SiLIF-DBS: Neuromorphic Adaptive DBS Controller

• The paper demonstrates that the SiLIF-DBS controller, implemented in 130nm CMOS, uses a refractory-enabled neuromorphic LIF neuron to adaptively modulate stimulation based on beta-band activity.
• Its design integrates membrane, leak, and refractory dynamics into transistor physics, achieving 25% power consumption relative to open-loop DBS while suppressing pathological beta oscillations.
• The controller is validated in a cortico–basal ganglia computational model, showing improved suppression efficiency and outlining future directions for fully implantable neuromorphic DBS systems.

Searching arXiv for the cited SiLIF-DBS and related DBS controller papers. arxiv_search(query="(Siddique et al., 5 Jul 2026)", max_results=5, sort_by="relevance") Searching for the primary SiLIF-DBS paper and closely related neuromorphic DBS work. {"query":"(Siddique et al., 5 Jul 2026)","max_results":5,"sort_by":"relevance"} Attempting arXiv lookup for exact identifiers and related terms. The SiLIF-DBS controller is a neuromorphic, leaky integrate-and-fire silicon neuron used directly as the control element in an adaptive deep brain stimulation system for Parkinson’s disease. In its current formulation, it is implemented in 130 nm CMOS, driven by beta-band subthalamic nucleus local field potentials, and uses the Beta_ARV biomarker to generate stimulation only when needed rather than delivering continuous fixed-amplitude stimulation. The defining claim of the architecture is that the controller itself is realized as a refractory-enabled silicon LIF neuron whose membrane and refractory dynamics determine stimulation behavior, yielding a system that suppresses pathological beta activity while using only 25% of the power required by open-loop DBS and achieving a suppression efficiency of $5.85\%/\mu\text{W}$ (Siddique et al., 5 Jul 2026).

1. Concept and clinical role

The SiLIF-DBS controller was proposed as a response to a specific limitation of conventional open-loop DBS in Parkinson’s disease: continuous high-frequency stimulation, typically around 130 Hz with fixed amplitude, is delivered regardless of whether beta activity is high or low. In the formulation under discussion, this causes excessive battery drain, high pulse-generator duty cycle, increased power consumption, and stimulation-related side effects because the brain is stimulated even when symptoms are mild. The controller therefore belongs to the broader class of adaptive DBS systems that monitor a neural biomarker and adapt stimulation in real time, but it departs from software-centric control by implementing the controller as an analog neuron whose integration, leak, threshold, reset, and refractory behavior appear directly in transistor physics (Siddique et al., 5 Jul 2026).

Within this framework, pathological beta-band oscillations in the subthalamic nucleus, specifically $13$–$30$ Hz activity extracted from STN-LFPs, serve as the control input. The resulting controller is not merely inspired by neural dynamics; it uses a silicon LIF neuron as the actuator-facing decision element. This design places the energy–suppression tradeoff inside the neuron circuit itself. A plausible implication is that controller tuning can be performed in directly physical terms—membrane, leak, reset, and refractory parameters—rather than only through software hyperparameters.

2. Silicon neuron model and CMOS realization

The starting point is the current-based LIF neuron, extended with refractory conductance dynamics. In canonical form, the membrane dynamics are written as

$$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$

with refractory decay

$$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$

Spike generation follows standard LIF logic: when $V_{\text{MEM}}$ crosses threshold $V_{\text{th}}$, a spike is emitted, the membrane is reset, and refractory conductance is increased (Siddique et al., 5 Jul 2026).

In the SiLIF-DBS circuit, the membrane node is $V_{\text{MEM}}$ on capacitor $C_{\text{mem}}$, driven by synaptic current $I_{\text{syn}}(t)$, and the effective circuit dynamics become

$13$0

Here, $13$1 is implemented by transistor M1 and set by bias $13$2, while the reset/refractory shunt is represented by $13$3. Threshold detection is performed by the M3–M4 comparator, which produces node $13$4; a spike is generated when the first inverter INV1 switches. The refractory state is represented by $13$5 on capacitor $13$6, with approximate refractory time constant

$13$7

The physical realization is a single silicon neuron with explicit subcircuits for synaptic-current integration, leak control, threshold detection, spike generation, and reset/refractory dynamics. It is implemented in SkyWater 130 nm CMOS with supply $13$8. In the reported computational model, $13$9, $30$0, $30$1, $30$2, and pulse width $30$3. The key tuning knob is $30$4; by adjusting it, the effective refractory time constant changes, setting the steady-state firing rate. The paper reports firing frequency increasing monotonically with $30$5 over $30$6–$30$7 Hz, which is central to how the controller encodes stimulation strength (Siddique et al., 5 Jul 2026).

