Pith. sign in

REVIEW 3 major objections 4 minor 34 references

Bandit Algorithms for Deep Brain Stimulation

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A training-free multi-armed bandit can jointly tune deep brain stimulation frequency and amplitude, converge on implant-class microcontrollers in under two minutes, and beat a deep-RL baseline at suppressing the beta-band biomarker while us

desk verdict A solid MAB-for-DBS paper with real MCU power measurements, but the 'outperforms deep-RL' claim is confounded by a mismatched training objective and needs fixing. read the letter →

arxiv 2601.12699 v2 pith:NC26SZOH submitted 2026-01-19 cs.LG cs.SYeess.SY

Arkaprava Gupta , Nicholas Carter , William Zellers , Prateek Ganguli , Benedikt Dietrich , Vibhor Krishna , Parasara Sridhar Duggirala , Samarjit Chakraborty This is my paper
classification cs.LGcs.SYeess.SY
keywords DeepbrainstimulationMulti-armedbanditAdaptiveDBSParkinson'sdiseaseBeta-bandbiomarkerEmbeddedsystemsEnergy-efficientreinforcementlearningFrequency-amplitudetuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that adaptive deep brain stimulation (DBS) does not need heavy deep-reinforcement-learning controllers. It proposes a simple multi-armed bandit, T3P, that jointly tunes stimulation frequency and amplitude without any offline training. In a computational model of Parkinson's disease, the authors report faster convergence than other bandit methods and better suppression of the pathological beta-band biomarker than a deep-RL baseline, while using less stimulation power. They also show the algorithm runs on microcontroller-class hardware and converges in under two minutes. If true, this is a concrete path toward personalized, energy-efficient implants that clinicians can understand and adjust.

What carries the argument

The load-bearing component is the pruning rule inside T3P: after a single play of each arm, the algorithm keeps only the K arms with the highest observed rewards and discards the rest, because the authors observed that below-average arms never became optimal in their simulations. This reduction of the search space from 31 arms to K, combined with an epsilon-greedy policy that only starts decaying after warm-up, is what produces the fast convergence. A second mechanism, re-triggering the warm-up on a timer or on a threshold change in the beta-band biomarker, is meant to adapt to slow or unobservable brain-state changes. The reward function combines beta-band power (negative weight), the prese

What would settle it

Run T3P in the same basal ganglia-thalamic model but introduce a state change after convergence (for example, shift the dopamine-related applied current or the beta-band baseline) so that a previously pruned arm becomes optimal; if T3P does not select the new optimum before its timer re-triggers—or if it repeatedly fails from a variety of initial conditions—the pruning claim is contradicted. A second check: compare T3P against unpruned epsilon-greedy on many random state-change sequences; the paper predicts T3P's cumulative regret stays lower.

Watch

Extended reading notes

Core claim

The central claim is that a warm-up-and-prune extension of epsilon-greedy—the Time- and Threshold-Triggered Pruned Multi-Armed Bandit (T3P)—can find the best frequency–amplitude combination for DBS faster than competing bandits and more effectively than a state-of-the-art deep-RL agent. After playing every one of the 31 candidate arms once, T3P discards all but the top K arms, then runs epsilon-greedy among the survivors; it repeats this warm-up only after a countdown timer or a threshold-level change in the beta-band biomarker. In the basal ganglia-thalamic simulation, T3P converged to the same arm (155 Hz, 1000 µA/cm²) in nearly all runs, and the authors report that this setting suppresses

Load-bearing premise

The claim depends on the assumption that, after one play of each arm, no below-average arm can ever become the optimal one; if a pruned arm becomes best later, T3P cannot select it until the warm-up is re-triggered.

