An end-to-end CMOS–memristor integrated platform for neural signal processing

This paper presents the design and hardware validation of a CMOS–memristor hybrid neural signal processing platform. The system combines a CMOS neural amplifier, a level-crossing spike encoder, FPGA-based pulse conditioning, and a volatile TiOx memristive integrating sensor (MIS) to convert low-frequency neural activity into spike-based events and then into a compact resistance trajectory. In this study, previously recorded LFP data were preprocessed and replayed through the hardware signal chain, while saline-bath electrode coupling was used as a controlled intermediate validation step between direct electrical injection and future biological acquisition. The results show that the front-end can preserve event timing under electrode-mediated coupling and that temporally clustered spike activity can induce measurable MIS resistance modulation. Overall, the platform demonstrates a feasibility-oriented hardware pathway for sensor-proximal neural activity summarisation, while quantitative event-detection benchmarking, multi-device statistics, and multi-channel MIS validation remain necessary for future deployment-oriented studies.

Brain–computer interfaces (BCIs) translate neural activity into actionable outputs, enabling applications ranging from assistive communication to neuro-rehabilitation. Clinical intracortical BCIs have demonstrated high-performance communication in people with paralysis, motivating continued development of implantable neural interfaces that are both accurate and practical for long-term use (Pandarinath et al., 2017). Recent intracortical brain-to-text demonstrations show that motor cortical activity can be decoded into real-time text, highlighting the clinical value of reliable neural acquisition and efficient signal handling (Willett et al., 2021). A core bottleneck is not only decoding accuracy but also how neural signals are acquired, represented, and moved through the system under tight power and size constraints, especially in chronic or fully implantable settings (Sun et al., 2024; Wu et al., 2023). Recent memristor-chip BCI systems further show that memristive hardware can contribute beyond passive storage, enabling adaptive neural decoding with software-equivalent performance and substantially improved computational efficiency (Liu et al., 2025). However, such decoder-level demonstrations address the downstream interpretation stage, whereas practical implantable systems also require compact front-end pathways that condition low-amplitude neural signals and convert them into device-compatible representations before decoding. Scalable BMI platforms with thousands of recording channels illustrate the broader trend toward higher channel counts, which increases the importance of local compression and task-relevant feature formation before full-rate data transmission (Musk and Neuralink, 2019).

Extracellular electrophysiology spans multiple signal modalities—action potentials (APs), local field potentials (LFPs), electrocorticography (ECoG), and electroencephalography (EEG)—each reflecting different spatial scales and biophysical origins (Buzsáki et al., 2012; Moran, 2010; Miller et al., 2020). These signals are typically low-amplitude and susceptible to interference from the electrode–tissue interface and electronic noise, motivating signal conditioning and feature extraction prior to classification (Wu et al., 2023).

In conventional pipelines, neural waveforms are continuously sampled, digitized, and streamed to downstream digital processing. The need for front-end data reduction is reinforced by high-density silicon probes, which can record from large neural populations across multiple brain regions and therefore increase the raw data volume that must be handled by the acquisition chain (Jun et al., 2017). However, as channel counts scale, the energy and bandwidth cost of “sample-everything” architectures can rival (or exceed) the cost of neural signal acquisition itself, making data movement a first-order design constraint (Marblestone et al., 2013). In implantable devices, this pressure is amplified by limited battery capacity, strict thermal limits, and long-term reliability requirements (Even-Chen et al., 2020; Serrano-Amenos et al., 2023). These system-level constraints motivate near-sensor intelligence—moving compression and decision-relevant feature formation closer to the recording front-end, so that only essential information summaries are retained or transmitted.

Event-driven representations are a natural fit for this goal. Many neural recordings exhibit a temporal structure where salient activity is concentrated in intermittent epochs. Moreover, encoding schemes that emphasize events (and their density) can reduce redundant samples and lower effective data rate (Wu et al., 2023). In practice, the challenge is to realize event detection and temporal integration without reintroducing substantial overhead via extra analogue circuitry or complex digital timing logic. This concern is also reflected in recent work on hardware-efficient neural event detection, where reducing calibration and circuit overhead remains central to making spike-detection pipelines practical for implantable brain-machine interfaces (Zhang et al., 2023).

Memristive devices (often implemented as resistive random-access memory, RRAM) offer a route to in-sensor computing because their resistance or conductance evolves nonlinearly with stimulus history. This property effectively embeds memory and computation in device physics (Tzouvadaki et al., 2023). In addition to spike-oriented memristive sensing, array-level demonstrations have shown that multichannel neural signals can be processed in parallel by encoding waveform energy and variation into memristor conductance changes, thereby forming compact device-state features for downstream neural-state classification (Liu et al., 2020). In the memristive integrating sensor (MIS) paradigm (Gupta et al., 2016; Sbandati et al., 2024), neural activity is mapped onto a slowly evolving resistive state: subthreshold fluctuations produce little change, whereas sustained or high-density events can trigger pronounced resistive transitions. This mechanism provides implicit temporal integration and thresholding at the sensing layer, thereby yielding a compact activity proxy while suppressing sparse, noise-like events (Gupta et al., 2016). More recently, MIS-style processing has been adapted beyond spike-centric tasks to LFP-based event detection, demonstrating single-trial identification of behaviorally relevant auditory cues from rat LFPs via memristor state changes and lightweight thresholding (Sbandati et al., 2024).

