e.g., 3 mA). When the biomarker exceeds the second threshold, stimulation decreases to the low amplitude (e.g., 2 mA). e, A control policy utilizing a middle state as noise buffer. Stimulation is high when the control signal is below the bottom threshold and stimulation is low when the control signal is above the top threshold. When the control signal is between the two thresholds, it remains at the level of the stimulation amplitude prior to crossing the threshold (i.e., no changes are made).

Extended Data Fig. 3.. Neural biomarkers of medication effects identified in-clinic.

(a, b) All tables show the results from our within-patient non-parametric cluster-based permutation analyses using in-clinic recordings during two medication states (off vs. on) and stimulation conditions (low vs. high stimulation amplitude). P-values were Bonferroni-corrected for multiple comparisons. Note that p<10⁻³ indicates that the cluster was found in all 1000 permutations. This means the probability of observing this effect by chance is less than 1 in 1000. a, Statistics for the largest main effect of medication, stimulation, and their interaction for each patient and hemisphere when searching the whole frequency space (2-100 Hz) across brain regions. Frequencies represent the center frequency of 1-Hz wide power spectral density bins. For all four patients (five out six hemispheres), we found that gamma power (specifically, stimulation-entrained gamma in four hemispheres) in the STN or cortex was the best predictor of medication state (in pat-3L, there was no significant effect of medication in any frequency band in clinic, but at home symptom monitoring identified cortical stimulation-entrained gamma power as neural biomarker; Extended Data Fig. 4). Positive Cohen’s d values for the medication effect highlight that the neural biomarker was higher during on-medication states. Positive Cohen’s d values for the stimulation effect indicate that the neural biomarker was higher during on-stimulation states (independent of medication), which could result in undesirable self-triggering of the algorithm (threshold crossing of the neural biomarker linked to stimulation change itself, rather than to true fluctuations of medication states and symptoms). Therefore, for patient 1, we excluded 63 and 67 Hz from the subsequently used control signal (positive Cohen’s d main effect of stimulation). For patients 2, 3 and 4, we did not find stimulation effects that positively modulated biomarkers and therefore were unrestricted in biomarker selection. b, When constraining the anatomic location and frequency space to STN beta oscillations (13-30 Hz), STN spectral beta power was only predictive for medication state in two hemispheres (pat-2R and pat-4) and smaller in effect size than cortical/STN stimulation-entrained gamma oscillations for all patients.

Extended Data Fig. 4.. Neural biomarkers of symptoms identified at-home.

We identified predictors of the most bothersome symptom (pat-1: bradykinesia, pat-2: lower limb dystonia), or the opposite symptom that limits the therapeutic window (pat-3 and pat-4: dyskinesia). a, Heatmaps of t-values derived from stepwise linear regressions using 1 Hz power bands between 2-100 Hz in the STN (left), anterior cortical montage (middle) and posterior cortical montage (right) to predict symptoms measured with continuous upper extremity wearable monitors for patients 1, 3 and 4 (patient 2’s bothersome symptom did not involve the upper extremity). b-d, Results from the linear regression (left) and linear discriminant analysis (LDA; right). P-values were Bonferroni-corrected for multiple comparisons (289 predictors). b, Both methods provide converging evidence that stimulation-entrained gamma power centered at half the stimulation frequency (65 Hz) in the STN and cortex optimally distinguishes hypo- and hyperkinetic symptoms. c, When constraining the anatomic location and frequency space to STN beta oscillations (13-30 Hz), frequency bands identified as most predictive were less discriminative than cortical/STN stimulation-entrained gamma oscillations (LDA: AUC<0.7). Regression models resulted in smaller magnitude coefficients, with only one hemisphere demonstrating a significant negative association with hyperkinetic symptoms (pat-3L). d, STN beta frequency bands were also poorly predictive of wearable bradykinesia scores (AUC<0.6), again with only one hemisphere demonstrating a significant effect in the regression model (corresponding to positive relationship with hypokinetic symptoms; pat-3L). e, Comparison of LDA results for STN and cortical gamma activity in predicting bothersome symptoms. Neural signals selected for adaptive stimulation are shaded in grey. In three out of six hemispheres (pat-2L, pat-2R, pat-4), stimulation-entrained gamma activity in the STN distinguished between hypo- and hyperkinetic symptoms. For pat-2, STN stimulation-entrained spectral gamma power was the optimal biomarker used for aDBS in both hemispheres. In pat-4, stimulation-entrained gamma activity in the STN was a strong predictor of residual motor signs but slightly underperformed compared to cortical signals. f, Visual illustration of AUC values comparing STN and cortical gamma activity in predicting bothersome symptoms. For pat-4, the predictive value of stimulation-entrained spectral gamma power was only slightly reduced compared to cortical signals.