Editorial extensions

If this is right

  • An implanted DBS device could personalise stimulation in vivo without an offline training session: the bandit finds the frequency–amplitude setting within about two minutes of turning on.
  • Because the controller is a small table of reward estimates rather than a neural network, the battery budget for inference drops, and the device can be re-targeted by a clinician who simply removes or limits arms.
  • Jointly tuning frequency and amplitude, rather than frequency alone, becomes practical with the same simple machinery, widening the parameter space that adaptive DBS can explore.
  • The same warm-up–prune–re-trigger scheme could transfer to other closed-loop implant problems where a small number of stimulation parameters must be set under tight energy constraints.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the pruning rule is the fragile part; if a pruned below-average arm becomes optimal because of disease progression, medication, or noise, T3P cannot recover it until a re-trigger. A natural test is to run T3P on a non-stationary version of the BGT model with changing dopamine input and compare cumulative regret against an unpruned epsilon-greedy.
  • Editorial inference: the energy measurements suggest the biomarker FFT, not the bandit itself, dominates compute; reducing the 100 kHz sampling rate or replacing the full FFT with a cheaper band-power estimator could cut power further, something the paper leaves to future work.
  • Editorial inference: the comparison against deep-RL hinges on the same BGT model and on the authors' beta-band computation; in real patients, where LFP noise is higher and the reward signal less clean, the sample-efficiency advantage of MAB could narrow or invert.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes T3P, a pruned epsilon-greedy multi-armed bandit controller for adaptive deep brain stimulation (DBS) that jointly selects stimulation frequency and amplitude. The reward is based on the beta-band power spectral density (P_beta), stimulation-off periods, and energy consumption. The algorithm is evaluated on a computational basal ganglia-thalamic (BGT) model against eight MAB baselines and a TD3 deep-RL baseline taken from the authors' prior work. The paper also reports energy measurements from running T3P on ESP32-S3 and ESP32-P4 microcontrollers, claiming convergence in under two minutes and suitability for implantable hardware. The central claims are that T3P converges faster than existing MAB methods and outperforms deep-RL baselines in suppressing P_beta while using less stimulation power.

Significance. If the claims were fully supported, the contribution would be significant: a training-free, transparent, joint frequency-amplitude DBS controller that converges quickly on commodity microcontrollers would be a practical step toward adaptive, personalized DBS. The hardware implementation with Joulescope measurements, the direct comparison among MAB variants, and the explicit reward design are clear strengths. However, the headline comparison against deep RL is confounded by differences in the reward function, biomarker computation, timestep, and baseline provenance, so the 'outperforms deep-RL' claim is not yet established. The MAB-vs-MAB comparison and hardware feasibility results are the most solid parts of the paper.

major comments (3)
  1. [§5.4, Table 1] The claim that T3P 'outperforms deep-RL baselines' is not supported by a controlled comparison. The TD3 baseline from [4] was trained with a reward based on the PSD of S_Gi in the 1–20 Hz band plus energy, a 100 ms timestep, and a different P_beta computation (individual neuron membrane potentials), whereas T3P optimizes P_beta (13–35 Hz) computed from the LFP over 1000 ms rounds. The paper acknowledges these differences, but still reports lower P_beta for T3P in Table 1. This difference could reflect the objective mismatch rather than algorithmic superiority. To support the abstract's claim, the authors should re-run TD3 under matched conditions (same reward, same P_beta computation, same episode length) or explicitly limit the claim to 'comparable under different objectives.'
  2. [§4.3, Algorithm 1] The pruning rule is load-bearing but rests on two unvalidated empirical observations from one computational model: (i) that a single play per arm during warm-up sufficiently estimates each arm's reward, and (ii) that arms with below-average rewards in warm-up never become optimal. The paper itself motivates re-triggering because the brain state can change, so a below-average arm in the initial state could become optimal after a state change. After pruning, the algorithm cannot recover such an arm until a re-trigger restarts the warm-up. The paper provides no sensitivity analysis, no noisy-reward experiments, and no formal guarantee. This assumption should be stress-tested across state changes, patient heterogeneity, and reward noise before the algorithm's convergence and adaptability claims are accepted.
  3. [§5.3, Fig. 6] The regret analysis identifies the 'optimal arm' as (155 Hz, 1000 µA/cm²) 'based on empirical observations' from the same simulation framework used to evaluate the algorithms. No independent method (e.g., a longer exhaustive grid search) is described, and the reported results are averages over 10 runs without error bars or confidence intervals. This weakens the quantitative claim that T3P converges faster than other MAB algorithms. Please report variance across runs and describe how the optimal arm was determined independently of the learning runs.
minor comments (4)
  1. [§5.3, Fig. 5] The caption says 'The values on the x-axis are the average instantaneous rewards,' but the x-axis is labeled 'Rounds' and the y-axis is 'Instantaneous Reward.' Please correct the wording.
  2. [Algorithm 1] The symbol δ is used both for the epsilon decay step and for the re-trigger threshold in the algorithm input list, which is confusing. Use different names, e.g., δ_epsilon and δ_threshold.
  3. [Table 2] The text states each round is 1000 ms, but the measured round lengths in Table 2 range from 1.1 to 2.4 s. Please clarify whether the difference is due to processing overhead or whether the algorithm's decision interval is longer than stated.
  4. [References] References [9] and [10] are identical; one should be removed or replaced with the correct distinct citation. Also fix the typo in the header of Table 3 ('W AND' should be 'WAND').