Despite this progress, a practical gap remains between concept demonstrations and deployable front-ends: device–circuit-system co-design is required to ensure that (i) neural signals are conditioned into a memristor-compatible stimulus regime, (ii) event timing is preserved where needed, and (iii) readout strategies match device volatility/retention so that informative transitions are not missed. Addressing these issues is essential for developing reliable MIS-based systems that can be used in the target applications such as low-power neural implants.

In this work, we implement and experimentally evaluate an end-to-end, spike-based memristive front-end for local neural activity encoding. Low-frequency neural signals are first conditioned and converted into event trains through a Delta modulator. Importantly, the present study evaluates a replay-based hardware pathway rather than live in vivo acquisition: prerecorded LFP data are preprocessed offline, replayed through the analogue/event-driven hardware chain, and then coupled through saline for controlled electrode-mediated validation. When required, an auxiliary digital pulse-conditioning stage extends the pulse width while preserving event timing to ensure reliable stimulation of the memristive integrating sensor (MIS). Then, the device is periodically read at low voltage, such that its resistance trajectory provides a compact representation of recent neural activity (Sbandati et al., 2024; Gupta et al., 2016). To examine system performance in a controlled electrode-coupled bench scenario, we further performed saline-bath experiments that emulate extracellular neural recording, and observed reliable spike conversion together with repeatable activity-dependent resistance modulation. These results demonstrate that the proposed architecture can capture salient temporal variations in neural activity through memristive integration and provide a compact intermediate representation for downstream neural interfaces. This work is compared with prior art in Table 1.

Neural waveform source

Local field potential (LFP) recordings sampled at 40 kHz were obtained from a previously published (Sbandati et al., 2024) extracellular recording dataset acquired during in vivo experiments and were used here as the analogue input to the hardware pipeline. To support both rapid iteration and final evaluation, two datasets were prepared: a short ∼ 5 s segment containing a representative significant neural event, and the corresponding full-length recording.

Experimental scope and processing timeline

To avoid ambiguity between biological acquisition, offline preparation, and hardware operation, the experimental workflow in this study is separated into four distinct stages. First, the neural source data were prerecorded in vivo LFP recordings obtained from a previously published rat experiment (Sbandati et al., 2024). Second, the traces were preprocessed offline using low-pass filtering, normalisation, and amplitude scaling to generate the waveform used for hardware replay. Third, the processed waveform was replayed through the bench-top front-end hardware chain, including the CMOS neural amplifier, level-crossing spike generator, and FPGA-based pulse-conditioning stage. Fourth, the pulse-conditioned output was applied to the memristive integrating sensor (MIS), whose state was monitored through periodic low-voltage resistance readout using the ArC TWO platform.

Therefore, “end-to-end” in this manuscript refers to the complete bench-top hardware pathway from replayed neural waveform to memristive resistance trajectory, rather than live closed-loop in vivo acquisition. The saline-bath experiment should likewise be interpreted as a controlled electrode-coupling validation step, used to test whether electrode-mediated attenuation and waveform smoothing destabilise the amplifier–encoder pathway, rather than as a biological recording experiment.

Input signal conditioning and preprocessing

Neural traces were preprocessed offline in MATLAB using a sixth-order Butterworth low-pass filter with a cutoff frequency of 130 Hz, implemented in a Sallen–Key topology. The cutoff was selected to suppress high-frequency components while preserving the dominant LFP dynamics and to remain consistent with common LFP-oriented memristive processing configurations (Sbandati et al., 2024). After filtering, both datasets were normalized and amplitude-scaled for hardware replay; the filtered waveform was also mapped to a ± 1.5 V range to satisfy downstream analogue and MIS operating requirements.

Hardware platform

The experimental platform comprised a neural amplifier chip mounted on a custom PCB, a level-crossing spike generator chip, an FPGA development board for timing-preserving pulse conditioning, and an ArC TWO instrumentation platform for memristor biasing and resistance readout (Foster et al., 2022). The ArC TWO platform provides FPGA-coordinated source-measure and digital-control functionality through a programmable multi-channel architecture, enabling scripted pulse delivery, device readout, and experiment sequencing. A Python control layer coordinated stimulus delivery, parameter configuration, readout scheduling, and logging of time-stamped resistance measurements.

Neural amplifier configuration

Low-noise AC-coupled CMOS amplifiers are widely used in neural recording front-ends because they can reject large electrode DC offsets while preserving weak extracellular activity in the target frequency band (Harrison and Charles, 2003). The neural amplifier served as the analogue front-end conditioning stage between the preprocessed LFP waveform and the threshold-based spike generator (Jiang et al., 2023). Its role was to provide band-limited amplification of the low-amplitude neural input while preserving the temporal structure required for subsequent event-based encoding. The amplifier offers two selectable high-pass corners (1 Hz and 500 Hz), two low-pass corners (500 Hz and 10 kHz), and two gain settings (100 and 1,000). In this work, the 1 Hz high-pass, 500 Hz low-pass, and 1,000x gain configuration was selected so that the overall transfer characteristic matched the LFP band while providing sufficient gain for reliable spike generation. To emulate physiologically relevant input levels, the waveform delivered by the signal source was attenuated by three cascaded 20 dB attenuators, corresponding to a total attenuation of 60 dB. This reduced a 100 mV input to 100 uV at the amplifier input, placing the stimulus within the amplitude range expected for extracellular LFP recordings. The high-gain amplifier stage then restored the signal to a level suitable for the downstream spike generator, enabling controlled bench-top evaluation under neural-like operating conditions while preserving the relevant LFP waveform dynamics.