3. Biomarker pathway and control law

The SiLIF-DBS control loop is driven by beta-band STN-LFPs. The processing chain is band-pass filtering with a Chebyshev filter over $30$8–$30$9 Hz, rectification, and short-window averaging to compute Beta_ARV, with updates every controller period $$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$0:

$$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$1

Although the implementation is discrete, the paper states the Beta_ARV conceptually as an average rectified value over a short window on the order of tens of milliseconds (Siddique et al., 5 Jul 2026).

Within each interval $$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$2, Beta_ARV sets a constant synaptic current $$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$3 proportional to the biomarker. Stronger beta therefore drives larger $$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$4, causing faster membrane buildup and more spikes. The controller then computes a smoothed spike-rate estimate

$$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$5

with $$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$6 and $$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$7, and maps that rate to a stimulation amplitude command

$$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$8

using $$\frac{dV_{\text{MEM}}}{dt} = \frac{1}{C_{\text{mem}}}\big[I_{\text{syn}}(t) - (g_L + g_{\text{ref}}(t))(V_{\text{MEM}} - E_L)\big],$$9, $$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$0, and $$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$1 (Siddique et al., 5 Jul 2026).

This control law has a specific interpretation. When Beta_ARV exceeds a threshold, $$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$2 increases, the neuron fires more rapidly, $$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$3 grows, and $$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$4 rises toward $$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$5. When beta activity decreases, spike rate and $$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$6 drop, producing weaker or no stimulation. The reported result is temporally sparse, amplitude-modulated DBS bursts rather than continuous fixed-amplitude stimulation. The downstream DBS waveform is represented as

$$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$7

with pulses occurring at a fixed carrier frequency. A common misconception is that the SiLIF-DBS paper already presents a fully laid out implantable stimulator chain; in fact, the pulse generator is described as a software/hardware pulse generator in the computational framework, with hardware compatibility to existing DBS stimulators assumed rather than fully realized on the same die (Siddique et al., 5 Jul 2026).

4. System-level validation and quantitative results

For system-level evaluation, the controller is embedded in a Parkinsonian cortico–basal ganglia framework following Fleming et al. and others. The loop is closed by generating raw STN-LFP from the network, extracting beta-band activity, computing Beta_ARV, feeding that quantity to the SiLIF-DBS computational neuron, and injecting the resulting amplitude-modulated DBS waveform back into the model. In this setting, the controller is explicitly benchmarked against open-loop DBS, an on-off controller, and a dual-threshold controller (Siddique et al., 5 Jul 2026).

Controller	Power (% of Open-loop)	Suppression Efficiency (%/$$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$8W)
Open-loop DBS	100.0	1.80
On-off	41.0	8.10
Dual-threshold	52.0	5.70
SiLIF-DBS (Ours)	25.0	5.85

The quantitative interpretation of this table is precise. SiLIF-DBS uses only 25% of the power required by open-loop DBS. Its suppression efficiency of $$\frac{dg_{\text{ref}}}{dt} = -\frac{g_{\text{ref}}}{T_{\text{ref}}}.$$9 is more than three times better than open-loop and slightly higher than dual-threshold control, while the paper characterizes its stimulation as smoother than pure on–off control. Figures described in the source material show strong beta oscillations under DBS-off and substantial attenuation under SiLIF-DBS, with Beta_ARV frequently above threshold in the DBS-off condition and near or below threshold under SiLIF-DBS. The low-power profile is attributed to analog neuromorphic implementation, subthreshold or low-voltage operation, event-driven behavior, compact architecture, and the single tuning parameter $V_{\text{MEM}}$0 (Siddique et al., 5 Jul 2026).

These results should nevertheless be read in light of the evaluation setting. The validation is model-based rather than in animal or human experiments. The reported suppression and power figures are therefore properties of the controller embedded in the computational cortico–basal ganglia model, not yet of an in-vivo implanted system. This distinction is central to an accurate understanding of the current state of the architecture.