Circularity Check

2 steps flagged · score 4.0 of 10

T3P itself is a genuine algorithm evaluated empirically, but the headline 'outperforms deep-RL baselines' rests on a same-group TD3 baseline from [4] with a different reward/biomarker, and the regret oracle is fitted to the same simulation data.

  1. self citation load bearing [Section 5.4 / Table 1; Section 2]
    "We used an implementation of the TD3 agent from a recent work [4] to draw a comparative study against our MAB approach. ... The results for the o-DBS setup and the TD3 Deep-RL are taken from [4]."

    The abstract's claim that T3P 'outperforms deep-RL baselines' is supported only by Table 1, whose TD3 and o-DBS numbers are taken from [4], a paper sharing authors with this work (Nicholas Carter). The paper itself notes TD3's reward was based on PSD of S_Gi in 1–20 Hz plus energy while T3P optimizes P_beta (13–35 Hz) plus energy, and that the P_beta computation differs between the works. Therefore the comparison is not a matched independent experiment; the headline result reduces to accepting the same group's prior numbers and biomarker choice.

  2. fitted input called prediction [Section 5.3, Fig. 6; Section 4.3]
    "The optimal arm chosen for regret calculation was the one that used a frequency of 155Hz and an amplitude of 1000 μA/cm2 based on empirical observations."

    This 'optimal' arm—used as the oracle for cumulative regret—is identified from the same simulation runs used to evaluate the algorithms. The T3P pruning rule is also derived from those runs ('we find that arms among these which returned below-average rewards never turned out to be the optimal arm'), so the faster-convergence/regret result is an in-sample evaluation: the oracle and the pruning heuristic are both fitted to the test distribution rather than predicted from it.

full rationale

The central contribution—T3P as a training-free, pruned bandit for joint frequency/amplitude control—is not definitionally circular: it is an algorithm evaluated against other bandits on a computational model, and the hardware energy measurements are independent. However, the strongest comparative claim in the abstract is not backed by a controlled, independent experiment. The TD3 baseline is taken from the same group's prior work [4] and was trained on a different reward and biomarker (PSD of S_Gi 1–20 Hz plus energy) than T3P (P_beta plus energy); evaluating both on P_beta therefore confounds algorithm quality with biomarker/reward choice. Additionally, the 'optimal arm' used for regret calculations is identified from the same simulation data, and the pruning heuristic is justified by the same empirical runs, making the convergence/regret advantage partly in-sample. These are real evidential weaknesses, but they are not definitional reductions: the algorithm's performance is not equivalent to its inputs by construction, and the paper contains independent content. Score 4 reflects one load-bearing self-citation plus an in-sample regret oracle, while acknowledging the method itself is a substantive, testable proposal.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the fidelity of the BGT computational model, the validity of the P_beta biomarker, and several hand-chosen parameters (epsilon, K, reward weights, decay schedule). The most fragile is the pruning assumption that below-average arms after one play can never become optimal; this is an in-sample observation, not a proven property.