Spike generation and sensitivity settings

Event-driven encoding was performed using an asynchronous delta-modulation (ADM)/level-crossing spike generator (Guo et al., 2026). Encoder sensitivity was adjusted via an on-board trimmer controlling the effective emission threshold band, enabling comparison between high- and low-sensitivity operation. The native spike generator output produced fixed-amplitude digital pulses with widths up to ∼ 600 ns, which were sufficient to represent discrete events. In the present study, these settings are used as operational high- and low-sensitivity configurations based on their relative spike density under the same waveform input. They should not be interpreted as fully calibrated threshold regimes; quantitative calibration of the threshold band and spike-rate response is identified as a future optimisation step.

Timing-preserving pulse conditioning

To ensure robust interfacing to high-impedance memristive devices (typically > 50 M Ω during read), an FPGA pulse-conditioning stage detected each spike rising edge and generated a wider output pulse while preserving the original event timing. In the experiments reported here, the pulse-conditioning output was configured to deliver 3.3 V, 1 m programming pulses, thereby decoupling event timing (set by the spike generator) from delivered energy (set by pulse width).

Saline-bath electrode coupling validation

To evaluate robustness under extracellular-like coupling, a saline-bath experiment was performed using a commercial 32-channel ECoG micro-electrode array (Cambridge Neurotech). A filtered LFP waveform was injected into physiological saline via a stimulation electrode, while a separate electrode pair sensed the propagated extracellular potential. The electrode-derived waveform was routed through the neural amplifier and spike generator, and oscilloscope recordings captured the amplified analogue trace and the corresponding spike output to assess level-crossing alignment and spurious event behaviour.

The memristive integrating sensor (MIS) was stimulated by the pulse-conditioned spike train, leveraging cumulative energy delivery to induce resistive-state modulation, consistent with MIS principles reported in prior memristive neural interfacing studies (Gupta et al., 2016; Tzouvadaki et al., 2023). Device state was monitored using periodic low-voltage readout by the ArC TWO platform. For the volatile MIS operating condition used in this work, the readout voltage was set to V read = 0.2 V, with T sample = 1.5 ms and T hold = 0.2 ms per cycle (corresponding to a readout frequency of ∼ 588 Hz). Read currents were converted to resistance values R (t) = V read / I (t) and logged to an external host as time-stamped trajectories for subsequent analysis.

Event annotation and alignment analysis

Significant-event windows were defined on the preprocessed LFP waveform to provide a reference timeline for qualitative alignment between (i) waveform deflection, (ii) spike density variations under different sensitivity settings, and (iii) MIS resistance modulations. The annotation philosophy followed the conservative event-windowing rationale adopted in memristor-based LFP cue detection (Sbandati et al., 2024), enabling consistent discussion of prominent events versus borderline waveform features.

System concept/overview

Figure 1 provides a compact roadmap for the experimental question behind this study: if a neural interface is constrained by telemetry bandwidth and energy budget, does it really make sense to digitize and stream every sample? Put another way, can the front-end instead emit only those events that reflect meaningful low-frequency dynamics and store their recent history in a device-level state? Continuous sampling and high-speed data transmission remain the default pathway in many neural recording systems, and their cost becomes increasingly difficult to justify in wireless or implantable settings where power-saving opportunities are limited and data movement can dominate the overall system budget as channel counts scale (Even-Chen et al., 2020; Marblestone et al., 2013). The architecture summarized in Figure 1 is built around an alternative assumption: for many applications that rely on low-frequency neural activity trends (e.g., LFP-like components). The informative structure is often distributed over time rather than carried by every individual sample, so an encoding pipeline that naturally suppresses weak background fluctuations while highlighting sustained activity may offer a more efficient front-end representation.

The proposed signal path operationalizes this idea through a sequence of hardware-compatible transformations. Low-amplitude neural waveforms are first amplified and conditioned to map the signal into the operating range of subsequent stages, ensuring that downstream event generation is driven by physiologically relevant low-frequency content rather than high-frequency noise. In response to sharp signal transitions, a level-crossing event encoder based on asynchronous Delta modulation (ADM) (Guo et al., 2026) converts the conditioned waveform into a spike-like pulse train, shifting the representation from continuous amplitude tracking to activity-dependent event timing. This encoding strategy is consistent with asynchronous level-crossing acquisition, where the output rate follows signal dynamics rather than a fixed sampling clock, making it attractive for sparse biomedical signals (Li et al., 2013). The principle of level-crossing sampling is to represent a signal by the time instants at which it crosses predefined amplitude levels, producing a nonuniform event sequence whose density depends on signal activity (Sayiner et al., 1996). This event stream is particularly well matched to memristive integrating sensors (MIS), because memristive state evolution is governed by thresholded and history-dependent energy accumulation. Isolated or sparse pulses typically fail to produce measurable state change, whereas temporally dense bursts can drive clear resistive modulation, yielding an intrinsic form of selectivity and temporal summarisation (Gupta et al., 2016; Sbandati et al., 2024). However, practical implementation hinges on stimulus compatibility. In our platform, a programmable pulse-width control stage, implemented digitally and illustrated as FPGA-managed pulse shaping in Figure 1, tunes the delivered pulse width to meet the energy requirements of high-impedance devices while preserving the original event timing, so that event timing is retained even when the native spike energy would otherwise be insufficient. The resulting device state is interrogated periodically at low read voltage, and the measured resistance trajectory forms a low-rate, time-evolving representation of recent event-rich activity rather than a reconstruction of the original waveform. Because the present study does not implement a complete ADC-based baseline or an integrated telemetry link, we do not claim a directly measured reduction in total ADC or data-link power. Instead, the demonstrated resistance-state representation provides a hardware proof of concept for future integrated implementations in which full-rate waveform digitisation, storage, or transmission could be duty-cycled or partially avoided depending on the target decoding task, event definition, and readout protocol. The implemented hardware platform is shown in Figure 2. This sets up the three experimental blocks that follow: front-end validation, saline-coupling robustness, and MIS resistive encoding behaviour.