Extended Data Fig. 5.. Beta oscillations in the STN.

a, Power spectral density in the STN based on in-clinic recordings off medication and off stimulation for all six hemispheres. All but one hemisphere (pat-1) exhibited a peak in the beta frequency band (illustrated in yellow). b, Example of the suppressive effect of DBS on STN beta oscillations precluding use of beta band activity as a biomarker of medication state during active stimulation (pat-2L, all data collected during the same in-clinic recording session). Off stimulation, the spectral peak in the beta frequency range was suppressed by medication (13-21 Hz, Cohens’ d=−1.09, p<10⁻³). However, this medication effect diminished during active stimulation, even at low stimulation amplitudes (1.8mA, largest effect in the beta band: 15-18 Hz, Cohens’ d=0.31, p=0.026). Data are corrected for stimulation-induced broadband shifts.

Extended Data Fig. 6.. Effects of aDBS and cDBS on most bothersome symptom severity, additional motor symptoms, and sleep quality.

a-j, Bar plots illustrating the mean (±SEM) self-reported symptoms, aside from the most bothersome symptoms, across testing days. Each dot represents the rating for one testing day (blue: cDBS, red: aDBS). These ratings constituted secondary outcome measures to ensure that we are not aggravating other motor and non-motor symptoms. a-b, Patient self-reported motor symptom severity from daily questionnaires (1=least severe, 10=most severe). Note that patients rated symptom severity (shown here) independently of symptom duration; bar graphs for the latter are in Fig 5 a,b. (Patient 3 did not record ratings within the instructed range of 1-10 and their data are therefore not reported.) a, In addition to a decrease in the amount of daily hours with the most bothersome symptom (symptom duration, shown in Fig. 5a), patients 1, 2, and 4 also experienced a significant improvement of symptom severity (pat-1: p<10⁻⁵, pat-2: p=0.018, pat=4: p=0.003). b, No subject reported worsened severity of their opposite symptom (pat-1: p=0.18, pat-2: p=1, pat-4: p=0.19). c-h, Comprehensive list of the self-reported duration of motor symptoms from daily questionnaires. These bar graphs illustrate only symptoms that were not identified by the patient as the most bothersome or as the opposite symptom. For each patient’s most bothersome symptom, results are displayed in Figures 5a and panel a of this figure; and are labeled in c-h as not applicable (n/a). None of these “other” motor symptoms were worsened by aDBS, and patient 2 demonstrated significant improvement in the percentage of waking hours with dyskinesia (d, p=0.044) and gait disturbance (h, p<10⁻⁴). i-j, Self-reported sleep quality (1=poorest sleep, 10=best sleep) and duration from daily questionnaires. aDBS provided no change in patients’ sleep characteristics. The number of testing days for each patient and condition used for statistical tests are summarized in Fig. 6a. Asterisks illustrate results from two-sided Wilcoxon rank sum tests. P-values for all within-subject control analyses were adjusted for multiple comparisons using the false discovery rate procedure and are indicated as: *p<0.05, **p<0.01, ***p<0.001.

Extended Data Fig. 7.. aDBS algorithm dynamics during nighttime.

a, Percent time spent at each stimulation amplitude during the night. Each dot represents the mean values of one night of aDBS testing across high stimulation states (orange) and low stimulation states (blue) in one hemisphere. Graphs are standard box plots (center: median; box limits: upper and lower quartiles; whiskers: minima=25th percentile-1.5 times the interquartile range, maxima=75th percentile+1.5 times the interquartile range). Each patient spent most of the night in the high stimulation state. b, Mean (±SEM) total electrical energy delivered (TEED) during aDBS and cDBS overnight, showing increased TEED during aDBS, similar to daytime analyses (stimulation main effect: β=27.7, p<10⁻²⁵, time main effect: β=0.05, p=0.377). Individually, TEED was increased in all hemispheres during aDBS (two-sided, one-sample Wilcoxon signed rank test, pat-1: p<10⁻⁶, pat-2R: p<10⁻⁵, pat-2L: p<10⁻⁵, pat-3R: p<10⁻⁶, pat-3L: p<10⁻⁶, pat-4: p<10⁻⁴). The number of testing nights for each patient and condition used for both illustrations are stated in Fig. 6a and are equivalent to the testing days. Asterisks illustrate results from two-sided one-sample Wilcoxon signed rank tests. P-values for TEED evaluations were adjusted for multiple comparisons using the false discovery rate procedure and are indicated as: *p<0.05, **p<0.01, ***p<0.001.