free parameters (5)
  • epsilon_start = 0.2
    Exploration probability for epsilon-greedy after warm-up; chosen by grid search in Section 5.2 (Fig. 4).
  • K (top arms kept after pruning) = 25
    Number of arms retained after warm-up; chosen by grid search in Section 5.2 (Fig. 4).
  • epsilon_decay_rate = 0.025 per episode (claimed)
    Linear decay of epsilon; exact schedule inconsistent between Algorithm 1 (delta*t) and text (0.025 per episode); no value for epsilon_min given.
  • reward weights alpha, beta, gamma = -0.7, 0.1, -0.2
    Hand-chosen coefficients in reward function R_instant (Section 4.2); not justified by clinical or simulation criteria.
  • trigger threshold delta = not specified
    Threshold for P_beta deviation that re-triggers the bandit (Algorithm 1 input); value never reported, so reproducibility is limited.
assumptions (5)
  • domain assumption The basal ganglia-thalamic model [13] accurately simulates PD pathophysiology and DBS response.
    All simulation results rest on this model; the paper does not validate against clinical data.
  • domain assumption P_beta (13-35 Hz band power of GPi LFP) is a valid biomarker for PD symptom severity and correlates with the error index.
    Cited from prior literature; used as the sole feedback signal; if P_beta does not track symptoms, the algorithm optimizes the wrong objective.
  • ad hoc to paper An arm's mean reward can be estimated from a single play in the warm-up phase.
    Stated in Section 4.3 as an empirical observation from experiments; used to justify pruning after one sample per arm.
  • ad hoc to paper Below-average arms in warm-up never become optimal.
    Pruning rule; derived from limited simulation runs on the BGT model; not an established property.
  • domain assumption Biphasic stimulation with ranges [55,180] Hz and [0,5000] uA/cm^2 is safe and effective.
    Parameter ranges based on prior DBS literature; used to define the arm set.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bandit Algorithms for Deep Brain Stimulation." pith.science (2026). https://pith.science/paper/NC26SZOH

@misc{pith2026260112699,
  author       = {Pith},
  title        = {Pith review of: Bandit Algorithms for Deep Brain Stimulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NC26SZOH}},
  note         = {Machine review of arXiv:2601.12699}
}
read the original abstract

Deep Brain Stimulation (DBS) is an effective treatment for Parkinson's disease, but conventional fixed-parameter stimulation can reduce battery life and cause side effects while failing to adapt to changing neural dynamics. Recent reinforcement learning approaches improve adaptability, yet most rely on deep neural networks that require offline training and are computationally too expensive for implantable hardware. This paper presents a resource-conscious adaptive DBS framework based on a Time- and Threshold-Triggered Pruned Multi-Armed Bandit (T3P MAB) algorithm. The proposed method jointly tunes stimulation frequency and amplitude, avoids prior training, and remains transparent enough to support clinician-guided adjustment. Using a computational basal ganglia-thalamic model, we show that T3P converges faster than competing MAB methods and outperforms deep-RL baselines in suppressing pathological beta-band activity while reducing stimulation power. We implemented it on different microcontrollers and report detailed energy measurements, showing convergence in under two minutes and suitability for resource-constrained implantable systems. These results support lightweight bandit-based control as a practical path toward personalized, energy-efficient DBS.

Figures

Figures reproduced from arXiv: 2601.12699 by the authors.