Front-end robustness validation under saline electrode coupling

The saline-bath experiment in Figure 3 was motivated by a pragmatic question that emerges once a neural front-end is taken out of an idealised, purely electrical testbench context: will the analogue conditioning and level-crossing event encoder remain stable when the input is no longer delivered as a clean, low-impedance voltage waveform, but instead is mediated by electrode coupling through a conductive volume? In extracellular electrophysiology, field potentials are shaped by the spread of transmembrane currents through tissue-like media, so what arrives at the acquisition electrodes is already a filtered and spatially averaged version of the underlying source activity (Buzsáki et al., 2012). This is precisely where front-end failure modes can hide: modest attenuation can move the signal closer to the encoder threshold, waveform smoothing can alter the slope at threshold crossings, and coupling-dependent noise pickup can seed spurious events. Because low-frequency signals such as LFPs and related modalities are widely used in BCI pipelines, and because practical acquisition chains must handle them under non-ideal coupling conditions (Moran, 2010; Sun et al., 2024; Wu et al., 2023), an intermediate validation stage that explicitly includes electrode coupling is valuable even before any biological experiments are considered.

In our implementation, physiological saline provides a controlled conductive medium in which the coupling pathway can be exercised while keeping the remainder of the signal chain unchanged (Figure 3). A pre-processed, bandwidth-limited LFP waveform was delivered into the bath through a stimulus electrode, and a separate set of recording electrodes sensed the resulting potential after propagation through the volume. This configuration deliberately separates “what we inject” from “what the front-end sees”: the injected stimulus sets a repeatable ground truth, while the recorded waveform captures coupling-induced effects such as amplitude loss and mild smoothing that are qualitatively consistent with how field potentials can be reshaped at cortical surface or intracranial interfaces (Miller et al., 2020). Importantly, the saline-bath does not aim to replicate the full complexity of tissue microstructure; rather, it provides a tractable way to introduce electrode-mediated transfer characteristics—and thus stress-test the event encoder—without introducing confounds from biological variability.

The saline-coupled signal was routed directly into the neural amplifier and then into the spike generator, mirroring the intended front-end operating chain. During the experiment, we monitored two signals in real time using an oscilloscope: the amplified LFP waveform (Channel 1) and the spike output of the event encoder (Channel 2), as illustrated in Figure 3. This dual-channel observation allowed the stability of event generation to be assessed relative to the analogue waveform that actually drove threshold crossings. Under the tested saline-coupling condition, the analogue trace exhibited the expected qualitative changes—attenuation accompanied by modest smoothing—yet the spike stream remained qualitatively well behaved. The output spikes were visually aligned with prominent threshold-crossing features in the conditioned waveform, indicating that the analogue conditioning stage preserved sufficient dynamic range and that the encoder did not collapse into a saturated or quiescent regime. In addition, oscilloscope monitoring did not reveal obvious isolated spikes during visually quiescent baseline segments. This observation should be interpreted as a qualitative front-end robustness check rather than a quantitative false-positive-rate measurement. Formal false-spike statistics would require threshold-defined spike labels, longer recordings, and event-level comparison between expected and observed spike times. Within the scope of this study, the saline experiment therefore demonstrates that electrode-mediated coupling did not visibly destabilize the amplifier–encoder pathway under the tested condition, while quantitative false-positive and false-negative analysis remains necessary for future benchmarking.

The MIS stage that follows relies on the density and temporal clustering of events to drive thresholded, history-dependent state modulation; as a result, robustness at the event-generation boundary is a prerequisite for attributing subsequent resistive transitions to neural activity rather than to coupling artifacts. The saline-bath validation therefore serves as a bridge between direct electrical injection and full end-to-end memristive encoding: it confirms that electrode-mediated propagation effects, at least in this controlled setting, do not prevent the analogue front-end and level-crossing encoder from producing clear, reproducible, activity-aligned spikes. With this robustness established, the subsequent MIS experiments can be interpreted in terms of selective integration and state-based summarization, rather than as a byproduct of fragile thresholding under non-ideal coupling.

MIS output and event-aligned resistive encoding

Figure 4 shows the encoding behaviour of the MIS, whereby a stream of spike-encoded activity is transformed into a compact resistance trajectory that tracks salient low-frequency events over time. Panel A shows a 25 s segment of the filtered LFP waveform, where each blue-shaded region denotes a significant-event window defined using the same auditory-cue-associated LFP event criterion as in Sbandati et al. (2024). In that definition, significant events correspond to LFP patterns elicited by the conditioned auditory cue, typically appearing as prominent positive and negative waveform peaks within the event window. The same event windows are overlaid on the MIS output in panel B to enable direct visual comparison. Across the majority of highlighted windows, the resistance trace departs from its quasi-stationary baseline and exhibits pronounced downward transitions, whereas the intervals between events remain comparatively stable aside from smaller readout fluctuations. This repeated co-occurrence supports the interpretation that the MIS is not simply reacting to isolated samples, but is instead modulated when the input contains temporally concentrated activity consistent with the selected event definition.

