🧬 biology

Neuromorphic Energy-Aware Learning for Adaptive Deep Brain Stimulation

This paper introduces an energy-aware learning framework that co-optimizes actuator stimulation and inference efficiency by incorporating actuator energy into reinforcement learning rewards and deploying a compressed spiking policy on neuromorphic hardware, achieving significant reductions in both pathological oscillations and total power consumption for adaptive deep brain stimulation.

Original authors: Jason Eshraghian, Binh Nguyen, Colleen Josephson, Mircea Teodorescu, Gert Cauwenberghs

Published 2026-06-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jason Eshraghian, Binh Nguyen, Colleen Josephson, Mircea Teodorescu, Gert Cauwenberghs

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Problem: The "Heavy Lifter" vs. The "Smart Brain"

Imagine you have a robot designed to help a person with Parkinson's disease. This robot has two main jobs:

Think: It needs to constantly monitor the patient's brain signals to know when they are having a tremor.
Act: It needs to send tiny electrical pulses (stimulation) to the brain to stop the tremor.

For years, researchers have been trying to make the "Think" part (the computer brain) super efficient and low-power. They've built tiny, energy-saving chips that can do this thinking for pennies a day.

But here is the catch: The paper argues that making the "Think" part efficient isn't enough. The "Act" part (sending the electrical pulses) is actually the heavy lifter. It uses way more energy than the computer brain does. If you have a super-efficient brain but a wasteful arm, you still run out of battery.

The authors realized that to save battery life, the computer brain needs to learn how to be lazy with the arm. It shouldn't just think efficiently; it needs to learn to send fewer and smaller electrical pulses while still doing its job.

The Solution: A "Smart" Brain That Learns to Save Energy

The team created a new kind of computer brain called a Spiking Neural Network (SNN). Think of this not as a standard calculator that runs constantly, but like a biological brain. It only "fires" (uses energy) when it actually receives a signal, just like a neuron in your body.

They trained this brain using a special game called Reinforcement Learning.

The Goal: Stop the brain tremors (pathological oscillations).
The Twist: In the game, the brain gets "punished" not just for failing to stop the tremor, but also for using too much electricity to do it.

It's like training a dog. Usually, you only reward the dog for catching a frisbee. In this experiment, the trainer also says, "If you catch the frisbee but run around in circles wasting energy first, you get no treat." The brain learned that the best way to win is to be precise and use the minimum amount of energy necessary.

The Results: Smarter, Faster, and Longer-Lasting

The team tested this new brain in a computer simulation of a rat's brain (which mimics the human brain's Parkinson's symptoms). Here is what happened:

It worked: The new brain stopped the bad brain waves (tremors) by 45%.
It saved massive energy: Because it learned to be efficient, it reduced the electrical charge sent to the brain by 80% compared to the old "always-on" method.
It fits on a tiny chip: They shrunk this brain down to fit on a specialized, ultra-low-power chip (the SynSense XyloAudio 3).
- The Comparison: If you tried to run this same brain on a standard powerful computer chip (like the one in a high-end smartphone or laptop), it would drain a battery in less than an hour.
- The Reality: On their tiny, specialized chip, the brain uses so little power that a standard implant battery could last over a year without needing a surgery to replace it.

The "Magic Trick": Distilling the Brain

The brain they trained was initially too big and complex to fit on the tiny implant chip. So, they used a technique called Knowledge Distillation.

Imagine a master chef (the big, complex brain) teaching an apprentice (the small chip brain). The master chef knows everything but is slow and uses too many ingredients. The apprentice learns to cook the exact same delicious meal but uses fewer ingredients and works faster. The paper shows that the apprentice could cook the meal just as well, but with 3.6 times fewer steps and using a fraction of the energy.

Why This Matters

The paper concludes that for medical implants to work for years without battery changes, you can't just make the computer brain efficient. You have to teach the brain to be efficient about the actuator (the part that delivers the electricity) too.

By combining a brain that thinks like a biological neuron (spiking) with a reward system that punishes energy waste, they created a system that is:

Therapeutically effective: It stops the symptoms.
Energy efficient: It saves 80% of the stimulation energy.
Hardware ready: It runs on a chip small and cool enough to be implanted in a human for years.

In short, they built a "smart thermostat" for the brain that doesn't just turn the heat on and off, but learns exactly how much heat is needed to keep the house comfortable without burning the house down.

Technical Summary: Neuromorphic Energy-Aware Learning for Adaptive Deep Brain Stimulation

Problem Statement

The deployment of neural network controllers in physical closed-loop systems, such as Implantable Pulse Generators (IPGs) for Deep Brain Stimulation (DBS), faces a critical energy bottleneck. While neuromorphic and edge computing have successfully reduced the inference cost of neural controllers, the energy consumed by the actuator (the neurostimulator delivering electrical charge) often rivals or exceeds the controller's inference cost. In DBS for Parkinson's Disease (PD), continuous stimulation (cDBS) consumes power on the order of hundreds of microwatts to a milliwatt, a regime comparable to efficient neuromorphic controllers. Consequently, optimizing inference efficiency alone is insufficient; the control policy must also minimize the physical charge delivered to the tissue to extend battery life and prevent thermal damage. Furthermore, conventional adaptive DBS (aDBS) approaches rely on window-aggregated features (e.g., mean spectral power) that discard fine-grained temporal spike structure, and they often utilize hardware (e.g., GPUs) that exceeds the sub-milliwatt power envelope required for chronic implants.