Figure 1
Figure 1. Deep Brain Stimulation Types of DBS: Most modern clini￾cal approaches use continuous DBS (cDBS) that involves stimulating the brain of the patient with a fixed amplitude and frequency. Pa￾tients must visit their neurologist multiple times a year so that the frequency of stimulations can be tuned to best fit their health needs. The effectiveness of stimulation is typically evaluated by monitoring neural activity in t… view at source ↗
Figure 2
Figure 2. The Basal Ganglia-Thalamic Model 𝐼𝑇 , 𝐼𝐶𝑎, and 𝐼𝐴𝐻𝑃 denote the leak, sodium, potassium, low-threshold calcium, high-threshold calcium, and after hypo-polarization potas￾sium channel currents, respectively. Variables of the form 𝐼𝑅, where 𝑅 ∈ {𝐺𝑃𝑖,𝐺𝑃𝑒, 𝑆𝑇 𝑁 ,𝑇𝐻}, represent current flowing in from neurons in region R. 𝐼𝑎𝑝𝑝 represents the external currents influential to the BG regions. Decreasing the contribution of 𝐼… view at source ↗
Figure 3
Figure 3. RL-BGT interaction. In the context of DBS, the RL controller operates on the current neural state of the pa￾tient, represented as a vector of electrophysiological biomark￾ers. Formally, let the brain state at time step 𝑡 be denoted by 𝑠𝑡 = [𝑏1 (𝑡), 𝑏2 (𝑡), . . . , 𝑏𝑛 (𝑡)] ∈ S, where each 𝑏𝑖(𝑡) corre￾sponds to a biomarker for DBS. The RL agent observes 𝑠𝑡 and selects a stimulation action 𝑎𝑡 = [𝑓𝑡 , 𝐴𝑡 ] ∈ A, where 𝑓𝑡… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Hyper-parameter tuning of T3P MAB using cumulative rewards. 10 20 30 40 50 60 70 Rounds 0.6 0.5 0.4 0.3 0.2 0.1 Instantaneous Reward Bayes UCB CLUCB Discounted UCB Neural UCB UCB TS -Greedy -Neural TS Ours [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 7
Figure 7. Figure 7: Neuron activity in the TH (left) and GPi (right) in a [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 9
Figure 9. Figure 9: Illustration of the adaptability of the T3P algorithm when a patient [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: ESP32-S3 MCU used for our T3P MAB. We found the 𝑃𝛽 calculation to be the most resource intensive step of the al￾gorithm. This is because every time we want to find the 𝑃𝛽 after a round, we have a total of 100, 000 float values to be pro￾cessed. However, in order to pe…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 7 canonical work pages

  1. [4]

    Nicholas Carter et al . 2025. In-Vivo Training for Deep Brain Stimulation. arXiv:2510.03643 [cs.LG] https://arxiv.org/abs/2510.03643

    arXiv 2025

  2. [1]

    Elie M Adam et al. 2021. Deep brain stimulation in the subthalamic nucleus for Parkinson’s disease can restore dynamics of striatal networks.bioRxiv(2021). arXiv:https://www.biorxiv.org/content/early/2021/08/29/2021.08.29.458121.full.pdf doi:10.1101/2021.08.29.458121

    work page doi:10.1101/2021.08.29.458121 2021

  3. [2]

    Al-Thaddeus Avestruz et al. 2009. A 5 W/Channel Spectral Analysis IC for Chronic Bidirectional Brain–Machine Interfaces.Solid-State Circuits, IEEE Journal of43 (01 2009), 3006 – 3024. doi:10.1109/JSSC.2008.2006460

    arXiv 2009

  4. [3]

    Arezu Bagheri et al. 2013. Massively-Parallel Neuromonitoring and Neurostimula- tion Rodent Headset With Nanotextured Flexible Microelectrodes.IEEE Transactions on Biomedical Circuits and Systems7, 5 (Oct 2013), 601–609. doi:10.1109/TBCAS. 2013.2281772

    arXiv 2013

  5. [5]

    Chia-Hung Cho et al. 2024. Closed-Loop Deep Brain Stimulation With Reinforce- ment Learning and Neural Simulation.IEEE Tran. Neural Systems & Rehabilitation Engg.32 (2024), 3615–3624. doi:10.1109/TNSRE.2024.3465243

    arXiv 2024

  6. [6]

    Fleming, Eleanor Dunn, and Madeleine M

    John E. Fleming, Eleanor Dunn, and Madeleine M. Lowery. 2020. Simulation of Closed-Loop Deep Brain Stimulation Control Schemes for Suppression of Patho- logical Beta Oscillations in Parkinson’s Disease.Frontiers in NeuroscienceVolume 14 - 2020 (2020). doi:10.3389/fnins.2020.00166

    arXiv 2020

  7. [7]