A more nuanced behaviour emerges around the ∼ 10 s region, where the LFP trace shows a smaller waveform deflection that is not highlighted as a significant event in panel A, yet a corresponding resistance perturbation is visible in panel B. Under the current annotation scheme, this feature is conservatively treated as an annotation-independent perturbation, or equivalently as a false-positive candidate for the MIS readout. In previous MIS-based LFP cue-detection work, event windows were defined using a repeatable criterion designed to isolate prominent cue-associated waveform patterns for downstream decision making, while sub-threshold or ambiguous waveform deflections were excluded to reduce false detections (Sbandati et al., 2024). Several mechanisms may therefore contribute to the observed response. First, the waveform deflection may represent borderline LFP activity that falls below the conservative significant-event criterion but still contains enough local amplitude or temporal structure to drive the level-crossing encoder. Second, the spike generator may emit threshold-crossing events outside the annotated windows when the local waveform slope or amplitude crosses its effective threshold band. Third, the volatile MIS device may exhibit relaxation-related fluctuations, readout noise, or operating-point-dependent resistance changes that resemble event-driven modulation. This ambiguity does not invalidate the observed event-aligned responses, but it shows that visual alignment alone is insufficient to establish task-level detection performance. The ∼ 10 s region is therefore used as a boundary case that motivates future threshold-defined event metrics, including false-positive and false-negative analysis, rather than being interpreted as a confirmed neural event.

Panels C–F then connect waveform structure to spike generation behaviour under different sensitivity settings, illustrating how the front-end can trade responsiveness for noise immunity. Panel C shows a zoomed significant event recorded in the saline experiment, captured after amplification; with the spike generator configured at a higher sensitivity, the event encoder produces a dense set of spikes (panel E). Such density is expected for level-crossing-style encoders, where a tighter effective threshold band yields more frequent crossing events, increasing responsiveness to waveform structure (Guo et al., 2026). Panel D instead includes both non-event activity and a significant event, while the spike generator is set to a lower sensitivity; under this configuration, the output in panel F becomes more selective, with only sparse spikes during the non-event segment and a clearer concentration of spikes during the highlighted event. At the same time, the lower-sensitivity setting produces fewer spikes overall during the event compared with the higher-sensitivity case, reflecting the intended suppression of weaker fluctuations and a reduced tendency to emit spikes in noise-dominated regions.

Taken together, these observations support the system-level rationale for coupling event-driven encoding to memristive integration. A high-sensitivity encoder can preserve subtle waveform structure, but it may pay for that responsiveness by emitting additional spikes outside the most salient epochs, whereas a lower-sensitivity setting reduces spurious activity at the cost of under-representing weaker waveform activity. The MIS offers a complementary selectivity mechanism because its state change is driven by history-dependent accumulation: temporally dense spike bursts are more likely to trigger measurable resistance modulation than isolated events, enabling a form of temporal compression in which the resistance trajectory emphasizes salient activity and suppresses background fluctuations (Gupta et al., 2016; Sbandati et al., 2024). In this context, the alignment between annotated significant events and resistive transitions (with the ∼ 10 s borderline case as a valuable exception) provides direct evidence that spike density and event structure can be mapped into a low-rate device-state representation suitable for subsequent low-power readout and analysis.

To quantify the correspondence between annotated neural events and the memristive state trajectory, the normalized resistance readout was converted into binary event decisions using compact features extracted over finite temporal windows. Because each feature descriptor integrates resistance dynamics over a different temporal scale, the analysis-window length was optimized separately for each strategy by selecting the window that produced the highest AUC within the evaluated sweep. The resulting optimal windows were 0.92 s for batch, 0.76 s for maxdiff, and 0.98 s for maxdrop. Among the three evaluated strategies, the batch-based descriptor yielded the clearest separation between event and non-event windows (Figure 5A). This indicates that, for the present spike-driven volatile device response, accumulated state displacement over a short analysis window is more informative than single-point maximum-difference or maximum-drop criteria.

At the selected operating point, the resulting classifier showed an asymmetric error profile, with stronger rejection of non-event periods than recovery of all annotated events (Figure 5B). This behaviour is consistent with a conservative sensor-proximal encoding stage: large event-driven resistance perturbations are detected reliably, whereas weaker or more ambiguous waveform deflections may remain below the decision boundary. The result therefore supports the role of the MIS as a low-dimensional activity summarizer rather than a waveform-preserving recorder, while also identifying sensitivity improvement as a key target for subsequent device- and threshold-level optimisation.

Device-level pulse-read operating window of the MIS

To quantify the operating window of the memristive integrating sensor independently of the upstream analogue front-end, controlled pulse-read measurements were performed directly using the ArC TWO platform. In this experiment, the neural amplifier, spike generator, FPGA pulse-conditioning stage, and saline interface were bypassed, allowing the response of the MIS itself to be isolated. Three volatile TiO x memristive devices were tested, with initial baseline resistances of approximately 34, 40, and 69 M Ω. For each device, a 3 × 3 stimulation matrix was applied, using pulse amplitudes of 3.0, 3.3, and 3.6 V and pulse widths of 500 μ s, 1 m, and 2 m. Each pulse condition was repeated eight times in pulse-read mode, where the voltage pulse was followed immediately by a low-voltage resistance readout. The normalised resistance modulation was calculated Equation 1 as follows: where R 0 is the median baseline resistance obtained from the read-only measurements before stimulation, and R pulse - read is the resistance extracted immediately after the applied pulse.