Methodology

The authors propose a framework for Energy-Aware Learning that co-optimizes therapeutic efficacy and stimulation energy within a closed-loop system.

Biophysical Simulation Environment:
- The system is trained in a dopamine-depleted biophysical model of the Cortico-Basal Ganglia-Thalamocortical (CBGT) circuit, based on the 6-OHDA lesioned rat model.
- The model comprises 80 neurons across eight populations (including STN, GPe, GPi, Striatum, Thalamus) using Hodgkin-Huxley and Izhikevich dynamics.
- The state observation is the raw 80-channel spike raster, which is spatially downsampled to 16 channels via average pooling to match hardware constraints.
Reinforcement Learning Agent (Teacher):
- A Deep Spiking Q-Network (DSQN) is trained using off-policy Q-learning.
- Reward Function: The reward is a discontinuous function that prioritizes suppressing pathological $\alpha$ - $\beta$ oscillations (7–35 Hz) in the GPi. If the $\beta$ -power exceeds a threshold ( $\tau_\beta$ ), the agent is penalized. If below the threshold, the reward is a weighted sum of therapeutic maintenance and an energy-aware penalty based on the Root Mean Square (RMS) of the stimulation current. This forces the agent to learn charge-efficient strategies.
- Action Space: The agent modulates three discrete DBS parameters: Frequency ( $f$ ), Amplitude ( $A$ ), and Pulse Width ( $\delta$ ).
Sparsity-Constrained Knowledge Distillation:
- To deploy the policy on ultra-low-power hardware, a "teacher" SNN is distilled into "student" SNNs.
- The distillation loss combines Kullback-Leibler (KL) divergence between teacher and student Q-values with an explicit sparsity penalty. This penalty regularizes the average firing rate of hidden layers toward a target sparsity ( $\rho$ ), directly minimizing Synaptic Operations (SynOps) to reduce dynamic power.
Hardware Deployment:
- The distilled student policy is deployed on the SynSense XyloAudio 3 neuromorphic processor.
- Performance is benchmarked against an Artificial Neural Network (ANN) and a Gated Recurrent Unit (RNN) baseline running on an NVIDIA Jetson Orin Nano.

Key Results

Therapeutic Efficacy

The SNN controller reduced pathological $\alpha$ - $\beta$ band power by 45.2% (95% CI [42.6, 47.7]) relative to the unstimulated PD baseline.
The controller successfully detected the pathological state and modulated stimulation parameters (ramping frequency to 180 Hz while decreasing amplitude) to break pathological synchrony between the STN and GPe.

Stimulation Energy Efficiency

Compared to continuous DBS (cDBS), the energy-aware SNN policy achieved an 80.0% reduction in cumulative stimulation charge.
It outperformed a clinical dual-threshold heuristic (current aDBS standard), which achieved only a 57.0% reduction. The SNN delivered 53.5% less charge than the heuristic while maintaining deeper suppression of $\beta$ -power.
Baseline Failures:
- The ANN (rate-coded) converged to a degenerate policy of near-zero stimulation (94.8% charge reduction) but failed to suppress $\beta$ -oscillations.
- The RNN failed to learn adaptive down-regulation, locking parameters at maximum values and consuming 50.6% more energy than cDBS.

Hardware Efficiency and Power Budget

Inference Power: The Xylo SNN consumed 0.52 mW (including I/O and analog front-end), whereas the Jetson ANN and RNN consumed 5.7 W and 5.6 W, respectively.
Energy per Inference: The Xylo SNN achieved a 28.1× reduction in energy per inference compared to the ANN and 72.5× compared to the RNN.
Latency Trade-off: The SNN inference latency was 315.10 ms (vs. 0.81 ms for ANN), but this remains within the physiological timescale of $\beta$ -band envelope variations (~1 s).
System-Level Impact: When combining inference and stimulation power, the Xylo-based system draws a total of ~0.57 mW. This projects to >1 year of continuous operation on a standard 5 Wh non-rechargeable IPG battery. In contrast, a Jetson-based system would deplete the battery in <1 hour.

Significance and Claims

The paper argues that hardware efficiency is a learnable property of the neural controller. By incorporating actuator energy directly into the reinforcement learning reward, the policy discovers charge-efficient strategies that standard objectives miss.

The authors claim that this framework addresses the "first-order" power demands of implantable neuromodulation by co-optimizing the controller and the actuator. They demonstrate that:

Sub-milliwatt inference is a prerequisite: Without it, the controller's power draw renders any stimulation savings irrelevant.
Low-power inference alone is insufficient: The stimulator remains a significant power cost; thus, the control policy must actively minimize charge delivery.
Spiking dynamics act as an inductive bias: The SNN's leaky integrate-and-fire dynamics and sparse, binary activations prevent convergence to the degenerate policies (either no stimulation or max stimulation) found by the ANN and RNN, respectively.

The work validates that neuromorphic computing, when combined with energy-aware learning, can resolve the power constraints that have previously precluded the use of complex neural network controllers in chronic, battery-powered implants. The results are grounded in a rodent biophysical model, with the authors noting that while absolute values are approximations, the fractional savings (e.g., 80% charge reduction) are expected to translate to clinical settings.