    Qitong Gao et al. 2020. Model-Based Design of Closed Loop Deep Brain Stimulation Controller using Reinforcement Learning. In11th International Conference on Cyber- Physical Systems (ICCPS). doi:10.1109/ICCPS48487.2020.00018

    arXiv 2020

  8. [8]

    Qitong Gao et al. 2022. Offline Policy Evaluation for Learning-based Deep Brain Stimulation Controllers. In2022 ACM/IEEE 13th International Conference on Cyber- Physical Systems (ICCPS). 80–91. doi:10.1109/ICCPS54341.2022.00014

    arXiv 2022

Show all 34 references
  1. [10]

    Qitong Gao et al. 2023. Offline Learning of Closed-Loop Deep Brain Stimulation Controllers for Parkinson Disease Treatment. InProceedings of the ACM/IEEE 14th International Conference on Cyber-Physical Systems (with CPS-IoT Week 2023)(San Antonio, TX, USA)(ICCPS ’23). Associat...

    2023

  2. [11]

    Aurélien Garivier and Eric Moulines. 2011. On Upper-Confidence Bound Policies for Switching Bandit Problems. InAlgorithmic Learning Theory, Jyrki Kivinen, Csaba Szepesvári, Esko Ukkonen, and Thomas Zeugmann (Eds.). Springer Berlin Heidelberg, Berlin, Heidelberg, 174–188

    2011

  3. [12]

    2024.𝜖-Neural Thompson Sampling of Deep Brain Stimulation for Parkinson Disease Treatment

    Hao-Lun Hsu, Qitong Gao, and Miroslav Pajic. 2024.𝜖-Neural Thompson Sampling of Deep Brain Stimulation for Parkinson Disease Treatment. In2024 ACM/IEEE 15th International Conference on Cyber-Physical Systems (ICCPS). 224–234. doi:10. 1109/ICCPS61052.2024.00027

    2024

  4. [13]

    Ilija Jovanov et al. 2018. Platform for Model-Based Design and Testing for Deep Brain Stimulation. In9th International Conference on Cyber-Physical Systems. 263–

    2018

  5. [14]

    Emilie Kaufmann, Olivier Cappe, and Aurelien Garivier. 2012. On Bayesian Upper Confidence Bounds for Bandit Problems. InProceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics (Proceedings of Machine Learning Research, Vol. 22), Neil D. ...

    2012

  6. [15]

    Abbas Kazerouni et al. 2017. Conservative contextual linear bandits. InProceedings of the 31st International Conference on Neural Information Processing Systems(Long Beach, California, USA)(NIPS’17). Curran Associates Inc., Red Hook, NY, USA, 3913–3922

    2017

  7. [16]

    2020.Bandit Algorithms

    Tor Lattimore and Csaba Szepesvári. 2020.Bandit Algorithms. Cambridge University Press

    2020

  8. [17]

    Lempka et al

    Scott F. Lempka et al. 2018. Characterization of the stimulus waveforms generated by implantable pulse generators for deep brain stimulation.Clinical Neurophysiol- ogy129, 4 (2018), 731–742. doi:10.1016/j.clinph.2018.01.015

    2018 doi

  9. [18]

    Xilin Liu et al. 2015. The PennBMBI: Design of a General Purpose Wireless Brain- Machine-Brain Interface System.IEEE transactions on biomedical circuits and systems9 (03 2015). doi:10.1109/TBCAS.2015.2392555

    2015

  10. [19]

    Meili Lu et al . 2020. Application of Reinforcement Learning to Deep Brain Stimulation in a Computational Model of Parkinson’s Disease.IEEE Transac- tions on Neural Systems and Rehabilitation Engineering28, 1 (2020), 339–349. doi:10.1109/TNSRE.2019.2952637

    2020

  11. [20]

    Marras et al

    C. Marras et al. 2018. Prevalence of Parkinson’s disease across North America.npj Parkinson’s Disease4, 1 (10 Jul 2018), 21. doi:10.1038/s41531-018-0058-0

    2018 doi

  12. [21]