The pulse-read measurements reveal a clear transient resistive response across all three devices, as shown in Figure 6A. At the reference operating condition used in the system-level experiments, 3.3 V and 1 m, the three devices exhibited repeatable resistance drops followed by rapid recovery during subsequent read-only operations. This behaviour confirms that the MIS response under these conditions is volatile and event-driven, rather than the result of permanent low-resistance-state programming. The recovery of the read-only points toward the baseline also indicates that the pulse-read measurement is essential for capturing the short-lived device-state perturbation.

The dependence of the MIS response on pulse amplitude and pulse width is summarised in Figures 6B,C. Across the three devices, the mean normalised resistance modulation increased from 10.9 ± 5.3 % at 3.0 V and 500 μ s to 20.9 ± 6.2 % at 3.6 V and 2 m. At a fixed pulse amplitude, increasing the pulse width consistently increased the measured resistance modulation. Similarly, at a fixed pulse width, increasing the pulse amplitude produced a larger transient resistance drop. This trend supports the interpretation that the volatile MIS response is controlled by the effective stimulation strength delivered to the device. The 3.3 V, 1 m condition produced a mean response of 16.9 ± 4.6 % across devices, placing the pulse setting used in the main system experiments within a reliable operating region.

Device-to-device variability is visible in Figure 6D. Although the absolute magnitude of Δ R / R 0 differs between devices, all three devices show a measurable pulse-induced resistance modulation at the reference condition. The variability is therefore primarily reflected in response amplitude rather than in the presence or absence of switching behaviour. This is consistent with the expected device-level dispersion of filamentary or interface-mediated resistive switching processes in metal-oxide memristors. Importantly, the monotonic response trend with pulse width and pulse amplitude remains preserved when the results are aggregated across devices, demonstrating that the observed behaviour is not restricted to a single selected device.

These controlled pulse-read measurements provide a quantitative basis for the selected MIS stimulation parameters. They show that the resistance response is not merely a qualitative visual effect, but a reproducible transient modulation whose magnitude can be tuned by pulse amplitude and pulse duration. The results also define the practical operating window of the volatile MIS: weaker stimulation produces smaller resistance perturbations, while stronger stimulation increases the response amplitude but may also increase device-to-device spread. The 3.3 V, 1 m setting therefore represents a balanced operating point that produces a robust transient response while avoiding the need for stronger stimulation conditions.

Power and energy estimation

Power consumption is an important constraint for implantable neural interfaces, where continuous waveform acquisition, digitisation, and data transmission can dominate the system energy budget (Even-Chen et al., 2020; Marblestone et al., 2013). To evaluate the proposed spike-driven MIS platform, the per-channel power budget is divided into five functional blocks: the neural analogue front-end, the event-driven spike encoder, the pulse-conditioning logic, the memristive read-write process, and the resistance readout and digital backend. The benchtop instruments used in this work, including the ArC TWO platform, the Basys3 board, the signal generator, the oscilloscope, and laboratory power supplies, are not included in the implant-oriented power estimate, because they are validation instruments rather than application-specific integrated hardware.

The total per-channel power can be expressed as follows (Equation 2): here, P AFE is the neural amplifier power, P enc is the spike encoder power, P pulse is the pulse-conditioning power, P MIS is the memristive device read-write power, and P readout represents the resistance digitisation and lightweight digital processing power.

For the analogue front-end, the measured per-channel power is approximately 19.8 uW, which is consistent with the 20.4 uW per-channel consumption reported for the 16-channel reconfigurable CMOS neural front-end used as a reference architecture (Jiang et al., 2023). The spike generation stage follows an event-driven level-crossing principle, where output events are generated only when the input signal crosses predefined thresholds, instead of continuously digitising the full waveform (Sayiner et al., 1996; Li et al., 2013). The complete adaptive auditory encoder reported by Dongxu Guo et al. consumes 299.85 uW per channel, but this value includes both the ADM spike-generation circuit and the threshold-generation/adaptive-threshold control circuitry (Guo et al., 2026). In the present experiment, the adaptive threshold generator was not used. Therefore, only the active spike-generation circuit is included in the present power budget. Based on the post-simulation average supply current, the spike-generator power is estimated as 1.5 V multiplied by 1.474 uA, corresponding to 2.211 uW. This value is more representative of the circuit block actually used in this spike-driven MIS experiment than the full adaptive auditory encoder power.

The memristive device contribution is estimated from the measured read and write conditions. During readout, a subthreshold voltage is applied across the memristor for a short hold time. By approximating the device as a resistance, the instantaneous read power is calculated in Equation 3 as follows: the average read power is obtained by multiplying the instantaneous read power by the read duty cycle (Equation 4): using a read voltage of 0.2 V, a read resistance of approximately 55 Mohm, a sampling interval of 1.5 m, and a hold time of 0.2 m, the instantaneous read power is calculated in Equation 5 as follows: the average read power is therefore calculated in Equation 6: the corresponding read energy per cycle is calculated in Equation 7:

the write operation is event-driven and only occurs when incoming spikes stimulate the volatile TiOx memristor. The write energy per spike is estimated as follows in Equation 8: the corresponding average write power is calculated in Equation 9: where f spike is the average spike rate. In this work, the write voltage is 3.3 V, the pulse width is 1 m, the average write current is approximately 4 nA, and the average spike rate is approximately 1.0–1.2 Hz. Therefore, Equation 10 calculates energy per spike as follows: the average write power is calculated in Equation 11: combining the read and write contributions gives the average device-level MIS power in Equation 12 as follows: this result shows that the volatile memristive element itself operates in the sub-nW regime. Under the present operating conditions, the device-level power is dominated by periodic readout, while the write contribution remains small because programming is only triggered by sparse spike events.