Extended Data Fig. 8.. Flowchart of biomarker identification analyses.

We identified neural biomarkers using standardized in-clinic and at-home recordings in patients’ naturalistic environments. Non-parametric cluster-based permutation analysis identified candidate spectral biomarkers from in-clinic data by assessing main effects of medication state, stimulation amplitude, and the interaction. Next, the predictability of neural biomarkers as robust aDBS control signals of symptom state was tested using at-home recordings. For patients where the most bothersome symptom was monitored by a wearable device (e.g., upper extremity bradykinesia or dyskinesia), linear stepwise regression was used to take advantage of the continuous nature of the symptom measurements. The most predictive frequency bands and recording sites were selected based on t-values. If the patient’s most bothersome symptom could not be captured by wearable monitors, the patient’s motor diaries and streaming app entries instead labeled the presence of symptoms. A linear discriminant analysis (LDA) based method identified the most predictive frequency band and recording site from these discretely labeled neural signal data, as measured by the area under the receiver operating curve (AUC). We also applied the LDA-based approach to symptoms measured by wearable monitors by mapping the continuous wearable scores to discrete symptom labels using a patient-specific dichotomization. This dichotomization allowed for subsequent offline assessment of the prediction accuracy based on multiple neural biomarkers combined as shown in Fig. 4e (note for online aDBS only single power band classifiers were implemented, as multiple power band classifiers were not found to be superior).

Figure 1.. Configuration of implanted hardware, algorithmic model and patient demographics.

a, Illustration of the adaptive paradigm starting with real-life sensing of brain activity (blue) that reflects changes in patient’s mobility–in this example slowness of movement (bradykinesia). Neural activity is sensed continuously on-board the DBS device from either the STN or sensorimotor cortex using depth or subdural electrodes, respectively. Here, we illustrate an example of a cortical control signal for fully-embedded adaptive implementation. Once a change in the brain signal across a predefined threshold is detected, the stimulation amplitude increases or decreases automatically (red) at the target brain region (STN). This adaptation of stimulation amplitude to the patient’s needs leads to improved symptoms–in this example, faster movement. b, Localization of depth leads in the STN with active contacts colored in red across patients in normalized Montreal Neurological Institute (MNI) space. STN is highlighted in orange and the red nucleus in red. c, Location of cortical leads that entered the aDBS pipeline for all patients in normalized MNI space overlaid on a common brain atlas. d, Patient characteristics including age, gender, disease duration, UPDRS-III off medication score, pre- and post-surgery levodopa equivalent daily dose (LEDD, mg) and residual motor fluctuations on clinically optimized cDBS, including the body side, the most bothersome symptom and the symptom in the opposite dopaminergic state. e, Timeline of the study protocol, including optimization of cDBS by a movement disorder neurologist, biomarker identification, aDBS algorithm design, and blinded comparisons between effects of aDBS and cDBS on symptoms. Stimulation conditions were applied for at least one month each in short, randomized blocks of 2-7 days.

Figure 2.. Workflow for data-driven biomarker identification and aDBS implementation.

We employed a seven-step workflow, individualized for each patient: Identification of bothersome residual symptoms on cDBS and required stimulation amplitude limits for better symptom control (steps 1 and 2), in-clinic and at-home neural recordings with simultaneous symptom monitoring for biomarker identification (steps 3-4), refining parameters for patient-tailored adaptive algorithms using supervised short-term (step 5) and longer-term (step 6) at-home testing, and finally blinded, randomized comparisons between cDBS and aDBS in multiple blocks of 2-7 days per condition (total of one month per condition) in patients’ real-life environments (step 7).