    Brigitte Piallat et al. 2009. Monophasic but not biphasic pulses induce brain tissue damage during monopolar high-frequency deep brain stimulation.Neurosurgery 64 (2009), 156–162. doi:10.1227/01.NEU.0000336331.88559.CF

    2009

  13. [22]

    Marco Pirini et al . 2009. A computational modelling approach to investigate different targets in deep brain stimulation for Parkinson’s disease.Journal of Computational Neuroscience26, 1 (01 Feb 2009), 91–107. doi:10.1007/s10827-008- 0100-z

    2009 doi

  14. [23]

    Jonathan Rubin and David Terman. 2004. High Frequency Stimulation of the Subthalamic Nucleus Eliminates Pathological Thalamic Rhythmicity in a Compu- tational Model.J. Computational Neuroscience16 (2004), 211–35. doi:10.1023/B: JCNS.0000025686.47117.67

    2004

  15. [24]

    Muhammad Salam, Jose Perez-Velazquez, and Roman Genov. 2015. Seizure Sup- pression Efficacy of Closed-Loop Versus Open-Loop Deep Brain Stimulation in a Rodent Model of Epilepsy.IEEE Transactions on Neural Systems and Rehabilitation Engineering24 (01 2015), 1–1. doi:10.1109/TNS...

    2015

  16. [25]

    Ruslana Shulyzki et al. 2015. 320-Channel Active Probe for High-Resolution Neu- romonitoring and Responsive Neurostimulation.IEEE Transactions on Biomedical Circuits and Systems9, 1 (2015), 34–49. doi:10.1109/TBCAS.2014.2312552

    2015

  17. [26]

    So et al

    R.Q. So et al. 2012. Relative contributions of local cell and passing fiber activation and silencing to changes in thalamic fidelity during deep brain stimulation and lesioning: a computational modeling study.J. Computational Neuroscience32, 3 (2012), 499–519. doi:10.1007/s108...

    2012 doi

  18. [27]

    Stanslaski et al

    S. Stanslaski et al. 2012. Design and Validation of a Fully Implantable, Chronic, Closed-Loop Neuromodulation Device With Concurrent Sensing and Stimulation. IEEE transactions on neural systems and rehabilitation engineering : a publication of the IEEE Engineering in Medicine ...

    2012

  19. [28]

    Scott Stanslaski et al. 2018. A Chronically Implantable Neural Coprocessor for Inves- tigating the Treatment of Neurological Disorders.IEEE Transactions on Biomedical Circuits and Systems12, 6 (2018), 1230–1245. doi:10.1109/TBCAS.2018.2880148

    2018

  20. [29]

    Fei Su et al. 2019. Model-Based Evaluation of Closed-Loop Deep Brain Stimula- tion Controller to Adapt to Dynamic Changes in Reference Signal.Frontiers in NeuroscienceVolume 13 - 2019 (2019). doi:10.3389/fnins.2019.00956

    2019

  21. [30]

    George Thomas and Barbara Jobst. 2015. Critical review of the responsive neu- rostimulator system for epilepsy.Medical Devices: Evidence and Research8 (10 2015),

    2015

  22. [31]

    Stavros Zanos et al. 2011. The Neurochip-2: An Autonomous Head-Fixed Computer for Recording and Stimulating in Freely Behaving Monkeys.IEEE Transactions Arkaprava Gupta, Nicholas Carter, William Zellers, Prateek Ganguli, Benedikt Dietrich, Vibhor Krishna, Parasara Sridhar Dugg...

    2011

  23. [32]

    Andy Zhou et al. 2017. WAND: A 128-channel, closed-loop, wireless artifact-free neuromodulation device. (08 2017). doi:10.48550/arXiv.1708.00556

    2017 doi arXiv

  24. [33]

    Dongruo Zhou, Lihong Li, and Quanquan Gu. 2020. Neural Contextual Bandits with UCB-based Exploration. InProceedings of the 37th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 119), Hal Daumé III and Aarti Singh (Eds.). PMLR, 11492–...

    2020

  25. [274]

    doi:10.1109/ICCPS.2018.00033

    2018

  26. [405]

    doi:10.2147/MDER.S62853

    doi

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.