In the current modular prototype, the total active signal-chain power is dominated by the spike encoder rather than by the memristive device. Excluding laboratory instrumentation and board-level static power, the present prototype path consumes approximately:

The value calculated in Equation 13 should be interpreted as a prototype-level upper bound. In an integrated implementation, the FPGA pulse extension could be replaced by a compact edge-triggered pulse stretcher, and the encoder could be redesigned as a low-bandwidth biomedical level-crossing circuit. Previous MIS work has shown that the readout and digital backend can be implemented at nW or sub-nW levels when the resistance readout rate is low and the resistance values are coarsely quantised (Sbandati et al., 2024).

Compared with conventional ADC-based neural acquisition, the proposed architecture shifts part of the processing burden from continuous waveform digitisation to sparse event generation and memristive temporal integration. Conventional neural-interface pipelines commonly follow an amplifier–ADC–digital-processing structure, in which the full neural waveform is digitised before event detection or classification, making data movement and downstream digital processing important power and bandwidth constraints in scalable neural interfaces (Sun et al., 2024; Wu et al., 2023; Even-Chen et al., 2020). For an LFP-oriented acquisition chain, a 500 Hz signal bandwidth is a reasonable front-end design assumption, consistent with low-frequency neural recording configurations and reconfigurable LFP front-end specifications (Buzsáki et al., 2012; Jiang et al., 2023). A Nyquist-rate ADC would therefore require at least 1 kS/s sampling. Using a representative 10-bit ADC resolution, the corresponding raw digitised data rate is per channel before downstream event detection.

In the proposed MIS-based pathway, digitisation is shifted after event-driven encoding and memristive temporal integration. The downstream ADC therefore digitises the low-rate resistance trajectory rather than the full neural waveform. Prior LFP-MIS work used memristive state changes as the event-related biomarker and estimated the post-MIS readout digitisation using a 5-bit ADC, with an effective readout frequency of 7.15 Hz for the resistance-derived feature stream (Sbandati et al., 2024). Interpreting this as a post-integration signal bandwidth gives a Nyquist sampling rate of 14.3 S/s, corresponding to per channel. Under these assumptions, the effective digitised output data-rate reduction is corresponding to a > 100 × compression ratio. Equivalently, the post-front-end digitisation rate is reduced from 1 kS/s to 14.3 S/s, giving an approximately 70 × reduction in sampling/readout rate.

This estimate should be interpreted as a protocol-level comparison rather than a direct system-level measurement of the present modular prototype, which does not include an integrated ADC or telemetry link. Nevertheless, it quantifies the central architectural advantage of the MIS approach: the memristive sensing element compresses recent neural activity into a low-rate resistance trajectory, allowing subsequent digitisation and digital processing to operate on a compact event-derived state variable rather than on the continuously sampled neural waveform.

Key findings

The results demonstrate that a level-crossing neural front-end can transform low-frequency LFP activity into spike-based events that are subsequently integrated by a memristive integrating sensor (MIS) into a compact device-state trajectory. By comparing the significant-event windows in the filtered waveform with the corresponding MIS response, we observe measurable resistance modulations that align with these events in time (Figures 4A,B). This supports the interpretation that the MIS does not simply preserve the input waveform, but instead acts as a history-dependent state variable whose resistance is displaced when spike activity becomes temporally concentrated.

This behaviour is consistent with prior MIS operation, where thresholded and history-dependent switching provides selectivity to event density and enables temporal compression (Gupta et al., 2016; Sbandati et al., 2024). In the present system, this selectivity is introduced at two different levels. The level-crossing encoder determines which waveform variations are converted into spikes, while the MIS determines which spike patterns accumulate sufficient stimulus history to produce measurable state change. This layered selectivity is evident in the sensitivity-dependent measurements in Figures 4C–F. A higher sensitivity setting preserves more waveform structure by producing a denser spike train, whereas a lower sensitivity setting suppresses weaker background fluctuations and concentrates spikes around the significant-event window. The MIS then further filters sparse activity through its intrinsic accumulation dynamics, so that temporally clustered stimulation is more likely to be expressed as a resistance transition.

The unhighlighted waveform feature near ∼ 10 s provides a useful boundary case. Although this feature is not labelled as a significant event, a resistance perturbation is still visible in the MIS trajectory. This observation does not necessarily indicate an annotation error; rather, it highlights the fact that event definition and device response are coupled design choices. Therefore, in the absence of a quantitative classifier, we do not interpret this perturbation as a detected neural event; rather, we report it as a boundary case that motivates future false-positive analysis. A waveform pattern that falls below the selected event criterion may still deliver enough cumulative stimulation to perturb the device state, particularly when the device operates near its effective switching boundary. Conversely, volatility and readout dynamics may also produce perturbations that resemble event-driven modulation. In either case, the result suggests that event definitions in MIS-based front-ends should not be treated as purely signal-processing choices, but as system-level parameters that interact with encoder sensitivity, pulse energy, device operating point, and readout interval.

The saline-bath validation further supports the robustness of the front-end under a more realistic signal-transfer condition. A front-end that performs well under direct electrical injection may still fail when the input is mediated by electrode coupling, where attenuation, smoothing, and noise pickup can shift waveforms relative to fixed thresholds. In the saline experiment, the system continued to produce clear and repeatable threshold-aligned spikes after the LFP waveform was propagated through a conductive medium and sensed by recording electrodes (Figure 3). Although saline does not reproduce the full complexity of neural tissue, it introduces an extracellular-like transfer path and therefore provides an intermediate validation step between ideal bench-top injection and future biological experiments.