Figure 3.. Examples of stimulation-entrained gamma oscillations in both in-clinic and at-home recordings.

a, Example spectrogram of cortical activity in the high dopaminergic state during systematic variations in stimulation amplitude (black dotted line), illustrating the phenomenon of stimulation-induced entrainment of gamma oscillations at half of the stimulation frequency (130 Hz, pat-1). Levodopa-induced finely-tuned gamma oscillations occur at 80-90 Hz when stimulation is off or stimulation amplitudes are low but become entrained to half the stimulation frequency (65 Hz) when stimulation exceeds a certain amplitude (1.5 mA in this example). b-c, Examples of biomarker identification using standardized in-clinic neural recordings (b, pat-2L, c, pat-1). Plots show power spectra during high and low dopaminergic states (labeled on and off medication, mean±SEM), i.e., periods during which hyper- and hypokinetic symptoms would emerge, respectively. Recordings are collapsed across low and high stimulation amplitude conditions, which were both amplitudes at which finely-tuned gamma oscillations entrained to half the stimulation frequency. We found that medication yielded the largest effect on stimulation-entrained gamma power at half the stimulation frequency in the STN (b, pat-2L) and motor cortex (c, pat-1, anterior montage) when controlling for effects of stimulation (Extended Data Fig. 2). Significant clusters are highlighted in gray. d-e, At-home recordings during constant stimulation amplitude and patients’ normal medication schedule in the STN (d, pat-2L) and motor cortex (e, pat-1, anterior montage). Patients marked their medication intake (red dashed line) and on- and off-set of their most bothersome symptom in their motor diary (completed in 30-minute intervals) and the streaming application. Both patients had bothersome hypokinetic symptoms associated with low dopaminergic states, i.e., lower limb-dystonia (d, pat-2) and bradykinesia (e, pat-1). For both, stimulation-entrained gamma oscillations occur ~45 minutes after medication intake (red arrows), corresponding to a typical latency of onset for dopaminergic medication. When the patient marked their most bothersome symptom (indicated by the black dashed line) in their motor diary, stimulation-entrained gamma oscillations disappeared, indicating a transition to a low dopaminergic state.

Figure 4.. Data-driven biomarker identification during active stimulation for all hemispheres.

a-b, Results of the within-subject nonparametric cluster-based permutation analysis for in-clinic recordings. a, Graphs illustrate the effect size (Cohen’s d) of the main effect of medication on spectral power of neural oscillations as a function of frequency in the STN (left), the anterior cortical montage (middle), and posterior cortical montage (right) for all patients. Red and blue colors represent positive effects (high>low dopaminergic state) and negative effects (low>high dopaminergic state), respectively. For all patients, we found stimulation-entrained gamma oscillations in the STN (pat-2, both hemispheres) and cortex (pat-1, left hemisphere, pat-3, right hemisphere and pat-4, left hemisphere) at half the stimulation frequency (130 Hz) to be the optimal biomarker for medication-related fluctuations during active stimulation (Extended Data Fig. 3). We did not find any significant effects in the left hemisphere of pat-3 (not shown). b, The effect sizes for cortical and STN stimulation-entrained gamma oscillations (right) were superior to those for STN beta oscillations (left) for all patients (mean±SEM across permutations; one-sided Wilcoxon signed rank test: p=0.031; Extended Data Fig. 3). c-e, Results of the within-subject linear discriminant analysis for at-home recordings using power spectral density at the three brain sites to predict the occurrence of the most bothersome or opposite symptom. c, The three graphs illustrate the initial area under the curve (AUC) prior to bandwidth optimization as a function of frequency for the STN (left), the anterior cortical montage (middle), and posterior cortical montage (right) for all patients. d, Across patients, we show that stimulation-entrained gamma oscillations in the STN (pat-2, both hemispheres) or cortex (pat-1, left hemisphere, pat-3, both hemispheres, and pat-4 left hemisphere) were the best predictors of the occurrence of the most bothersome or opposite symptom and superior to beta oscillations (mean±SEM across permutations; one-sided Wilcoxon signed rank test: p=0.016; Extended Data Fig. 4). e, The combined use of STN/cortical gamma and STN beta bands provided minimal improvement in the AUC of at-home symptom prediction using linear discriminant analysis (mean±SEM across permutations).