5. Relation to adjacent closed-loop DBS architectures

The SiLIF-DBS controller sits at the intersection of at least four nearby lines of work. First, the 2024 neuromorphic Parkinson’s disease study introduced on-off LIF and dual LIF controllers that used STN beta-band average rectified value to adjust DBS current in simulation. That work reported that the on-off LIF controller and dual LIF controller reduced the power consumption of closed-loop DBS systems by 19% and 56%, respectively, while suppression efficiency in Figure 1 was $V_{\text{MEM}}$1 for on-off LIF and $V_{\text{MEM}}$2 for dual LIF. In this sense, SiLIF-DBS can be read as a transistor-level continuation of the earlier proposal that a silicon implementation of LIF-based neuromorphic control could be embedded inside the IPG (Biswas et al., 2024).

Second, the 2022 phase-locked DBS SoC provides a complementary architecture rather than a competing controller. That chip integrates a 16-channel low-noise chopper-stabilized analog front-end, programmable digital FIR filtering and analytic signal generation, a lightweight phase estimator, a neural connectivity processor, a 4-channel charge-balanced neurostimulator, and control logic for multi-mode phase-locked stimulation. It realizes recorded LFPs $V_{\text{MEM}}$3 digital biomarker extraction $V_{\text{MEM}}$4 decision logic $V_{\text{MEM}}$5 phase-locked stimulation, consumes $V_{\text{MEM}}$6 total, and can be viewed as the “signal processing and timing front-end” for a future SiLIF-DBS controller in which neuromorphic LIF circuits replace threshold-and-rule decision logic (Shin et al., 2022).

Third, the 2026 energy-aware learning study generalizes the neuromorphic DBS concept from a single analog silicon neuron to a deep spiking Q-network. That controller used a biophysical cortico-basal ganglia-thalamic model, learned discrete adjustments of frequency, pulse width, and amplitude, reduced pathological alpha-beta oscillations by 45.2%, reduced stimulation charge by 80.0% relative to continuous DBS, and was distilled onto the SynSense XyloAudio 3 neuromorphic processor at 0.52 mW inference power. The paper explicitly frames this as co-optimization of stimulation energy and inference efficiency. This suggests a broader taxonomy in which the SiLIF-DBS controller is the single-neuron, analog-CMOS end of a wider neuromorphic DBS design space that also includes multi-layer spiking RL controllers (Nguyen et al., 26 Jun 2026).

Fourth, the clinically tested offline RL framework for Parkinson disease treatment provides an algorithmic counterpoint. That study adapted stimulation amplitude every $V_{\text{MEM}}$7 seconds from windowed beta-amplitude history, was evaluated in four PD patients using the RC+S DBS system, and reported 20–55% reduction in stimulation energy with non-inferior symptom control relative to continuous DBS. Relative to this line of work, SiLIF-DBS shifts complexity away from a tablet-hosted policy and into transistor-level membrane and refractory dynamics, thereby targeting implantable control rather than external computation (Gao et al., 2023).

6. Limitations, interpretation, and future directions

Several limitations are explicit. The SiLIF-DBS controller is validated in a computational cortico–basal ganglia model rather than in animal or human experiments. The biomarker pathway is intentionally simple: beta-band filtering, rectification, and Beta_ARV. Control is implemented by one neuromorphic neuron, even though real brain–device interactions are multi-channel and multi-target. DBS pulses are delivered at a fixed carrier frequency and only amplitude is modulated. Detailed discussion of noise, mismatch, temperature drift, and long-term reliability of the silicon neuron in vivo is not yet provided (Siddique et al., 5 Jul 2026).

These constraints delimit what the term “SiLIF-DBS controller” currently denotes. It does not yet denote a full closed-loop implant with integrated multi-channel sensing, artifact management, and physical in-vivo validation. In that broader sense, the 2022 phase-locked DBS SoC shows what a fully integrated sensing and stimulation substrate can look like: multi-region LFP acquisition, instantaneous phase extraction, PLV/PAC/SE computation, control logic, and a charge-balanced high-voltage stimulator on chip (Shin et al., 2022). A plausible implication is that a future fully integrated SiLIF-DBS system would combine that class of sensing-and-stimulation SoC with the silicon LIF decision element developed for Parkinsonian adaptive control.

The future directions stated for SiLIF-DBS are correspondingly concrete: multi-neuron arrays, more complex biomarker-driven control, hardware-in-the-loop and in-vivo validation, mixed-signal integration with implantable sensing front-ends and pulse generators, and exploration of alternative neuromorphic primitives beyond LIF. Taken together, these directions define the present status of the SiLIF-DBS controller: a transistor-level neuromorphic controller for adaptive DBS whose main contribution is to realize the controller itself as a silicon LIF neuron, while leaving the full closed-loop implant as the next integration target (Siddique et al., 5 Jul 2026).