Limitations

The present study should be interpreted as a feasibility-oriented hardware validation rather than a fully integrated neural interface demonstration. The main limitations arise from the device dynamics, prototype implementation, and scale of validation, rather than from the event-driven encoding principle itself. The present revision therefore frames the study as a proof-of-concept hardware validation rather than a complete event-detection benchmark. The current evidence supports feasibility of the signal pathway, but does not yet establish task-level sensitivity, specificity, or false-positive/false-negative rates across long recordings.

A first limitation is the retention behaviour of the volatile memristive devices used in the MIS stage. Although volatility is useful for short-term temporal integration and for suppressing weak isolated activity, it also means that spike-induced resistance changes may partially relax before they are captured by periodic readout. This is particularly important when the device is operated near its effective switching boundary or when the incoming spike density is low. As a result, the readout interval is not a secondary measurement detail, but a system-level parameter that determines which resistance transitions are observable. Future implementations will therefore need to co-optimize device volatility, pulse energy, encoder sensitivity, and readout interval to avoid missed transitions while preserving the desired temporal integration behaviour.

A second limitation is the prototype-oriented implementation of the pulse-conditioning stage. In the present platform, auxiliary digital pulse conditioning is used to extend the native spike-generator output so that each event can reliably stimulate the high-resistance MIS device. This approach is useful for controlled experiments because it preserves event timing while allowing the delivered pulse width to be tuned. However, it is not intended as the final deployment architecture. A practical integrated system would replace this function with a minimal low-power pulse-shaping circuit, or merge spike generation and pulse conditioning into a compact mixed-signal front-end.

A third limitation concerns scalability and device variability. The multi-channel electrode array was used to create an electrode-coupled saline recording pathway; it should not be interpreted as a demonstration of simultaneous multi-channel MIS processing. The experiments reported here validate a single-channel signal path, whereas practical neural interfaces require stable operation across multiple channels, devices, and recording conditions. Scaling the architecture will require strategies for handling device-to-device variability, stochastic switching, calibration drift, and multiplexed readout. These issues are especially important for memristive systems, where the operating point of each device can influence whether a given spike pattern produces a measurable state transition.

Finally, the present analysis focuses on temporal alignment between significant-event windows, spike-density changes, and MIS resistance modulation. The significant-event definition follows the auditory-cue-associated LFP event criterion used by Sbandati et al. (2024), but this work does not yet provide a complete task-level detection framework. Quantitative evaluation using longer recordings and event-level metrics will be needed to assess sensitivity, specificity, false detections, and robustness across different operating conditions. Although the saline bath captures conductive coupling and waveform attenuation, real neural interfaces introduce additional material and biological constraints associated with electrode composition, surface properties, and tissue compatibility (Fattahi et al., 2014). The saline-bath experiment should therefore be interpreted as an intermediate electrical validation, since chronic implantation can involve tissue reaction, encapsulation, and electrode impedance changes that are not reproduced by physiological saline (Polikov et al., 2005).

Future outlook

The next stage of development should focus on co-optimizing the encoder, pulse-conditioning stage, MIS operating point, and readout protocol as a single coupled system. Encoder sensitivity controls which waveform features are promoted into spikes, pulse width and amplitude determine the stimulus delivered to the device, and the readout interval determines whether transient resistance changes are captured before relaxation. Tuning these parameters jointly would allow the system to preserve sensitivity to weaker but structured events while reducing spurious responses to background fluctuations.

A second direction is hardware integration. The present platform demonstrates the feasibility of the full signal chain using modular components: offline signal preparation, analogue conditioning, threshold-based spike generation, auxiliary pulse conditioning, and scripted resistance readout. A deployment-oriented system should retain only the essential functions and consolidate them into a purpose-built mixed-signal front-end. In such an implementation, minimal pulse shaping could be integrated close to the spike generator, while multiplexed readout and shared support circuitry could reduce per-channel overhead. This integration pathway is complementary to conventional embedded neural processors that perform digital feature extraction or online spike sorting after front-end acquisition; in contrast, the present architecture shifts part of feature formation into an event-driven and device-state-based analogue pathway (Chen et al., 2023).

Finally, future validation should move from controlled replay and saline-coupled tests toward longer-duration and multi-condition datasets. This would make it possible to quantify detection performance, characterize failure modes near the event boundary, and evaluate stability across different signal amplitudes, coupling conditions, and device states. Long-term high-density recording platforms such as Neuropixels 2.0 further emphasize that future front-ends must remain stable across extended recording periods and multi-channel operating conditions (Steinmetz et al., 2021). Extending the platform to multi-channel operation would also allow investigation of calibration strategies and adaptive readout policies that compensate for memristive variability without reverting to continuous full-rate waveform transmission.

Summary

This work demonstrates that combining level-crossing event encoding with MIS-based temporal integration can condense salient low-frequency neural activity into a compact resistance trajectory. The main finding is not simply that the MIS responds to neural-derived stimulation, but that the full front-end pathway can preserve event timing, tolerate saline-mediated electrode coupling, and translate temporally clustered spike activity into measurable device-state modulation. The borderline waveform feature near ∼ 10 s further highlights an important design principle: event labels, encoder sensitivity, pulse delivery, device dynamics, and readout interval must be considered together. These results support further investigation of spike-based MIS front-ends as a route toward compact neural signal representations, while the present study remains a single-channel, replay-based proof-of-concept requiring quantitative detection benchmarks, multi-device validation, and integrated low-power implementation before implantable or multi-channel deployment claims can be made.