Figure 5.. Effects of aDBS compared to cDBS on both subjective and objective metrics of motor symptoms and quality of life.

a-c, Self-reported symptom duration from daily questionnaires for each subject. Every bar represents the mean (±SEM) across testing days per condition and patient. Each dot represents the rating for one testing day (blue: cDBS, red: aDBS). aDBS resulted in a decreased percentage of awake hours experiencing the most bothersome symptom (a) without exacerbating the opposite symptom (b). For visualization, Pat-3’s six outliers (100% awake hours, five with cDBS, one with aDBS) are not plotted in (a). Quality of life, as measured by the EQ-5D, was improved for three of four patients. Patient 3 reported very high quality of life scores, with minimal variance for both cDBS and aDBS. d-g, Effect of DBS condition across multiple motor signs (mean±SEM across testing days) illustrated in radar plots. Personalized bothersome and opposite symptoms are bold with the most bothersome symptom underlined. Control analyses showed that adaptive stimulation did not worsen any other motor symptoms, but instead patient 2 experienced decreased time with gait disturbance and dyskinesia (p<10⁻⁴ and p=0.044, respectively, blue asterisks). All other motor symptoms were not affected (Extended Data Fig. 6a-h, Supplementary Table 1). Further, aDBS did not adversely affect sleep or any other monitored non-motor symptoms (Extended Data Fig. 6i-j, Supplementary Table 2). Note the subject-specific axis scales. h-i, Wearable monitor scores (mean±SEM fluctuation scores across testing days) demonstrating the decreases in symptom intensity fluctuations between low- and high-dopaminergic states. Only patients 1, 3 and 4 are displayed, as patient 2’s bothersome and opposite symptoms were not measurable by a wearable device. Laterality refers to the brain hemisphere where aDBS was applied (and therefore contralateral motor sign measurement). The number of testing days for each patient used for statistical tests are summarized in Fig. 6a (wearables). Asterisks illustrate results from two-sided Wilcoxon rank sum tests. P-values for within-subject control analyses (d-g) were adjusted for multiple comparisons using the false discovery rate procedure and are indicated as: *p<0.05, **p<0.01, ***p<0.001.

Figure 6.. Characteristics and technical performance of adaptive DBS algorithms.

a, Final stimulation parameters used for blinded, randomized comparisons between stimulation conditions including the control signal for aDBS. All parameters except stimulation amplitude were identical between cDBS and aDBS. b-c, Examples of two control algorithms using subthalamic (b, pat-2R) and cortical (c, pat-1) stimulation-entrained gamma activity at half the stimulation frequency as control signals. The upper subplots illustrate the control signal as a function of time with thresholds (black) that are used to determine changes in stimulation amplitude. The lower subpanels illustrate the stimulation amplitudes responding to fluctuations in the neural signal. In all patients, we used a biomarker indicating high dopaminergic states, such that stimulation amplitude increases when the biomarker amplitude dropped below and decreases when the biomarker amplitude exceeds a threshold. Note that stimulation-entrained gamma activity persisted at low stimulation amplitudes well after the reduction of stimulation amplitude in the on medication state. Timing of dopaminergic medication intake is marked by dashed red vertical lines. d,e, Dynamics of algorithm performance showing adaptive changes on a time course of minutes-hours. Each dot represents the mean values of one day of aDBS testing across high stimulation states (orange) and low stimulation states (blue) in one hemisphere. Graphs are standard box plots (center: median; box limits: upper and lower quartiles; whiskers: minima=25th percentile-1.5 times the interquartile range, maxima=75th percentile+1.5 times the interquartile range). d, Duration of each stimulation amplitude state in a day. Pat-3’s left hemisphere’s high stimulation amplitude state has three outliers not currently plotted which include 4.32, 5.15, and 7.32 hours. e, Percent time spent at each stimulation amplitude. f, Mean (±SEM) total electrical energy delivered (TEED) during aDBS and cDBS across testing days and hemispheres. During awake hours, aDBS resulted in greater TEED compared to cDBS during the day. The number of testing days for each patient and condition used for (d-f) are stated in (a). Asterisks illustrate results from two-sided one-sample Wilcoxon signed rank tests. P-values for TEED evaluations were adjusted for multiple comparisons using the false discovery rate procedure and are indicated as: *p<0.05, **p<0.01, ***p<0.001.