Memristor-Driven Spike Encoding for Fully Implantable Cochlear Implants

TL;DR

This work addresses the energy and form-factor limitations of fully implantable cochlear implants by proposing a biomimetic front-end that combines FFT-free, frequency-selective piezoelectric MEMS sensing with a $VO_2$ memristor-based relaxation oscillator to directly generate neuromorphic spikes. The system uses a cantilever array covering $f_i^{res.}$ in the $200$–$700$ Hz range, where the oscillator frequency $f_{osc}$ encodes the stimulus amplitude, yielding spike rates from approximately $100$ Hz to $1$ kHz and enabling rapid, low-power temporal coding. A crucial result is the demonstration that a single channel can convert mechanical displacement into biomimetic, biphasic stimulation after adding a parallel LR network, addressing safety concerns for CI electrodes. Overall, the approach promises a compact, energy-efficient front-end for fully implantable Cochlear Implants, with potential latency reductions and opportunities for phase-locking and multi-channel integration in future devices, aligning with neural-inspired temporal encoding strategies.

Abstract

Objective: This work aims to demonstrate a low-power, biomimetic auditory sensing concept for fully implantable cochlear implants. The approach draws inspiration from the frequency selectivity and temporal encoding of the cochlea, and uses neuromorphic spike generation to replace conventional signal processing blocks. The goal is to establish a compact, energy-efficient front-end architecture suitable for future implantable systems. Methods: An auditory sensing unit was implemented, consisting of a piezoelectric MEMS cantilever mechanically coupled to a single VO$_2$ nanogap Mott memristor-based oscillator. This configuration enables FFT-free, frequency-selective sensing and direct spike generation, forming a biomimetic auditory front end. The concept was experimentally examined using controlled mechanical excitation. Results: The sensing unit exhibited frequency-selective detection of mechanical vibrations in the nanometer to tens-of-nanometers displacement range and generated biomimetic spiking waveforms. Spike rate-encoding of the input amplitude was demonstrated, with output spiking frequencies tunable between approximately 100 Hz and 1 kHz depending on the excitation level. The waveform was finally converted to a biphasic shape suitable for cochlear implant stimulation. Significance: Temporal encoding is fundamental to natural auditory signal processing in the nervous system. By implementing this principle through neuromorphic spike encoding, the proposed approach can provide significant benefits for cochlear implants. In addition, the circuit has the potential to reduce footprint, energy consumption, and latencies compared with current commercial solutions.

Figures (8)

• Figure 1: Concept of our bio-inspired auditory sensing system. Vibroacoustic stimuli composed of several different frequency signals ($S=\sum S(f)$) are sensed by piezo-MEMS cantilevers, each having a well-defined resonance frequency, $f_i^{\rm res.}$, realizing frequency-selective sensing. The output signals of the cantilever array are carried to VO$_2$ memristor-based relaxation oscillator circuits ($i$ number of channels), which emit neural spikes proportional to the amplitude of the incoming stimulus in the $i^{\rm th}$ channel, $f_i^{\rm osc.}\sim S(f_i^{\rm res.})$. The auditory sensing system realizes rate-encoding in $i$ channels, creating spiking waveforms suitable for further processing by the nervous system, aimed at using in FICIs.
• Figure 2: Measurement setups for the characterization of piezo-MEMS cantilevers. (a) Schematics of the interferometric measurement setup used for the mechanical characterisation of cantilevers. (b) Photo of the custom-built sample holder for the electromechanical test: (A) piezoelectric actuator, (B) shielded electrical connections to the piezoelectric actuator, (C) custom PCB holding the chip with cantilevers, (D) flexible ribbon cables to capture the cantilever's signal, (E) custom PCB for routing the connections to the cantilevers, (F) BNC connector, (G) jumper used for selecting the desired electrode of a cantilever. Scale bar: 2 cm.
• Figure 3: Fabrication and characterization of VO$_2$ memristors. (a) Illustration of the planar memristor design showing layer thicknesses. (b) Scanning electron microscopy image of the nanogap region of a representative VO$_2$ device. Scale bar: 100 nm. (c) Circuit schematic for $I(V)$ characterization of the VO$_2$ samples. Drive voltage $V_{\rm drive}$ is applied to the memristor and a series resistor $R$, while current $I$ is monitored. Bias voltage is calculated as $V_{\rm bias} = V_{\rm drive}-R\cdot I$. (d) Typical $I(V)$ curves of a VO$_2$ device exhibiting volatile resistance switching, $R=380~\Omega$. Current is also displayed as a function of bias voltage (blue curve) and drive voltage (gray curve). The set process/IMT (reset process/MIT) is indicated by a pink (green) arrow on the $I(V_{\rm bias})$ curve. Note that these transitions are not resolved on the 12 kSa/s sampling rate of the measurement; there are no datapoints in the sections indicated with arrows.
• Figure 4: Relaxation oscillator made of a VO$_2$ memristor. (a) Current waveform $I_{\rm osc.}(t)$ of an oscillator. (b) Magnified $I(V_{\rm bias})$ characteristics of a memristor (positive quadrant) shown with blue colors (transitions blurred with light blue). Grey line marks the possible equilibrium currents defined by the load line of the oscillator circuit according to the equation $(V_{\rm in} - V_{\rm bias} )/ R$. Equilibrium points defined by the continuation of the HRS and LRS are not reached, as marked by the red crosses. (c) Voltage waveform $V_{\rm osc.}(t)$ of an oscillator. The green shaded area and pink/green labels of $V_{\rm set}$/$V_{\rm reset}$ in panels (b-c) highlight the operation regime of the memristor during oscillation. Pink/green encircled points in panels (a-c) mark the points where set/reset transitions occur, and black arrows mark the direction. (d) Circuit schematics of the relaxation oscillator (output current/voltage and the memristor element are highlighted in blue).
• Figure 5: Circuit schematics of a single channel of the proposed auditory sensing system. A piezo-MEMS cantilever with a well-defined frequency resonance senses a sinusoidal vibroacoustic stimulus provided at the same frequency. As a result, an AC voltage signal is generated by the cantilever, which is amplified ($V_{\rm MEMS}^{\rm AC}$) and rectified ($V_{\rm MEMS}^{\rm DC}$) by a custom-built circuitry. The $V_{\rm MEMS}^{\rm DC}$ signal serves as an input DC voltage for the oscillator circuit, which emits current spikes at its output. These spikes are measured with the $50~\Omega$ input terminal of a digital oscilloscope.